Report generation method and system based on multi-agent architecture

Abstract

The application provides a report generation method and system based on a multi-agent architecture, which first receives and analyzes a report generation instruction, obtains a theme requirement, a framework specification and initial reference materials, then starts a multi-agent collaborative framework, generates an agent task allocation table, a resource retrieval agent performs network resource directional retrieval according to the task allocation table, generates an associated resource set, a content extraction agent performs structured conversion on the associated resource set and the initial reference materials to obtain structured content units with chapter codes. A report integration agent classifies and logically concatenates the structured content units according to the framework specification to generate a preliminary report. Finally, the multi-agent collaborative framework performs content verification and optimization on the preliminary report to generate a final report text that meets the framework specification. Thus, the report generation can be automated, intelligentized and efficient, and the report quality can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to a report generation method and system based on a multi-agent architecture. BACKGROUND

[0002] In the traditional report generation process, a series of work from material collection, content writing to report integration is usually completed by manual operation. Manual operation is not only inefficient, but also prone to errors and omissions when dealing with a large amount of information and complex frameworks. With the development of information technology, although some automated report generation tools have appeared, these tools are mostly single-functioned and can only complete part of specific tasks, such as simple data extraction or template filling. For example, some tools can only extract data from a pre-set database and fill it into a fixed template, but cannot flexibly retrieve network resources according to the report theme requirements, nor can they perform in-depth analysis and structured processing on the collected materials. In addition, the existing tools have obvious deficiencies in handling the logical coherence and content integrity of the report, and it is difficult to generate high-quality reports that meet the professional framework specifications. SUMMARY

[0003] Therefore, the purpose of the present application is to provide a report generation method and system based on a multi-agent architecture.

[0004] According to a first aspect of the present application, a report generation method based on a multi-agent architecture is provided, which comprises:

[0005] receiving a report generation instruction, parsing the report generation instruction to obtain a theme requirement and a corresponding framework specification, obtaining initial reference materials associated with the theme requirement, and the framework specification including chapter hierarchical relationship and content module division standard;

[0006] starting a multi-agent collaborative framework, calling a resource retrieval agent, a content extraction agent and a report integration agent, generating an agent task allocation table based on the chapter hierarchical relationship in the framework specification, the agent task allocation table including the task identification, chapter code and task execution requirement corresponding to each agent;

[0007] performing a network resource directional retrieval operation according to the task identification and chapter code in the agent task allocation table through the resource retrieval agent, generating an associated resource set matching the chapter hierarchical relationship, calling the content extraction agent to perform content structured conversion processing on the associated resource set and the initial reference materials, and generating a structured content unit with chapter code;

[0008] The report integration agent is invoked to receive the structured content units, and the report integration agent classifies the structured content units into modules and logically concatenates the structured content units according to the content module classification criteria in the framework specification, to generate a preliminary report conforming to the chapter hierarchy;

[0009] The multi-agent collaborative framework is used to perform content verification and optimization on the preliminary report, to generate a final report text conforming to the framework specification.

[0010] According to a second aspect of the present application, a report generation system based on a multi-agent architecture is provided, the report generation system based on the multi-agent architecture comprising a machine-readable storage medium and a processor, the machine-readable storage medium storing machine-executable instructions, and the processor, when executing the machine-executable instructions, implements the report generation method based on the multi-agent architecture.

[0011] According to a third aspect of the present application, a computer-readable storage medium is provided, the computer-readable storage medium storing computer-executable instructions, and when the computer-executable instructions are executed, the report generation method based on the multi-agent architecture is implemented.

[0012] According to any one of the above aspects, the technical effects of the present application are as follows:

[0013] By receiving a report generation instruction and parsing the subject requirement and the framework specification, an explicit target and structural guidance are provided for report generation, a multi-agent collaborative framework is started, different types of agents are invoked and an agent task allocation table is generated, reasonable allocation of tasks and orderly cooperation between agents are achieved, the resource retrieval agent performs directional retrieval of network resources according to the task allocation table, the content extraction agent performs structured conversion on the retrieved resources and the initial reference materials to generate structured content units with chapter codes, the content specification and processability are improved, the report integration agent classifies the structured content units into modules and logically concatenates the structured content units according to the framework specification, to generate a preliminary report, ensuring the completeness and logicality of the report, finally, the multi-agent collaborative framework performs content verification and optimization on the preliminary report, further improving the quality of the report, so that the report generation process is automated, intelligentized and efficient, a high-quality report conforming to professional requirements can be generated, and the human and time costs are greatly saved. BRIEF DESCRIPTION OF DRAWINGS

[0014] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments. It should be understood that the following drawings only show some of the embodiments of the present application, and therefore should not be regarded as limiting the scope. For those skilled in the art, other related drawings can also be obtained without creative labor.

[0015] Figure 1 The flowchart of the report generation method based on the multi-agent architecture provided by the embodiments of the present application is shown.

[0016] Figure 2 The component structure diagram of the report generation system based on the multi-agent architecture for implementing the report generation method based on the multi-agent architecture described above is shown. DETAILED DESCRIPTION

[0017] The embodiments of the present application will be described below in conjunction with the drawings in the present application. It should be understood that the embodiments described below in conjunction with the drawings are exemplary descriptions for explaining the technical solutions of the embodiments of the present application, and do not limit the technical solutions of the embodiments of the present application.

