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

Abstract

The invention discloses a policy text analysis method and device, computer equipment and a storage medium. The method comprises the steps of obtaining policy information to be analyzed; analyzing the policy information to be analyzed in a data enhanced recall mode to obtain retrieval recall data; performing task splitting on the policy information to be analyzed to obtain a plurality of sub-tasks; determining an analysis agent matched with the plurality of sub-tasks, and analyzing the retrieval recall data in parallel through the analysis agent to obtain sub-task candidate results of the plurality of sub-tasks; and integrating the plurality of sub-task candidate results to obtain an analysis result of the policy information to be analyzed. According to the method and the device, through various data enhancement recall modes, the recall rate of the actual associated text blocks is effectively improved on the data set after data enhancement.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of intelligent government affairs, and in particular to a policy text analysis method and device, computer equipment and a storage medium. BACKGROUND

[0002] A large language model (hereinafter referred to as a large model) is a technology based on artificial intelligence deep learning, which can understand, summarize, generate and translate human language by pre-training on a large amount of data, and has shown logical reasoning and creativity.

[0003] Based on the excellent language understanding ability shown by the large model, more and more text processing work has begun to combine the large model to improve work efficiency. In the government field, the interpretation and processing of policy text are the key work content of civilian civil servants. In the context of the government department accelerating the digital transformation, using the large model to assist in completing related work has become an important means for the intelligentization and scientization of government work.

[0004] Although the large model has powerful capabilities, it is not always correct. At present, it is essentially a probability-based prediction system based on the Transformer model architecture, and many times it will fabricate false information, which is commonly known as the "hallucination" problem. Therefore, how to get more accurate and comprehensive analysis results is a big difficulty. SUMMARY

[0005] The embodiments of the present application provide a policy text analysis method, device, computer equipment and storage medium, which effectively improves the recall rate of actual associated text blocks on the data set after data enhancement through multiple data enhancement recall methods.

[0006] In a first aspect, the embodiments of the present application provide a policy text analysis method, which comprises: obtaining policy information to be analyzed; analyzing the policy information to be analyzed by using a data enhancement recall method to obtain retrieval recall data; task decomposition is performed on the policy information to be analyzed to obtain a plurality of subtasks; determining an analysis agent matched with the plurality of subtasks, and obtaining subtask candidate results of the plurality of subtasks by parallel analysis of the retrieval recall data through the analysis agent; integrating the plurality of subtask candidate results to obtain an analysis result of the policy information to be analyzed.

[0007] In some embodiments, the analyzing the policy information to be analyzed by using a data enhancement recall method to obtain retrieval recall data comprises: perform retrieval analysis on the to-be-analyzed policy information based on a preset retrieval engine to obtain relevant text of the to-be-analyzed policy information; and / or, perform matching analysis on the to-be-analyzed policy information based on a preset policy dictionary to obtain paraphrase information of the to-be-analyzed policy information; and / or, perform question segmentation on the to-be-analyzed policy information based on a preset model to obtain core phrases of the to-be-analyzed policy information; and / or, determine associated information of the to-be-analyzed policy information based on label information embedded in advance in each type of policy information in a database, and the retrieval recall data includes the relevant text, the paraphrase information, the core phrases, and the associated information.

[0008] 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 first type text, second type text, and third type text, the retrieval analysis on the to-be-analyzed policy information based on the preset retrieval engine to obtain relevant text information of the to-be-analyzed policy information includes: perform key information matching on the to-be-analyzed policy information based on the sparse retrieval layer to recall the first type text related to the to-be-analyzed policy information; perform semantic understanding on the to-be-analyzed policy information based on the dense retrieval layer to recall the second type text semantically related to the to-be-analyzed policy information; analyze the third type text related to the to-be-analyzed policy information based on the knowledge graph layer.

[0009] In some embodiments, before the task decomposition on the to-be-analyzed policy information to obtain a plurality of sub-tasks, the method further includes: identifying intent information of the to-be-analyzed policy information based on a preset intent recognition model; if the intent information matches a preset intent scenario, determining an intent fence corresponding to the preset intent scenario as an intent fence of the intent information; the obtaining of the sub-task candidate results of the plurality of sub-tasks by the analysis agent in parallel includes: obtaining the sub-task candidate results of the plurality of sub-tasks by the analysis agent in parallel based on the intent fence.

[0010] In some embodiments, the obtaining of the analysis result of the to-be-analyzed policy information by integrating the plurality of sub-task candidate results includes: integrating the plurality of sub-task candidate results to obtain analysis basis information of the plurality of sub-task candidate results; The analysis result of the policy information to be analyzed is selected from the plurality of subtask candidate results based on the analysis basis information.

[0011] In some embodiments, the selecting the analysis result of the policy information to be analyzed from the plurality of subtask candidate results based on the analysis basis information comprises: identifying conflicting results in the plurality of subtask candidate results by a debating agent; determining a reserved result in the conflicting results based on the analysis basis information by an arbitration agent; obtaining the analysis result of the policy information to be analyzed based on the reserved result and non-conflicting results in the plurality of subtask candidate results.

[0012] In some embodiments, before the determining the analysis agent matching the plurality of subtasks, the method further comprises: if a target subtask is added, adding an analysis agent corresponding to the target subtask.

[0013] In a second aspect, the embodiments of the present application provide a policy text analysis device, and the device comprises: a data acquisition module configured to acquire policy information to be analyzed; a data retrieval module in communication connection with the data acquisition module and configured to analyze the policy information to be analyzed in a data enhancement recall manner to obtain retrieval recall data; a task splitting module in communication connection with the data retrieval module and configured to split the policy information to be analyzed into a plurality of subtasks; an intelligent analysis module in communication connection with the task splitting module and configured to determine an analysis agent matching the plurality of subtasks, and to obtain subtask candidate results of the plurality of subtasks by parallel analyzing the retrieval recall data through the analysis agent; a result integration module in communication connection with the intelligent analysis module and configured to integrate the plurality of subtask candidate results to obtain an analysis result of the policy information to be analyzed.

[0014] In a third aspect, the embodiments of the present application further provide a computer device comprising a processor and a memory, wherein the memory stores at least one computer program, the at least one computer program is loaded and executed by the processor to implement the operations performed by the policy text analysis method.

[0015] In a fourth aspect, the embodiments of the present application further provide a computer readable storage medium, wherein at least one computer program is stored in the computer readable storage medium, and the at least one computer program is loaded and executed by a processor to implement the operations performed by the policy text analysis method according to any one of the preceding aspects.

[0016] By using the scheme of the embodiments of the present application, the recall rate of the actual associated text block is effectively improved on the data set after data enhancement in the manner of recalling multiple data enhancements, and under the mixed intention problem, multiple sub-tasks are effectively disassembled to improve the accuracy of the answer, and the task efficiency is improved by multi-task coordination and parallel processing. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0018] Figure 1 is a flowchart of a policy text analysis method provided in the embodiments of the present application; Figure 2 is a flowchart of a policy text analysis method provided in the embodiments of the present application; Figure 3 is a flowchart of a policy text analysis method provided in the embodiments of the present application; Figure 4 is a flowchart of a policy text analysis method provided in the embodiments of the present application; Figure 5 is a structural schematic diagram of a policy text analysis device in the embodiments of the present application; Figure 6 is a structural schematic diagram of a terminal in the embodiments of the present application; Figure 7 is a structural schematic diagram of a server in the embodiments of the present application. DETAILED DESCRIPTION

[0019] In order to make the purpose, technical solutions and advantages of the embodiments of the present application more clear, the embodiments of the present application will be further described in detail with reference to the drawings.

[0020] It can be understood that the terms "first", "second", and the like used in the present application can be used herein to describe various concepts, but unless specifically 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 the present application, a first impact parameter can be referred to as a second impact parameter, and similarly, a second impact parameter can be referred to as a first impact parameter.

[0021] At least one refers to one or more than one, for example, at least one event type can be one event type, two event types, three event types, or any integer greater than or equal to one event type. A plurality refers to two or more than two, for example, a plurality of event types can be two event types, three event types, or any integer greater than or equal to two event types. Each refers to each of at least one, for example, each event type refers to each of the plurality of event types, if the plurality of event types is 3 event types, each event type refers to each of the 3 event types.

[0022] It can be understood that in the embodiments of the present application, data related to user information and the like are involved, and when the above embodiments of the present application are applied to specific products or technologies, user permission or consent needs to be obtained, and the collection, use and processing of related data need to comply with relevant laws, regulations and standards of countries and regions.

[0023] By accessing massive historical policy texts, social statistical data, academic research reports, and public opinion information, large models can provide data-driven insights for decision-makers. First, it can efficiently complete multi-dimensional, deep analysis of complex problems with its computing power advantage, greatly surpassing human capabilities in efficiency; second, large models also have the ability to simulate and predict effects, so that they not only can quickly generate policy drafts and provide multiple alternative solutions, but also can warn potential risks and identify policy blind spots, improving the efficiency and forward-looking nature of the decision-making process.

[0024] For such a very vertical scenario as the government affairs field, there is often a lack of high-quality information in the training data of the original model, resulting in subpar performance of large models in these vertical fields compared to general knowledge fields.

[0025] To minimize the impact of "hallucinations", the current common solution is to use retrieval augmented generation (Retrieval Augmented Generation, hereinafter referred to as RAG) and other technologies to provide additional information related to the task as context to the large model, so that the large model can combine the user task with this additional information to generate more accurate and relevant answers.

[0026] RAG is an agent technology that combines large language models and information retrieval. When a large model needs to generate text or answer questions, RAG first retrieves relevant information from an external knowledge base, and then uses the information as context for the model instructions, improving the quality of the large model output.

[0027] During the index construction phase, the functions of each module are as follows: 1. Document loading: Process raw documents of various formats, extract pure text and metadata through document parsing technology.

[0028] 2. Text segmentation: Based on specific logic, long text is divided into semantically related text chunks. The size of the text chunk will affect the efficiency and quality of subsequent retrieval recall and model answer.

[0029] 3. Text embedding: Use embedding models to convert each text chunk into a high-dimensional vector, mapping the semantic information of the text into a high-dimensional space, so that the correlation between texts can be quantified using cosine similarity and other calculation methods.

[0030] 4. Vector storage: Store the text chunk (and metadata) and its corresponding text vector in the vector database, supporting efficient semantic similarity retrieval.

[0031] During the agent runtime phase, the functions of each module are as follows: 1. Query embedding: Use the same embedding model as in the index construction phase to convert the user query into a vector.

[0032] 2. Semantic recall coarse screening: Use the index of the vector database to efficiently retrieve the Top-K text chunks with the highest correlation (usually using cosine similarity) to the user query vector.

[0033] 3. Document rearrangement fine screening: Use a more refined rearrangement model to reorder and filter the Top-K text chunks obtained by coarse screening, to obtain the Top-N text chunks with the highest correlation to the user query.

[0034] 4. Construct enhanced prompt words: Combine the user query and the Top-N most relevant text chunks obtained after rearrangement into a structured prompt word, enhancing the background information that the large model can refer to when generating answers.

[0035] Although RAG uses text embedding and vector database retrieval technology to provide additional information context for large models, when the document block retrieval results are not good and the additional information is insufficient, the "hallucination" problem of large models is still serious. The general RAG solution lacks detection of such "hallucination" cases and cannot correct "hallucination" problems.

[0036] In order to reduce the probability of the "hallucination" problem in the RAG scheme, the quality of the document block retrieval result needs to be improved, and the general semantic correlation degree measurement method used in the existing scheme is limited by the quality of the vertical field training data, and the actual effect in the government field still has a lot of room for improvement.

[0037] For example, "business environment" and "building a fair market competition for enterprises" have a low correlation under the general semantic correlation degree measurement. However, in the government field, the connotation of "business environment" includes "policy environment, market environment, rule of law environment, and humanistic environment", and it mainly faces enterprises, and has a very close correlation with the latter.

[0038] When solving complex tasks that need to be decomposed into steps, the general RAG scheme is limited by the model's own ability and the length of the context, and often has problems such as instruction following failure, missing key information, and incorrect generated result format.