[0018] Those skilled in the art can understand that, unless specifically stated, the singular forms "a", "an" and "the" used herein also include the plural forms. It should be further understood that the terms "comprise" and "include" used by the embodiments of the present application mean that the corresponding features can be implemented as the presented features, information, data, steps, operations, elements and / or components, but do not exclude other features, information, data, steps, operations, elements, components and / or their combinations supported by the present technical field. It should be understood that when it is said that one element is "connected" or "coupled" to another element, the one element can be directly connected or coupled to the other element, or it can mean that the one element and the other element are connected through an intermediate element. In addition, the "connection" or "coupling" used herein can include wireless connection or wireless coupling. The term "and / or" used herein means that at least one of the terms defined by the term, for example, "A and / or B" can be implemented as "A", or as "B", or as "A and B".

[0019] In order to make the purposes, technical solutions and advantages of the present application more clear, the embodiments of the present application will be further described in detail below in conjunction with the drawings. The technical solutions of the embodiments of the present application and the technical effects generated by the technical solutions of the present application will be described below through the description of several exemplary embodiments. It should be pointed out that the following embodiments can be mutually referenced, borrowed or combined, and the same terms, similar features and similar implementation steps in different embodiments will not be described repeatedly.

[0020] Figure 1 The flowchart of the report generation method and system based on the multi-agent architecture provided by the embodiments of the present application is shown. It should be understood that in other embodiments, the order of some steps of the report generation method based on the multi-agent architecture of the present embodiment can be shared according to actual needs, or some steps can be omitted or maintained. The detailed steps of the report generation method based on the multi-agent architecture include:

[0021] Step S110: receiving a report generation instruction, parsing the report generation instruction to obtain a topic requirement and a corresponding framework specification, obtaining initial reference materials associated with the topic requirement, and the framework specification including chapter hierarchical relationship and content module division standard.

[0022] In the medical scientific research digital academic scene, when researchers need to generate an academic review on the mechanism and progress of diagnosis and treatment of a specific disease, a report generation instruction can be triggered. After receiving the report generation instruction, it is parsed to determine the topic requirement, such as "circRNA mechanism research progress in Alzheimer's disease". At the same time, the corresponding framework specification is parsed, which is formulated according to the general paradigm of academic review and the professional requirements of the field.

[0023] The chapter hierarchical relationship in the framework specification clearly defines the structural hierarchy of the report, from the main chapter to the sub-chapter, forming a clear hierarchical system. For example, the main chapter may include "disease molecular mechanism overview", "circRNA related research progress", "experimental techniques and methods", "potential clinical applications", "future research directions", etc.; each main chapter has corresponding sub-chapters, such as "circRNA related research progress" may have "circRNA biosynthesis and function", "circRNA and neurodegenerative diseases" and the like; sub-chapters may also have more detailed sub-chapters.

[0024] The content module division standard specifies the content category that each module should cover, such as the "disease molecular mechanism overview" module needs to include the molecular pathological mechanism of the disease, related signal pathways, genetic factors, etc.; the "circRNA related research progress" module needs to include the biological characteristics of circRNA, the latest experimental data, functional verification, etc.

[0025] After obtaining the topic requirement, initial reference materials related to the topic can be obtained. These initial reference materials may include related experimental data sets, published related academic papers, authoritative research reports, etc., which will serve as the basic materials for report generation.

[0026] Step S120: Start the multi-agent collaborative framework, call the resource retrieval agent, the content extraction agent and the report integration agent, generate an agent task allocation table based on the chapter level relationship in the framework specification, the agent task allocation table contains the task identifier corresponding to each agent, chapter code and task execution requirements.

[0027] After obtaining the topic requirements, framework specification and initial reference materials, start the multi-agent collaborative framework. The multi-agent collaborative framework can provide a collaborative working environment for each agent to ensure efficient information exchange and task collaboration between them.

[0028] Then, call the resource retrieval agent, the content extraction agent and the report integration agent. The resource retrieval agent has the ability to accurately obtain the required information from network resources; the content extraction agent can deeply process the obtained resources and initial reference materials to extract key information; the report integration agent can integrate the processed content into a complete report according to certain logic and structure.

[0029] Generate an agent task allocation table based on the chapter level relationship in the framework specification. Each node in the chapter level relationship has its specific content requirements and processing methods, so different agents need to be assigned corresponding tasks.

[0030] The task identifier in the agent task allocation table is used to uniquely identify each task to avoid confusion; the chapter code corresponds to the chapter level relationship in the framework specification to ensure accurate matching of tasks and specific chapters; the task execution requirements specify the goals to be achieved and the standards to be followed for each task, for example, the resource retrieval agent needs to retrieve authoritative academic literature within a certain period of time when retrieving resources related to a chapter.

[0031] Step S121: Analyze the framework specification to obtain the chapter level relationship, which includes main chapter nodes, sub-chapter nodes and sub-chapter nodes, each node has a unique chapter code and chapter theme description.

[0032] Analyze the framework specification to obtain the chapter level relationship, which is the basis for generating the agent task allocation table. By carefully analyzing the framework specification, the hierarchical structure between the main chapter nodes, sub-chapter nodes and sub-chapter nodes is sorted out.