[0039] When solving such complex problems, a more detailed agent workflow process needs to be designed to help the large model better understand the task goal, so as to successfully complete the step-by-step decomposition of the complex task and solve it one by one.

[0040] Please refer to Figure 1 and Figure 2 The application provides a policy text analysis method, and the specific process can be as follows S110~S150, the method comprises: S110, obtaining policy information to be analyzed.

[0041] Specifically, the policy information to be analyzed is policy text information input by a user through a terminal device and needs to be recognized and analyzed. For example, the policy information to be analyzed obtained is the similarities and differences between the reward policies of A city and B city for new energy vehicle enterprises.

[0042] S120, analyzing the policy information to be analyzed in a data enhancement recall manner to obtain retrieval recall data.

[0043] Specifically, the policy information to be analyzed is analyzed in a data enhancement recall manner to obtain retrieval recall data. The traditional RAG scheme relies on general semantic similarity calculation, which leads to the fact that the semantic correlation of special terms, abbreviations, and label words in the government field cannot be accurately identified, and the retrieval quality is insufficient. The embodiment utilizes different forms of enhanced data to improve the recall effect.

[0044] In one embodiment, the step includes: S210, retrieving and analyzing the to-be-analyzed policy information based on a preset search engine to obtain relevant text of the to-be-analyzed policy information; and / or, S220, matching and analyzing the to-be-analyzed policy information based on a preset policy dictionary to obtain paraphrase information of the to-be-analyzed policy information; and / or, S230, question segmentation of the to-be-analyzed policy information based on a preset model to obtain core phrases of the to-be-analyzed policy information; and / or, S240, determining associated information of the to-be-analyzed policy information based on label information embedded in advance in various policy information in a database, and the retrieval recall data includes the relevant text, the paraphrase information, the core phrases, and the associated information.

[0045] Specifically, the to-be-analyzed policy information is retrieved and analyzed based on a preset search engine to obtain relevant text of the to-be-analyzed policy information, as shown in Figure 2 As shown in the figure, the preset search engine adopts a three-way recall architecture of "sparse + dense + knowledge graph" to improve the recall effect.

[0046] A policy dictionary is introduced to enhance the data of the natural language explanation of the special vocabulary in the government affairs field. The policy original document and the user's question often contain special vocabulary or commonly used abbreviations in the government affairs field. Through natural language explanation and the introduction of a policy professional dictionary, the model can accurately grasp the semantic of the vocabulary and significantly improve the understanding ability of the user's non-standardized question in the query expansion and retrieval stage. The preset policy dictionary includes but is not limited to synonym / dictionary, professional vocabulary paraphrase dictionary, and abbreviation dictionary.

[0047] Synonym / dictionary: Different words with the same or similar semantics in the policy document are set as synonym / dictionary, which can make the model more accurately understand the meaning of the words and avoid missing relevant content due to different word expressions, so as to give more accurate answers. For example: "Beijing" = "capital".

[0048] Professional vocabulary paraphrase dictionary: natural language explanation of professional terms in policy documents to accurately grasp the precise semantics of the vocabulary so as to make reasoning in accordance with the true meaning rather than just staying at the literal understanding level. For example: "high-precision center" refers to a major scientific and technological innovation platform led by Beijing Municipal Government, relying on universities in Beijing area, involving multiple innovation subjects, and operating as a relatively independent entity.

[0049] Abbreviation dictionary: explanation and supplement of abbreviations to make the large model better understand the abbreviations in the policy.

[0050] The policy information to be analyzed is matched and parsed based on a preset policy dictionary to obtain interpretation information of the policy information to be analyzed. The interpretation information is an explanation of the words in the policy information to be analyzed based on the preset policy dictionary. The introduction of professional knowledge can effectively avoid misinterpretation of obscure terms and unclear expressions in the latest policy data by the large model, and avoid the problem that the pre-trained large model has not been exposed to the latest data.

[0051] To effectively locate the key phrases in the user's problem and filter out the text blocks with incomplete references, the query is finely segmented. The policy information to be analyzed is segmented based on a preset model to obtain the core phrases of the policy information to be analyzed. The problem segmentation method is used to optimize the data enhancement method for extracting key words in the government field.

[0052] The preset model can be a phrase dictionary suitable for the problem field of the policy information to be analyzed. The phrases in the phrase dictionary are assigned weights, and the segmentation method of the policy information to be analyzed with the highest total score is determined by dynamic programming or greedy method.

[0053] The preset model can model the problem segmentation of the policy information to be analyzed as a supervised machine learning problem, which is a binary classification problem for characters. For each character, it is determined whether to continue the current segmentation or start a new segmentation. By manually annotating or collecting a set of problems with correct segmentation, a binary classifier can be trained to solve the problem.

[0054] For example, the user asks "Are college teachers eligible for new energy vehicle purchase subsidies?", and the core phrases "new energy vehicle purchase subsidies" and "college teachers" can be identified after segmentation to ensure that key information is not missed during retrieval.

[0055] In addition to having unique interpretations, government domain vocabulary often has specific attribute labels. For example, the term "business environment" implies that the target is "enterprise", so "enterprise" can be used as a label for the term "business environment". Text pairs with the same label have special relevance in the government domain semantics, and corresponding weights can be set in the recall algorithm to increase the relevance score of text pairs with the same label. Therefore, the database embeds the corresponding label information in advance, matches the label information with the policy information to be analyzed, and determines the relevant information in the database related to the policy information to be analyzed. Using the label embedding method, the data enhancement method for optimizing the recall effect of the government domain text block is optimized.

[0056] For specific scenarios, user questions can be summarized into a few core labels. At this time, the label embedding of documents and vocabulary can be abstracted as a multi-label classification problem, which can be completed by a pre-trained multi-label classifier.

[0057] For example, the query "measures to optimize the business environment" can automatically recall the clauses "enterprise burden reduction" and "simplified policy approval" even if these texts do not explicitly contain the word "business environment".