[0033] Each node has a unique chapter code, for example, the code of the main chapter "Chapter 1: Mechanism of circRNA in Alzheimer's Disease" may be "M1", the code of the sub-chapter "Sub-chapter 1: Biological function of circRNA" under it is "M1-F1", the code of the sub-chapter "Sub-chapter 1: Biological function of circRNA" under it is "M1-F1-Z1", and so on. This coding method can clearly reflect the position of the node in the hierarchical relationship.

[0034] The chapter theme description of each node clearly describes the core content of the node, such as the theme description of "Chapter 2: Analysis of the current status of circRNA research methods" is "a detailed introduction and analysis of various technical means currently used for circRNA research"; the theme description of "Sub-chapter 2: Introduction of traditional molecular biology techniques" is "to describe the operation process, scope of application, etc. of the commonly used traditional molecular biology techniques".

[0035] Step S122: Extract the chapter theme description of each node in the chapter hierarchical relationship, determine the content attribute identifier corresponding to each node by content attribute determination of the chapter theme description, and the content attribute identifier includes theoretical type, data type and case type.

[0036] After extracting the chapter theme description of each node from the chapter hierarchical relationship, the content attribute of these descriptions is determined. The basis of the determination is the content type reflected by the theme description.

[0037] If the chapter theme description mainly involves the exposition of theoretical knowledge, such as "Introduction of the principle of new diagnostic techniques", its content attribute identifier is theoretical type, because this part mainly explains the related theoretical knowledge.

[0038] When the chapter theme description involves the presentation and analysis of data, such as "Comparison of diagnostic accuracy of different diagnostic methods", its content attribute identifier is data type, and this kind of content needs to be supported by data for analysis.

[0039] And the chapter theme description like "Mechanism of circRNA in Alzheimer's Disease" mainly tells the specific scientific research content, so the content attribute identifier is case type.

[0040] Through the above determination, the corresponding content attribute identifier of each node is determined, so as to assign appropriate tasks to different intelligent agents in the subsequent process.

[0041] Step S123: Based on the content attribute identifier, query the intelligent agent capability matching rule to determine the intelligent agent type corresponding to each content attribute identifier, wherein the theoretical type corresponds to the resource retrieval intelligent agent, the data type and the case type correspond to the content extraction intelligent agent, and the integration of all type nodes corresponds to the report integration intelligent agent.

[0042] After determining the content attribute identifier of each node, the corresponding intelligent agent type is matched for each content attribute identifier according to the preset intelligent agent capability matching rule.

[0043] The intelligent agent capability matching rule specifies the intelligent agent suitable for different content attribute identifiers. The theoretical type needs to retrieve relevant theoretical knowledge from various resources, so the theoretical type corresponds to the resource retrieval intelligent agent. The resource retrieval intelligent agent can efficiently retrieve theoretical content from network resources.

[0044] Data type and case type content not only need to obtain relevant materials, but also need to extract and process the materials to meet the requirements of the report, so the corresponding content extraction intelligent agent has the ability to analyze and extract key information from data and cases.

[0045] The content of all type nodes needs to be integrated into a complete and coherent report, which is the responsibility of the report integration intelligent agent, which can integrate the content processed by different intelligent agents according to certain logic and structure.

[0046] Step S124: For each chapter node, generate a task description item combining its chapter code, content attribute identifier and corresponding intelligent agent type. The task description item includes chapter code, intelligent agent type identifier, content attribute identifier and task target description.

[0047] For each chapter node, combine its chapter code, content attribute identifier and corresponding intelligent agent type to generate a task description item.

[0048] For example, a chapter node has a chapter code of "M2-F2", a content attribute identifier of data type, and a corresponding intelligent agent type of content extraction intelligent agent. The generated task description item may be: chapter code "M2-F2", intelligent agent type identifier "content extraction intelligent agent", content attribute identifier "data type", and task target description "extract and analyze the relevant data of circRNA in the mechanism research progress of Alzheimer's disease, and extract key data information".

[0049] Each task description item clearly specifies the chapter corresponding to the task, the responsible intelligent agent, the content attribute, and the goal to be achieved, providing clear guidance for the intelligent agent to execute the task.

[0050] Step S125: Sort the task description entries based on the node order in the chapter hierarchy relationship, and generate an agent task allocation table indexed by chapter codes, containing agent type identification and task target description.

[0051] According to the order of nodes in the chapter hierarchy relationship, all generated task description entries are sorted. The basis for sorting is the logical order of chapters in the report, ensuring that the execution of tasks can be carried out in sequence according to the structure of the report.

[0052] After sorting, the agent task allocation table is generated. The agent task allocation table is indexed by chapter codes, and each index corresponds to the agent type identification and task target description of the chapter.

[0053] In this way, when the agent executes the task, it can quickly find the corresponding task requirements through the chapter code, and clearly know what work it needs to complete, thereby improving the efficiency and accuracy of the entire report generation process.

[0054] Step S130: The resource retrieval agent performs network resource directional retrieval operation according to the task identification and chapter code in the agent task allocation table, generates an associated resource set matching the chapter hierarchy relationship, and calls the content extraction agent to perform content structured conversion processing on the associated resource set and the initial reference material, generating structured content units with chapter codes.