[0058] Through the above data enhancement recall method, all retrieved recall data obtained are placed in the evidence pool for subsequent analysis of intelligent agents to call and parse, so that the problems and document blocks in the policy information to be analyzed can carry more context information in the government affairs field, and the precision of the whole link of recall, sorting and question answering is improved.

[0059] In this embodiment, not only unstructured policy texts are relied on, but also structured knowledge (policy knowledge graph, policy professional dictionary) is combined. The knowledge graph provides structured support for entities, clauses and relationships, avoiding reasoning omissions. The policy professional dictionary provides synonyms, abbreviations and term expansion, improving the understanding ability of intelligent agents for non-standardized expressions in the government affairs field. The double-wheel driving mechanism significantly reduces reasoning errors caused by term ambiguity and improves the professional depth of the system in policy interpretation and analysis.

[0060] In one embodiment, the preset retrieval engine includes a sparse retrieval layer, a dense retrieval layer, and a knowledge graph layer, and the related text includes a first type of text, a second type of text, and a third type of text. The step S210 of retrieving and analyzing the policy information to be analyzed based on the preset retrieval engine includes: S310, matching key information of the policy information to be analyzed based on the sparse retrieval layer, recalling the first type of text related to the policy information to be analyzed; S320, based on the dense retrieval layer, performing semantic understanding on the policy information to be analyzed, recalling the second type of text related to the semantic of the policy information to be analyzed; S330, based on the knowledge graph layer, analyzing the third type of text related to the policy information to be analyzed.

[0061] Specifically, as shown in Figure 2 In terms of hybrid retrieval framework, a "sparse + dense + knowledge enhancement" three-way recall architecture is adopted, i.e., the preset retrieval engine includes a sparse retrieval layer, a dense retrieval layer, and a knowledge graph layer. The sparse retrieval layer is based on the BM25 (Best Match 25) algorithm and performs well in precise matching scenarios (such as keyword and clause number retrieval) combined with policy corpus. The dense retrieval layer is based on the bge-m3 (BAAI General Embedding-M3, BAAI General Embedding Model M3) vector model, which has multi-language, multi-task and multi-granularity representation capabilities, and can deeply understand the complex semantics in policy question answering combined with the preset vector database to recall the policy content that does not match in literal but is related in semantics.

[0062] Knowledge graph layer: combined with the policy knowledge graph, discover the implicit associated policies in the text, such as the implicit semantic connection between "business environment" and "enterprise support". For example, the user query "Beijing's support policy for new energy vehicle enterprises", even if the word "support" does not appear in the policy text, the system can recall relevant documents containing "subsidy", "reward", "encourage investment" through dense retrieval and knowledge enhancement.

[0063] S130, task decomposition of the policy information to be analyzed is performed to obtain a plurality of sub-tasks.

[0064] Specifically, the characteristic information of the policy information to be analyzed is analyzed in combination with retrieval recall data, and the policy information to be analyzed is decomposed into a plurality of sub-tasks through the characteristic information, that is, a complex policy problem is decomposed into a plurality of sub-problems and solved one by one. For example, for the cross-dimension and cross-task characteristics of the policy problem of the policy information to be analyzed, the complex problem of the policy information to be analyzed is decomposed into a plurality of sub-tasks, such as clause extraction, cross-regional comparison, time verification, and compliance review.

[0065] For example, facing the user question "What are the similarities and differences between Beijing and Shanghai in the tax preferential policy for science and technology enterprises" in the policy information to be analyzed: sub-task 1: retrieve the tax preferential policy for science and technology enterprises in Beijing; sub-task 2: retrieve the tax preferential policy for science and technology enterprises in Shanghai; sub-task 3: analyze the common points of the policies of the two places; sub-task 4: analyze the differences between the two policies.

[0066] The user proposes "Compare the differences between A province and B province in new energy automobile subsidy policy, and judge whether there are any clauses conflicting with the central policy". The system decomposes it into three sub-tasks: 1) extract the subsidy amount, applicable object, and effective period; 2) compare the clauses across provinces; 3) conduct compliance review.

[0067] In one embodiment, the step further comprises: 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 that the intent fence corresponding to the preset intent scenario is the intent fence of the intent information.

[0068] Specifically, for typical scenarios of policy analysis, a plurality of "intent fences" are preset, and the intent fence is a set of technologies or rules used to constrain or guide the behavior boundary of the AI system, to ensure that the generated content meets the preset goals, values or safety requirements, and to prevent deviation from the core task or harmful output. A plurality of intent fences corresponding to intent scenarios are preset, such as a main region horizontal comparison intent fence, a policy definition search intent fence, a policy time evolution analysis intent fence, a policy summary intent fence, a policy similarity and difference analysis intent fence, and a policy existence verification intent fence.

[0069] When the user raises a complex question, the system first performs mixed intent recognition on the policy information to be analyzed based on a preset intent recognition model to determine intent information, and then matches the intent information based on the intent scene to schedule the task of the policy information to be analyzed to the corresponding fence workflow, ensuring that the large model "maintains boundaries" under specific tasks and avoids deviating from the answer.

[0070] First, the user question in the policy information to be analyzed is converted into a standard form, including intent understanding, follow-up expansion, and other links. Then, according to the standard form and the result of intent understanding, a preset or newly generated intent fence is matched. According to the intent fence process, the large model is planned and instructed to perform 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 the talent introduction policies in different regions", the task will be sent to the "horizontal comparison intent fence between main regions" to avoid the large model from misanswering the interpretation of a single policy. The workflow of the intent fence is shown in Figure 3 The user question in the policy information to be analyzed is identified, and it is determined whether the current question is a follow-up question. If it is, the intent understanding and rewriting are performed to maintain continuity and consistency in multiple rounds of dialogue. Then, the user intent is identified based on a preset intent recognition model. The user intent, i.e., the intent information, includes but is not limited to tool invocation intent, political involvement intent, privacy involvement intent, and business question and answer intent. After determining the user intent, the subsequent process is configured according to the intent. In addition, sensitive word detection is performed at the same time as intent recognition. If a sensitive word is hit, it can further assist in intent recognition. If no sensitive word is hit, the knowledge base is retrieved for post-processing, and then the retrieved question and answer questions and recommended topics are integrated to guide the user to continue asking questions.

[0071] S140, determining an analysis agent matched with the plurality of subtasks, and obtaining subtask candidate results of the plurality of subtasks by parallel analyzing the retrieval recall data through the analysis agent.

[0072] Specifically, a plurality of different types of analysis agents are preset to process different types of subtasks, for example: an extraction agent is responsible for structured information extraction and focuses on extracting amounts and conditions in clauses. A comparison agent is responsible for comparing differences in clauses in different regions and generating a difference table. A compliance agent is responsible for verifying compliance from the perspectives of time and level. A debate agent is responsible for raising questions when potential conflicts are found. An arbitration agent is responsible for finally summarizing and ruling on multiple opinions.

[0073] Therefore, based on the sub-tasks, the analysis agents matched with the sub-tasks are determined, and each analysis agent analyzes the retrieved recall data to obtain a sub-task candidate result of the sub-tasks in parallel. Unlike the traditional pipeline, the multiple agents can run in parallel in the same time window, greatly shortening the system response delay. In the policy scenario, this parallel execution is particularly suitable for complex requests involving multiple dimension analysis. For example, a user inquires, “Please check whether my company can enjoy both the local government's science and technology innovation subsidy and the central level's special fund support.” The system can run three agents, “eligibility condition extraction”, “local subsidy clause comparison” and “central policy compliance verification” in parallel, and output the results in a short time. The present embodiment designs a multi-task adaptation mechanism based on multi-agent collaboration to meet the needs of complex and diversified tasks in policy RAG applications. The basic idea is that multiple agents with different roles or capabilities collaborate and cross-verify in the same evidence pool, thereby improving the accuracy and robustness of complex task processing.

[0074] In one embodiment, the present step further includes: S510, if a target sub-task is added, an analysis agent corresponding to the target sub-task is added.

[0075] Specifically, the system architecture applied in the present embodiment has good scalability, and when a target sub-task is added, only a new analysis agent needs to be added without the need to reconstruct the entire system. Parallel computing is supported, which significantly reduces the response delay. In terms of interpretation, the debate and arbitration process of the multiple agents provides a clear reasoning link, and the user can not only see the result but also understand why this result is obtained.

[0076] S150, integrating the multiple sub-task candidate results to obtain an analysis result of the policy information to be analyzed.

[0077] Specifically, in formal modeling, each analysis agent outputs a sub-task candidate result based on the query and the evidence set, all sub-task candidate results enter an integration function, and common implementation methods include majority voting, weighted fusion, etc. The multiple sub-task candidate results are integrated to obtain an analysis result of the policy information to be analyzed, and the best one is selected from the multiple sub-task candidate results to avoid single model bias.

[0078] In one embodiment, the present step includes: S610, integrating the multiple sub-task candidate results to obtain analysis basis information of the multiple sub-task candidate results; and S620, selecting an analysis result of the policy information to be analyzed from the multiple sub-task candidate results based on the analysis basis information.

[0079] Specifically, the multiple sub-task candidate results are integrated, and analysis basis information of the multiple sub-task candidate results is obtained, the analysis basis information being reference information when the analysis agent derives the sub-task candidate results and reliable information of the derived sub-task candidate results, including but not limited to evidence coverage, consistency degree, and cross-task conflict situation.

[0080] Based on the analysis basis information, an analysis result of the policy information to be analyzed is selected from the multiple sub-task candidate results, for example, the analysis basis information is used as a weighting factor to weight the multiple sub-task candidate results to obtain the analysis result of the policy information to be analyzed. For example, an extraction agent identifies that "A province subsidy amount is 5000 yuan, and B province subsidy amount is 3000 yuan", and another extraction agent extracts that "B province subsidy amount is 3500 yuan". An integration function selects "3000 yuan" with higher support as the final result by comparing evidence coverage and consistency.

[0081] In one embodiment, step S620, based on the analysis basis information, selecting an analysis result of the policy information to be analyzed from the multiple sub-task candidate results, includes: S710, identifying conflict results in the multiple sub-task candidate results by a debate agent; S720, determining a reserved result in the conflict results based on the analysis basis information by an arbitration agent; and S730, based on the reserved result and non-conflict results in the multiple sub-task candidate results, obtaining the analysis result of the policy information to be analyzed.

[0082] Specifically, as shown in Figure 2 The cooperation of multiple analysis agents not only includes division of labor, but also includes "dialogue" and "debate". When there are contradictory results between analysis agents, a debate agent identifies conflict results in the multiple sub-task candidate results, actively points out the conflict, and attaches relevant evidence fragments for labeling, and submits to an arbitration agent. The arbitration agent compares the publishing time, hierarchical effectiveness and text authority of the evidence, and finally determines the reserved result in the conflict results.

[0083] For example, an extraction agent extracts "B province subsidy policy is valid until December 2025", and another agent retrieves "has been abolished in June 2025". The debate agent submits the conflict to arbitration, and the arbitration agent finally outputs "B province policy has been abolished" as the reserved result by judging that the latter is the latest document.

[0084] The embodiment also provides a product architecture of a policy text analysis device, which is used to implement the policy text analysis method described in the above embodiment. The product architecture mainly includes four layers.

[0085] Perception layer: responsible for receiving multi-modal input of policy information to be analyzed (text, table, image, 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.

[0086] Cognition layer: fusion of long-term memory (vector database, storing semantic representation of policy documents) and short-term memory (dialogue context tracking), maintaining continuity and consistency in multi-round dialogue.

[0087] Decision layer: based on the intent recognition result, i.e., the intent information of the policy information to be analyzed, the complex task is decomposed into several sub-tasks, and the planning path is dynamically adjusted.

[0088] Execution layer: calling external tools (retrieval engine, knowledge graph, calculation engine, etc.), and integrating multiple source results into the final output of the analysis result.