[0055] The resource retrieval agent will perform network resource directional retrieval operation according to the task identification and chapter code in the agent task allocation table. For example, for a task with chapter code "M3-F1" and task identification "searching for the mechanism of circRNA", the resource retrieval agent will perform directional retrieval of related content in medical academic databases, authoritative medical websites, etc.

[0056] Through retrieval, a large number of resources related to the chapter, such as literature and research reports, can be collected, and these resources will be organized into an associated resource set matching the chapter hierarchy relationship, each resource corresponding to a corresponding chapter code.

[0057] Then, the content extraction agent is called to process the associated resource set and the initial reference material together. The content extraction agent will analyze and extract key information from these materials, and perform structured conversion, finally generating structured content units with chapter codes, each unit corresponding to a specific chapter, preparing for subsequent report integration.

[0058] Step S131: Analyze the intelligent agent task allocation table to obtain the chapter code and task target description corresponding to the task identifier. Extract the core theme terms in the task target description. The core theme terms are key concept expressions in the chapter theme description.

[0059] Parse the intelligent agent task allocation table to obtain the chapter code and task target description corresponding to each task identifier. The task target description elaborates the specific goals that need to be achieved, such as "searching for relevant academic literature on the regulatory principles of new molecular mechanisms in the development of a rare disease".

[0060] Extract core theme terms from the task target description. These core theme terms are key concepts in the chapter theme description. For the above task target description, core theme terms may include "new molecular mechanism", "rare disease", "regulation", "principle", "academic literature", etc. These core theme terms will serve as the basis for subsequent retrieval, ensuring that the retrieved resources are highly relevant to the task target.

[0061] Step S132: Call the retrieval term expansion module of the resource retrieval intelligent agent to perform synonym expansion processing on the core theme terms to generate a retrieval term set containing core theme terms and synonyms.

[0062] Call the retrieval term expansion module of the resource retrieval intelligent agent to perform synonym expansion processing on the core theme terms. This is to expand the retrieval range and avoid missing relevant resources due to different term expressions.

[0063] For example, synonyms of "new molecular mechanism" may include "new action mechanism" and "new action mode"; synonyms of "principle" may include "mechanism" and "principle mechanism".

[0064] Through the above expansion, a retrieval term set containing core theme terms and their synonyms is generated, such as {"new molecular mechanism", "new action mechanism", "new action mode", "rare disease", "regulation", "principle", "mechanism", "principle mechanism", "academic literature"}, preparing for more comprehensive retrieval.

[0065] Step S133: Construct a retrieval request based on the retrieval term set and chapter code. The retrieval request contains the retrieval term set, chapter code, and content attribute identifier. Submit the retrieval request to the resource database through the preset network resource interface.

[0066] Construct a retrieval request using the retrieval term set and corresponding chapter code. In addition to the retrieval term set and chapter code, the retrieval request also includes content attribute identifiers, such as theoretical type.

[0067] For example, for the chapter code "M3-F1", the retrieval word set is the above-mentioned expanded set, the content attribute identifier is theoretical type, and the constructed retrieval request contains these information.

[0068] After that, through the preset network resource interface, the retrieval request is submitted to the resource database, such as the professional academic database in the field of medical scientific research, etc., to obtain related resources.

[0069] Step S134: receiving the candidate resource list returned by the resource database, performing metadata extraction on the resource entries in the candidate resource list, and obtaining resource titles, publication times, source identifiers and content abstracts.

[0070] After receiving the retrieval request, the resource database can return a candidate resource list, which contains a large number of resource entries related to the retrieval conditions.

[0071] Metadata extraction is performed on these resource entries. The extracted information includes resource titles such as "Exploration of a new molecular mechanism in the regulation of a rare disease", publication times such as a specific year, source identifiers such as the resource coming from which academic journal, and content abstracts that briefly summarize the core content of the resource, such as the research exploring a new molecular mechanism regulating the development of a rare disease through phenotype regulation.

[0072] Step S135: performing semantic matching of the resource titles and content abstracts with the task target description, screening resource entries whose semantic matching results meet preset conditions, storing the screened resource entries according to chapter codes, and generating an associated resource set matching the chapter hierarchical relationship.

[0073] The extracted resource titles and content abstracts are semantically matched with the task target description. This process analyzes the relevance between the titles and abstracts and the task target, and determines whether the resource meets the requirements of the task.

[0074] For example, for a task whose task target is to search for the research progress of new circRNA in Alzheimer's disease, it can be determined whether the resource title and abstract focus on this principle. Through semantic matching, resource entries whose matching results meet the preset conditions are screened, and these entries have higher relevance to the task target.

[0075] Then, the screened resource entries are stored according to their corresponding chapter codes, thereby generating an associated resource set matching the chapter hierarchical relationship. Each chapter code has corresponding resource entries, which facilitates subsequent content refining processing.

[0076] Step S136: extracting the theme keyword sequence in the task target description, the theme keyword sequence being composed of core theme terms and feature words corresponding to content attribute identifiers.

[0077] In the process of semantic matching of the resource entry, the subject keyword sequence in the task target description is first extracted. The subject keyword sequence includes not only the previously determined core subject term, but also the feature words corresponding to the content attribute identifier.

[0078] If the content attribute identifier is theoretical, the corresponding feature words can be "principle", "mechanism", "theoretical basis", etc.; the corresponding feature words of data type can be "data", "statistics", "numerical results", etc.; the corresponding feature words of case type can be "case", "instance", "specific case", etc.