[0089] As shown in Figure 4 , the embodiment also provides a multi-agent collaboration process. The input is policy information to be analyzed, the intent information obtained by analyzing the policy information to be analyzed is a complex policy problem; the complex policy problem is decomposed into several sub-tasks; different analysis agents are called to perform each sub-task in parallel to obtain candidate results of the sub-tasks, for example, an extraction agent extracts amounts / conditions / objects, a comparison agent compares differences in different regions / clauses, and a compliance agent verifies time and level compliance; multiple candidate results are collected; a debate agent detects contradictory / conflicting results; if there is a conflict, an arbitration agent compares evidence release time / level / authority, and if there is no conflict, the results are directly integrated; the final credible conclusion is the output of the analysis result of the policy information to be analyzed.

[0090] The embodiment also provides a policy question-answering scenario for the public or policy personnel. The user inputs a question, for example, "compare the similarities and differences between Beijing and Shanghai's policies on unsecured loans for small and medium-sized enterprises", the system first recalls relevant policy clauses by using the bge-m3 vector retrieval model combined with policy dictionaries and label embeddings to improve retrieval accuracy. Then, the candidate results are optimized and sorted by a rerank model to ensure high semantic matching with the user's question. On this basis, the system determines whether the user's question belongs to a fact query, a clause comparison, or a comprehensive analysis task through a hybrid intent recognition mechanism, and performs multi-step task decomposition and sub-task scheduling by an agent workflow. For example, when the question involves the comparison of different policies, the agent will automatically decompose it into three sub-tasks: "retrieve Beijing policy", "retrieve Shanghai policy", and "compare similarities and differences", and complete them in turn, finally integrating the results into an accurate and traceable answer. This embodiment can effectively improve the recall accuracy, semantic understanding, and answer explainability in the question-answering scenario, avoid biases in the answers of large models, and provide reliable policy consulting services for the public and policy personnel.

[0091] The embodiment also provides a policy full-text intelligent generation scenario. In this scenario, a user inputs a macro policy goal, such as “promoting the integration of regional digital economy and green low-carbon industry”, the system first retrieves typical clauses and structured knowledge highly relevant to the goal from a historical policy library based on a data enhancement recall mechanism, and automatically identifies policy framework elements (such as guiding ideology, applicable scope, supporting measures, and guarantee mechanisms). Subsequently, the system divides the complex “full-text generation” task into several subtasks, such as framework planning, paragraph generation, cross-paragraph consistency verification, and content integration, by combining the multi-step decomposition reasoning ability of the intelligent agent and the multi-task adaptation mechanism. Different sub-agents collaborate to complete each subtask, some sub-agents focus on fact alignment and recall result utilization, some sub-agents are responsible for paragraph generation and style unification, and others are responsible for cross-paragraph logical consistency verification. The main agent is responsible for task scheduling and result integration, ensuring that the overall generation process is consistent with the policy logic and maintains content continuity.

[0092] For example, if the previous paragraph proposes “complete regional digital transformation pilot by 2025”, and the subsequent paragraph stipulates “start pilot by 2027”, the system will identify the contradiction through the logic verification sub-agent, and adjust by calling the recall result and context planning, finally ensuring that the policy draft full-text structure is reasonable, clear, and there is no contradiction between paragraphs. This embodiment demonstrates the advantages of the present application in the combination of large model generation and intelligent agent collaboration, realizing complete support from clause-level knowledge utilization to full-text policy draft assistance, and providing intelligent tool support for policy making. The embodiment also provides a policy text analysis device, which can be integrated in a terminal device. For example, as shown in the figure, the policy text analysis device 900 includes: Figure 5 A data acquisition module 910 is configured to acquire policy information to be analyzed. A data retrieval module 920 is in communication connection with the data acquisition module 910, configured to analyze the policy information to be analyzed in a data enhancement recall manner to obtain retrieval recall data. A task splitting module 930 is in communication connection with the data retrieval module 920, configured to split the policy information to be analyzed into a plurality of subtasks. An intelligent analysis module 940 is in communication connection with the task splitting module 930, configured to determine an analysis intelligent agent matched with the plurality of subtasks, and obtain subtask candidate results of the plurality of subtasks by parallel analyzing the retrieval recall data through the analysis intelligent agent. A result integration module 950 is in communication connection with the intelligent analysis module 940, configured to integrate the plurality of subtask candidate results to obtain an analysis result of the policy information to be analyzed. ​

[0093] In some embodiments, the data retrieval module 920 is further configured to perform retrieval 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, match and analyze the policy information to be analyzed based on a preset policy dictionary to obtain paraphrase information of the policy information to be analyzed; and / or, perform question segmentation on the policy information to be analyzed based on a preset model to obtain core phrases of the policy information to be analyzed; and / or, determine associated information of the policy information to be analyzed based on label information embedded in advance in various policy information in a database, wherein the retrieval recall data includes the relevant text, the paraphrase information, the core phrases, and the associated information.

[0094] 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 first text, second text, and third text, and the data retrieval module 920 is further configured to match key information of the policy information to be analyzed based on the sparse retrieval layer to recall the first 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 to recall the second text semantically related to the policy information to be analyzed; and analyze the third text related to the policy information to be analyzed based on the knowledge graph layer.

[0095] In some embodiments, the task splitting module 930 is further configured to identify 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, determine that an intent fence corresponding to the preset intent scenario is an intent fence of the intent information; and the result integration module 950 is further configured to analyze the retrieval recall data in parallel through the analysis agent based on the intent fence to obtain sub-task candidate results of the multiple sub-tasks.

[0096] In some embodiments, the result integration module 950 is further configured to integrate the multiple sub-task candidate results to obtain analysis basis information of the multiple sub-task candidate results; and select an analysis result of the policy information to be analyzed from the multiple sub-task candidate results based on the analysis basis information.

[0097] In some embodiments, the result integration module 950 is further configured to identify conflicting results in the multiple sub-task candidate results through a debate agent; determine a reserved result in the conflicting results based on the analysis basis information through an arbitration agent; and obtain an analysis result of the policy information to be analyzed based on the reserved result and non-conflicting results in the multiple sub-task candidate results.

[0098] In some embodiments, the intelligent analysis module 940 is further configured to add an analysis intelligent agent corresponding to the target subtask if the target subtask is added.