[0079] For example, the content attribute identifier corresponding to the task target description is theoretical, and the core subject term is "new type of circRNA", "Alzheimer's disease", "regulation", "principle". The subject keyword sequence can be "new type of circRNA, Alzheimer's disease, regulation, principle, mechanism, theoretical basis".

[0080] Step S137: Extract the title keyword sequence in the resource title and the abstract keyword sequence in the content abstract. The title keyword sequence and the abstract keyword sequence are obtained by word segmentation processing and stop word filtering.

[0081] The resource title is subjected to word segmentation processing, and the title is split into individual words, and then the stop words (such as words without actual meaning such as "de", "in", "in", etc.) are filtered out, to obtain the title keyword sequence.

[0082] Similarly, the content abstract is also subjected to similar processing, and the stop words are filtered after word segmentation, to obtain the abstract keyword sequence.

[0083] For example, the resource title is "Exploration of the mechanism of action of a new type of molecule in the occurrence and development of a rare disease", and after processing, the title keyword sequence can be "new type, molecule, rare disease, occurrence and development, mechanism of action, exploration". The content abstract is also processed to obtain the corresponding abstract keyword sequence.

[0084] Step S138: Calculate the number of overlapping words between the subject keyword sequence and the title keyword sequence, denoted as the title matching number, and calculate the number of overlapping words between the subject keyword sequence and the abstract keyword sequence, denoted as the abstract matching number.

[0085] The subject keyword sequence and the title keyword sequence are compared, and the number of overlapping words between the two is counted. The number of words is the title matching number.

[0086] Similarly, the subject keyword sequence and the abstract keyword sequence are compared, and the number of overlapping words is counted to obtain the abstract matching number.

[0087] For example, the subject keyword sequence is "new molecular marker, a rare disease, mechanism research, signal pathway, theoretical basis", the title keyword sequence is "new, molecular mechanism, a rare disease, mechanism research, principle exploration", and the overlapping words of the two may be "new", "a rare disease", and "mechanism research". The title matching number is the number of these overlapping words. The abstract keyword sequence and the subject keyword sequence are compared to obtain the corresponding abstract matching number.

[0088] Step S139: The task target description, resource title, and content abstract are respectively subjected to semantic vector conversion to obtain a target semantic vector, a title semantic vector, and an abstract semantic vector. The cosine similarity between the target semantic vector and the title semantic vector is calculated as a title similarity value, and the cosine similarity between the target semantic vector and the abstract semantic vector is calculated as an abstract similarity value.

[0089] The task target description is subjected to semantic vector conversion to convert it into a multi-dimensional target semantic vector, which can represent the content of the task target description from the semantic level.

[0090] The resource title and the content abstract are respectively converted into a title semantic vector and an abstract semantic vector by the same method. Both of the vectors are multi-dimensional and represent the semantics of the title and the abstract, respectively.

[0091] Then, the cosine similarity between the target semantic vector and the title semantic vector is calculated. The cosine similarity value reflects the degree of association between the two in terms of semantics, and is taken as the title similarity value.

[0092] Similarly, the cosine similarity between the target semantic vector and the abstract semantic vector is calculated to obtain the abstract similarity value.

[0093] Step S1310: Determine a comprehensive matching index based on the title matching number, the abstract matching number, the title similarity value, and the abstract similarity value. The comprehensive matching index is obtained by multiplying the sum of the title matching number and the abstract matching number by the average of the title similarity value and the abstract similarity value. The resource entries with a comprehensive matching index greater than a preset threshold are screened.

[0094] The title matching number, the abstract matching number, the title similarity value, and the abstract similarity value are comprehensively considered to determine the comprehensive matching index. The calculation method is to first add the title matching number and the abstract matching number to obtain a sum, then calculate the average of the title similarity value and the abstract similarity value, and finally multiply the sum by the average to obtain the comprehensive matching index.

[0095] The comprehensive matching index calculated in this way can comprehensively reflect the matching degree of the resource title and abstract and the task target description. Then, the resource entries with a comprehensive matching index greater than a preset threshold are screened out, and these resource entries have a higher relevance to the task target and can better meet the demand of report generation.

[0096] Step S1311: merging the resource entry content corresponding to each chapter code in the associated resource set and the material content of the corresponding chapter theme in the initial reference material to obtain a comprehensive content pool classified by chapter code.

[0097] The resource entry content corresponding to each chapter code in the associated resource set is summarized, and the material content corresponding to each chapter theme is screened out from the initial reference material. For example, for chapter code "M3-F1", the associated resource set may contain multiple literature contents about the principle of new diagnostic techniques, and the initial reference material may have related technical notes and conference records sorted by clinical doctors, and the like, and these contents are merged together.

[0098] After the merging process, a comprehensive content pool classified by chapter code is formed. Each chapter code corresponds to an independent content pool, and the content pool contains all the associated resource content and initial reference material content required for the chapter.

[0099] Step S1312: calling the paragraph division module of the content extraction agent to perform paragraph boundary identification processing on the text content in the comprehensive content pool to obtain multiple independent paragraphs, each independent paragraph containing a paragraph start position identifier and a paragraph end position identifier.