[0099] In the above embodiments, the description of each embodiment has its own focus, and the parts not described in detail in a certain embodiment can be referred to the relevant description of other embodiments.

[0100] It should be noted that: the account risk detection device provided in the above embodiments is only exemplified by the division of the above functional modules, and in actual application, the above functions can be completed by different functional modules according to needs, that is, the internal structure of the computer device is 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 provided in the above embodiments belong to the same concept, and the specific implementation process is described in the method embodiments, which will not be repeated here.

[0101] The embodiments of the present application also provide a computer device, which comprises a processor and a memory, and the memory stores at least one computer program, which is loaded and executed by the processor to implement the operations performed in the account risk detection method of the above embodiments.

[0102] Optionally, the computer device is provided as a terminal. Figure 6 The structure schematic diagram of the terminal 700 provided by an example embodiment of the present application is shown.

[0103] The terminal 700 comprises a processor 701 and a memory 702.

[0104] The processor 701 can include one or more processing cores, such as a 4-core processor, an 8-core processor, and the like. The processor 701 can be implemented in at least one of a hardware form of a DSP (Digital Signal Processing), an FPGA (Field-Programmable Gate Array), a PLA (Programmable Logic Array). The processor 701 can also include a main processor and a coprocessor, the main processor being a processor for processing data in an awake state, also known as a CPU (Central Processing Unit), and the coprocessor being a low-power processor for processing data in a standby state. In some embodiments, the processor 701 can be integrated with a GPU (Graphics Processing Unit) for rendering and drawing content required to be displayed by the display screen. In some embodiments, the processor 701 can further include an AI (Artificial Intelligence) processor for processing machine learning related computing operations.

[0105] The memory 702 can include one or more computer-readable storage media that can be non-transitory. The memory 702 can also include a high-speed random access memory, and a nonvolatile memory such as one or more disk storage devices, flash storage devices. In some embodiments, the non-transitory computer-readable storage medium in the memory 702 is used to store at least one computer program for being executed by the processor 701 to implement the account risk detection method provided by the method embodiments in the present application.

[0106] In some embodiments, the terminal 700 can also optionally include a peripheral device interface 703 and at least one peripheral device. The processor 701, the memory 702, and the peripheral device interface 703 can be connected through a bus or a signal line. Each peripheral device can be connected to the peripheral device interface 703 through a bus, a signal line, or a circuit board. Optionally, the peripheral device includes at least one of a radio frequency circuit 704, a display screen 705, a camera component 706, and an audio circuit 707.

[0107] The peripheral interface 703 can be used to connect at least one I / O (Input / Output) related peripheral device to the processor 701 and the memory 702. In some embodiments, the processor 701, the memory 702 and the peripheral interface 703 are integrated on the same chip or circuit board; in some other embodiments, any one or two of the processor 701, the memory 702 and the peripheral interface 703 can be implemented on a separate chip or circuit board, and the present embodiments are not limited to this.

[0108] The radio frequency circuit 704 is configured to receive and send RF (Radio Frequency) signals, also known as electromagnetic signals. The radio frequency circuit 704 communicates with communication networks and other communication devices through electromagnetic signals. The radio frequency circuit 704 converts electrical signals to electromagnetic signals for transmission, or converts electromagnetic signals received to electrical signals. Optionally, the radio frequency 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 subscriber identity module card, and the like. The radio frequency circuit 704 can communicate with other devices through at least one wireless communication protocol. The wireless communication protocol includes, but is not limited to, a metropolitan area network, various generations of mobile communication networks (2G, 3G, 4G and 5G), a wireless local area network and / or a WiFi (Wireless Fidelity) network. In some embodiments, the radio frequency circuit 704 can also include NFC (Near Field Communication) related circuitry, and the present application is not limited to this.

[0109] The display screen 705 is configured to display a UI (User Interface). The UI can include graphics, text, icons, video, and any combination thereof. When the display screen 705 is a touch display screen, the display screen 705 is also capable of capturing touch signals on or above the surface of the display screen 705. The touch signals can be input to the processor 701 as control signals for processing. At this time, the display screen 705 can also be configured to provide virtual buttons and / or virtual keyboard, also known as soft buttons and / or soft keyboard. In some embodiments, the display screen 705 can be one, arranged on the front panel of the terminal 700; in other embodiments, the display screen 705 can be at least two, arranged on different surfaces of the terminal 700 or in a folding design; in other embodiments, the display screen 705 can be a flexible display screen, arranged on a curved surface or a folding surface of the terminal 700. Even, the display screen 705 can also be arranged in an irregular shape, i.e. a special-shaped screen. The display screen 705 can be made of LCD (Liquid Crystal Display), OLED (Organic Light-Emitting Diode), etc.

[0110] The camera assembly 706 is configured to capture images or videos. Optionally, the camera assembly 706 includes a front camera and a rear camera. The front camera is arranged on the front panel of the terminal 700, and the rear camera is arranged on the back of the terminal 700. In some embodiments, the rear camera is at least two, which are any one of a main camera, a depth-of-field camera, a wide-angle camera, and a telephoto camera, to realize the background blur function by fusing the main camera and the depth-of-field camera, the panoramic shooting and VR (Virtual Reality) shooting function by fusing the main camera and the wide-angle camera, or other fusion shooting functions. In some embodiments, the camera assembly 706 can also include a flash. The flash can be a single-color temperature flash or a dual-color temperature flash. The dual-color temperature flash refers to the combination of a warm light flash and a cold light flash, which can be used for light compensation under different color temperatures.

[0111] The audio circuit 707 can include a microphone and a speaker. The microphone is used to collect sound waves of a user and an environment, and convert the sound waves into an electrical signal input to the processor 701 for processing, or input to the radio frequency circuit 704 to realize voice communication. For the purpose of stereo sound collection or noise reduction, the microphone can be multiple, and arranged at different parts of the terminal 700. The microphone can also be an array microphone or an omnidirectional collection microphone. The speaker is used to convert an electrical signal from the processor 701 or the radio frequency circuit 704 into sound waves. The speaker can be a conventional diaphragm speaker, or a piezoelectric ceramic speaker. When the speaker is a piezoelectric ceramic speaker, not only can the electrical signal be converted into a sound wave audible to humans, but also can be converted into a sound wave inaudible to humans for ranging purposes. In some embodiments, the audio circuit 707 can also include a headphone jack.