[0100] The paragraph division module of the content extraction agent is called, and the paragraph division module has the ability to identify the paragraph boundary of the text content. The text content in the comprehensive content pool is analyzed word by word and sentence by sentence, and according to the features such as punctuation marks (such as period, question mark, exclamation mark, etc.), line feed, and semantic pause, the boundary of the paragraph is identified.

[0101] After processing, the text content is divided into multiple independent paragraphs. Each independent paragraph is assigned a paragraph start position identifier and a paragraph end position identifier, which are usually represented by the index position of the character in the entire text. For example, the start position identifier of a paragraph is the number of characters in the text, and the end position identifier is the number of characters, so that each paragraph can be accurately located in the text, and subsequent processing of the paragraph can be facilitated.

[0102] Step S1313: Perform core information extraction processing on each independent paragraph to identify opinion expression sentences, argument support sentences, and case reference sentences in the independent paragraph. An opinion expression sentence is a sentence containing a conclusive expression, an argument support sentence is a sentence containing a data or fact description, and a case reference sentence is a sentence containing a specific example description.

[0103] For each independent paragraph, core information extraction processing is performed. This process requires detailed analysis of each sentence in the paragraph to identify opinion expression sentences, argument support sentences, and case reference sentences.

[0104] An opinion expression sentence is a sentence containing a conclusive expression, usually expressing a judgment, claim, or insight. For example, in a paragraph about a new diagnostic technology, "the new diagnostic technology has higher specificity than traditional technology" is an opinion expression sentence that clearly expresses a conclusive judgment about the technology.

[0105] An argument support sentence is a sentence containing a data or fact description, used to support an opinion expression sentence. For example, "in a previous test, the technology achieved a certain level of detection accuracy for a certain type of sample" is an argument support sentence that provides support for the opinion by describing data.

[0106] A case reference sentence is a sentence containing a specific example description, explaining or supporting related content through specific cases. For example, "a hospital used the new technology to quickly and accurately diagnose a suspected patient" is a case reference sentence that illustrates the application of the technology through a specific hospital case.

[0107] Through the above identification, the sentences in the paragraph are classified, laying the foundation for subsequent structured processing.

[0108] Step S1314: Arrange the extracted opinion expression sentences, argument support sentences, and case reference sentences in logical order, supplementing the inter-sentence conjunctions to form structured paragraphs. The structured paragraph contains three components: opinion items, argument items, and case items.

[0109] The opinion expression sentences, argument support sentences, and case reference sentences extracted from the independent paragraph are arranged in a certain logical order. The usual logical order is to first present the opinion expression sentence, then use the argument support sentence to support the opinion, and finally use the case reference sentence to illustrate with examples, making the logic of the paragraph more clear and coherent.

[0110] Add appropriate inter-sentence conjunctions between the arranged sentences. For example, conjunctions such as "for example", "specifically" can be used between the viewpoint expression sentence and the argument supporting sentence; conjunctions such as "for example", "for example" can be used between the argument supporting sentence and the case reference sentence. By adding conjunctions, the transition between sentences is more natural and smooth.

[0111] After the above processing, a structured paragraph is formed, which contains three components: viewpoint items, argument items, and case items. The viewpoint items are composed of viewpoint expression sentences, the argument items contain argument supporting sentences, and the case items are composed of case reference sentences. Each part has a clear content orientation, making the structure of the paragraph more clear and facilitating subsequent integration and use.

[0112] Step S1315: Add the corresponding chapter code to each structured paragraph as a chapter identifier, and combine all structured paragraphs under the same chapter code to generate a structured content unit with chapter code.

[0113] Add the chapter code corresponding to the content of each structured paragraph as a chapter identifier, ensuring that each paragraph can be clearly attributed to the corresponding chapter. For example, structured paragraphs belonging to the "M3-F1" chapter will all be added with the chapter code "M3-F1" as an identifier.

[0114] Combine all structured paragraphs under the same chapter code to form a complete content unit. These paragraphs are arranged in logical order in the chapter and together constitute the main content required by the chapter.

[0115] Finally, a structured content unit with chapter code is generated, each structured content unit corresponds to a specific chapter code and contains the structured content required by the chapter.

[0116] Step S140: Call the report integration agent to receive the structured content unit, combine the content module division standard in the framework specification to classify the structured content unit into modules and process logically in series to generate a report draft conforming to the chapter hierarchy.

[0117] Call the report integration agent, which receives all structured content units with chapter codes generated previously. Then, combined with the content module division standard in the framework specification, the structured content units are classified into modules.

[0118] The content module division criteria define the section range included in each module of the report. The report integration agent divides the structured content units coded with different chapters into corresponding modules according to this criterion. For example, the "disease overview" module may include structured content units related to "M1", the "new diagnostic technology research" module may include structured content units related to "M3", and so on.

[0119] After completing the module classification, the structured content units within each module are logically concatenated. By analyzing the logical relationship between the content units, appropriate transitional content is added to make the content within the module coherent and connected.

[0120] After module classification and logical concatenation, the contents of all modules are combined to generate a preliminary draft of the report that conforms to the section hierarchy relationship. This preliminary draft has the basic framework and main content of the report.

[0121] Step S141: Analyze the content module division criteria in the framework specification to determine the standard module structure of the report, which includes an introduction module, a main body module, and a conclusion module, each corresponding to a specific main section node in the section hierarchy relationship.