[0112] Those skilled in the art can understand that the structure shown in the above embodiments is not a limitation on the terminal 700, and the terminal 700 can include more or fewer components than those shown in the figure, or combine certain components, or use a different component arrangement. Figure 6 Those skilled in the art can understand that the structure shown in the above embodiments is not a limitation on the terminal 700, and the terminal 700 can include more or fewer components than those shown in the figure, or combine certain components, or use a different component arrangement.

[0113] Optionally, the computer device is provided as a server. Figure 7 The server 800 can have a large difference due to different configurations or performances, and can include one or more processors (Central Processing Units, CPUs) 801 and one or more memories 802, wherein the memory 802 stores at least one computer program, and the at least one computer program is loaded and executed by the processor 801 to implement the method provided by each of the above methods. Of course, the server can also have a wired or wireless network interface, a keyboard, and an input and output interface, etc., so as to perform input and output, and the server can also include other components for realizing the functions of the device, which are not described here.

[0114] The embodiments of the present application also provide a computer readable storage medium, which stores at least one computer program, and the at least one computer program is loaded and executed by a processor to implement the operations performed by the account risk detection method of the above embodiments.

[0115] Those skilled in the art can understand that all or part of the steps of the above embodiments can be completed by hardware, or by a program instructing related hardware to complete, and the program can be stored in a computer readable storage medium, and the storage medium mentioned above can be a read-only memory, a magnetic disk or an optical disk, etc.

[0116] The above merely describes optional embodiments of the present application, and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A policy text analysis method, the method comprising: obtaining policy information to be analyzed; analyzing the policy information to be analyzed in a data enhancement recall manner to obtain retrieval recall data; task decomposing the policy information to be analyzed to obtain a plurality of subtasks; determining an analysis agent matched with the plurality of subtasks, and obtaining subtask candidate results of the plurality of subtasks by parallelizing the analysis agent to analyze the retrieval recall data; integrating the plurality of subtask candidate results to obtain an analysis result of the policy information to be analyzed.

2. The policy text analysis method of claim 1, wherein, The analyzing the policy information to be analyzed in a data enhancement recall manner to obtain retrieval recall data comprises: performing retrieval 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, matching and analyzing the policy information to be analyzed based on a preset policy dictionary to obtain paraphrase information of the policy information to be analyzed; and / or, performing question segmentation on the policy information to be analyzed based on a preset model to obtain core phrases of the policy information to be analyzed; and / or, determining associated information of the policy information to be analyzed based on label information embedded in advance in each type of policy information in a database, wherein the retrieval recall data comprises the relevant text, the paraphrase information, the core phrases, and the associated information.

3. The policy text analysis method of claim 2, wherein, The preset retrieval engine comprises a sparse retrieval layer, a dense retrieval layer, and a knowledge graph layer, the relevant text comprises a first type of text, a second type of text, and a third type of text, and the performing retrieval analysis on the policy information to be analyzed based on the preset retrieval engine to obtain relevant text information of the policy information to be analyzed comprises: matching key information of 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; 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; analyzing the third type of text related to the policy information to be analyzed based on the knowledge graph layer.

4. The policy text analysis method of claim 1, wherein, Before the task decomposing the policy information to be analyzed to obtain a plurality of subtasks, the method further comprises: identifying 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, determining an intent fence corresponding to the preset intent scenario as an intent fence of the intent information. The obtaining subtask candidate results of the plurality of subtasks by parallelizing the analysis agent to analyze the retrieval recall data comprises: based on the intent fence, obtaining subtask candidate results of the plurality of subtasks by parallelizing the analysis agent to analyze the retrieval recall data.

5. The policy text analysis method of claim 1, wherein, The integrating the plurality of subtask candidate results to obtain an analysis result of the policy information to be analyzed comprises: integrating the plurality of subtask candidate results to obtain analysis basis information of the plurality of subtask candidate results; based on the analysis basis information, selecting an analysis result of the policy information to be analyzed from the plurality of subtask candidate results.

6. The policy text analysis method of claim 5, wherein, The analysis result of the policy information to be analyzed is selected from the multiple sub-task candidate results based on the analysis basis information, and the analysis result of the policy information to be analyzed is obtained, comprising: The conflict result in the multiple sub-task candidate results is identified by a debate intelligent agent; The reserved result in the conflict result is determined by an arbitration intelligent agent based on the analysis basis information; The analysis result of the policy information to be analyzed is obtained based on the reserved result and the non-conflict result in the multiple sub-task candidate results.

7. The policy text analysis method of claim 1, wherein, Before the analysis intelligent agent matched with the multiple sub-tasks is determined, the method further comprises: If a target sub-task is added, an analysis intelligent agent corresponding to the target sub-task is added.

8. A policy text analysis apparatus characterized by comprising: The device comprises: A data acquisition module is configured to acquire policy information to be analyzed; A data retrieval module is in communication connection with the data acquisition module and is configured to analyze the policy information to be analyzed by using a data enhancement recall mode to obtain retrieval recall data; A task splitting module is in communication connection with the data retrieval module and is configured to split the policy information to be analyzed into multiple sub-tasks; An intelligent analysis module is in communication connection with the task splitting module and is configured to determine an analysis intelligent agent matched with the multiple sub-tasks, and to obtain sub-task candidate results of the multiple sub-tasks by parallel analysis of the retrieval recall data by the analysis intelligent agent; A result integration module is in communication connection with the intelligent analysis module and is configured to integrate the multiple sub-task candidate results to obtain an analysis result of the policy information to be analyzed.

9. A computer device, comprising: The computer device comprises a processor and a memory, and the memory stores at least one computer program, which is loaded and executed by the processor to implement the operations performed by the policy text analysis method according to any one of claims 1 to 7.

10. 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 the processor to implement the operations performed by the policy text analysis method according to any one of claims 1 to 7.

Images