[0122] Analyze the content module division criteria in the framework specification to determine the standard module structure of the report. The standard module structure usually includes an introduction module, a main body module, and a conclusion module.

[0123] The introduction module is mainly used to introduce the research background, purpose, and significance of the report. The corresponding main section node in the section hierarchy relationship may be "M1: Disease Overview", etc. Through the content of these chapters, the theme of the report is introduced.

[0124] The main body module is the core part of the report, containing a large amount of research content, analysis process, etc. The corresponding main section nodes may include "M2: Analysis of Existing Diagnostic Methods", "M3: Research on New Diagnostic Technologies", "M4: Clinical Application Cases", etc. These chapters elaborate the core content of the report in detail.

[0125] The conclusion module is used to summarize the main points of the report, research results, and future prospects, etc. The corresponding main section node may be "M5: Future Development Trends", etc. to conclude and summarize the entire report.

[0126] Each module corresponds to a specific main section node in the section hierarchy relationship, clearly defining the module's position in the report and the content range it contains.

[0127] Step S142: calling the module mapping component of the report integration agent, matching the chapter codes in the structured content units with the modules in the standard module structure, and determining the corresponding structured content unit subset for each module.

[0128] The module mapping component of the report integration agent is called, which can match the chapter codes in the structured content units with the modules in the standard module structure according to the preset mapping rule.

[0129] The mapping rule specifies the modules corresponding to different chapter codes. For example, the structured content units with chapter codes starting with "M1" correspond to the introduction module; those starting with "M2", "M3", and "M4" correspond to the main body module; and those starting with "M5" correspond to the conclusion module.

[0130] Through the matching operation of the module mapping component, the corresponding structured content unit subset for each module is determined, and each subset contains all the structured content units required by the module, preparing for subsequent module content processing.

[0131] Step S143: processing each module in order according to the order of the standard module structure, arranging the structured content unit subset corresponding to each module in the order of the chapter node in the chapter hierarchical relationship, and generating a module content sequence.

[0132] According to the order of the introduction module, the main body module, and the conclusion module in the standard module structure, each module is processed in turn. For each module, its corresponding structured content unit subset contains multiple structured content units, which correspond to different chapter nodes under the main chapter of the module.

[0133] According to the order of the chapter node in the chapter hierarchical relationship, the structured content unit subset is arranged. For example, in the "M3: New Diagnostic Technology Research" main chapter of the main body module, there are "M3-F1: Technical Principles" and "M3-F2: Technical Advantages" chapter nodes, and the corresponding structured content units are arranged in the order of these chapter nodes.

[0134] After arrangement, the module content sequence of each module is generated, and the structured content units in the sequence are arranged in the logical order of the chapters, making the structure of the module content more clear.

[0135] Step S144: performing logical concatenation processing on the structured content units in the module content sequence, inserting transition paragraphs between adjacent structured content units, and generating transition paragraphs based on the viewpoint expression sentences of the two structured content units before and after, to connect the logical relationship between different content units.

[0136] The structured content units in the module content sequence are logically concatenated, focusing on the logical relationship between adjacent structured content units. By analyzing the viewpoint expression sentences of the two structured content units before and after, the logical relationship between them is determined, such as progression, parallelism, causality, etc.

[0137] According to the determined logical relationship, a transition paragraph is generated. The content of the transition paragraph is based on the viewpoint expression sentences of the two structured content units before and after, and the logical relationship between them is described and connected. For example, if the viewpoint of the previous content unit is to introduce the principle of a new diagnostic technology, and the next one is to introduce its advantages, the transition paragraph can be expressed as "After understanding the principle of the new diagnostic technology, its advantages over traditional technology are more obvious, which can be reflected in the following aspects".

[0138] The generated transition paragraph is inserted between adjacent structured content units, so that each unit in the module content sequence can be connected coherently to form an organic whole.

[0139] Step S145: Combine the concatenated content of all modules, add module titles and chapter titles, and generate a complete text containing the content of the introduction module, the main module and the conclusion module, as the initial draft of the report conforming to the chapter hierarchy.

[0140] Combine the content of all modules after logical concatenation, and arrange them in the order of introduction module, main module, and conclusion module.

[0141] Add the corresponding module title to each module, such as "I. Introduction", "II. Main Content", "III. Conclusion", etc.; at the same time, add the chapter title to each structured content unit corresponding to the chapter, such as "2.1 Principle of New Diagnostic Technology", etc.

[0142] After adding the title and combining the processing, a complete text containing the content of the introduction module, the main module and the conclusion module is generated, which is the initial draft of the report conforming to the chapter hierarchy, with the complete structure and main content of the report.

[0143] Step S146: Extract the viewpoint expression sentences of the two adjacent structured content units in the module content sequence, and mark them as the preposition viewpoint sentence and the postposition viewpoint sentence.

[0144] When logically concatenating the module content sequence, first extract the viewpoint expression sentences in the two adjacent structured content units. Mark the viewpoint expression sentence in the first structured content unit as the preposition viewpoint sentence, and the viewpoint expression sentence in the second structured content unit as the postposition viewpoint sentence.

[0145] For example, in the module content sequence, the opinion expression sentence of the previous structured content unit is "new diagnostic technology has a significant advantage in detection speed", and the opinion expression sentence of the next one is "the technology also performs well in detection accuracy". These two sentences are respectively the pre-positioned opinion sentence and the post-positioned opinion sentence.

[0146] Extracting these two opinion sentences is the basis for generating the transition paragraph. The logical relationship between them is analyzed to determine the transition mode.

[0147] Step S147: semantic relationship analysis is performed on the pre-positioned opinion sentence and the post-positioned opinion sentence to determine the logical relationship type between the pre-positioned opinion sentence and the post-positioned opinion sentence, the logical relationship type including progressive relationship, parallel relationship, causal relationship and transition relationship.

[0148] The semantic relationship analysis of the pre-positioned opinion sentence and the post-positioned opinion sentence determines the logical relationship type between the two by analyzing the meaning, keywords and context of the sentences.

[0149] If the post-positioned opinion sentence is a more in-depth elaboration or supplement based on the pre-positioned opinion sentence, they have a progressive relationship. For example, the pre-positioned opinion sentence introduces a certain advantage of the technology, and the post-positioned opinion sentence further introduces a more prominent advantage of the technology in another related aspect.

[0150] If the pre-positioned opinion sentence and the post-positioned opinion sentence respectively express the same theme from different angles, and the two are equal in status, then they have a parallel relationship. For example, the opinion sentences describing the technology from the speed and accuracy angles respectively have a parallel relationship.

[0151] If the post-positioned opinion sentence is the result or influence caused by the pre-positioned opinion sentence, then the two have a causal relationship. For example, the pre-positioned opinion sentence explains a certain characteristic of the technology, and the post-positioned opinion sentence explains the effect brought about by this characteristic.

[0152] If the post-positioned opinion sentence is opposite or contrasted to the expression of the pre-positioned opinion sentence, then it is a transition relationship. For example, the pre-positioned opinion sentence mentions a certain deficiency of the technology, and the post-positioned opinion sentence explains its advantage in another aspect.

[0153] Step S148: based on the logical relationship type, call the transition sentence template library to select the transition sentence template matching the logical relationship type, the transition sentence template including progressive type template, parallel type template, causal type template and transition type template.

[0154] According to the determined logical relationship type, the transition sentence template library is called. The transition sentence template library contains various types of transition sentence templates, corresponding to different logical relationship types.

[0155] The progressive template is suitable for progressive relationships, such as templates in the form of “not only …, but also …”, “in addition to …, also …”, and the like; the parallel template is suitable for parallel relationships, such as “on the one hand …, on the other hand …”, “at the same time …”, and the like; the cause-and-effect template corresponds to cause-and-effect relationships, such as “because …, therefore …”, “because …, so …”, and the like; and the transition template is used for transition relationships, such as “although …, but …”, “however …”, and the like.

[0156] Step S149: Fill the core concept in the preceding opinion sentence and the core concept in the following opinion sentence into the corresponding positions of the transition sentence template to generate a preliminary transition paragraph.

[0157] The core concept in the preceding opinion sentence and the core concept in the following opinion sentence are extracted, which are usually key nouns or noun phrases in the sentences and can represent the core content of the sentences. For example, the core concept in the preceding opinion sentence can be “the detection speed of the new diagnostic technology”, and the core concept in the following opinion sentence can be “the detection accuracy of the new diagnostic technology”.

[0158] These core concepts are filled into the corresponding positions of the selected transition sentence template. For example, if the selected template is the progressive template “not only … is outstanding in some aspects, but also … is also obviously advantageous in another aspect”, the core concept of the preceding opinion sentence is filled into the first ellipsis position, and the core concept of the following opinion sentence is filled into the second ellipsis position to generate a preliminary transition paragraph.

[0159] The preliminary transition paragraph already has the basic content of the transition, but the expression may not be smooth and natural enough.

[0160] Step S1410: Adjust the sentence fluency of the preliminary transition paragraph, replace the placeholders in the transition sentence template with specific concept expressions, so that the transition paragraph is consistent with the expression style of the adjacent structured content units, and insert the adjusted transition paragraph between the adjacent structured content units.

[0161] The sentence fluency of the preliminary transition paragraph is adjusted, and the grammar of the sentence is checked to see if it is correct and the expression is smooth. If it is not smooth, it is modified. At the same time, the placeholders in the transition sentence template are replaced with specific concept expressions, making the content of the transition paragraph more specific and clear.

[0162] In addition, it is ensured that the transition paragraph is consistent with the expression style of the adjacent structured content units, such as maintaining a rigorous and objective style in an academic report. If the adjacent content units use formal terms and expression methods, the transition paragraph should also adopt the corresponding style.

[0163] After adjustment, the transition paragraph is inserted between the adjacent structured content units, enabling smooth transition between the two content units and enhancing the coherence and logic of the module content.

[0164] Step S150: performing content checking and optimization operations on the preliminary report draft through the multi-agent collaborative framework to generate a final report text that conforms to the framework specification, the content checking and optimization operations including chapter integrity checking, content expression consistency checking, and module connection adjustment.

[0165] The multi-agent collaborative framework organizes relevant agents to perform content checking and optimization operations on the preliminary report draft. These operations include chapter integrity checking, checking whether the chapters of the preliminary report draft are complete and conform to the chapter hierarchy relationship in the framework specification; content expression consistency checking, ensuring that the expression of the same concept or term in the report remains consistent; and module connection adjustment, further optimizing the connection between modules to make the logic of the entire report more coherent.

[0166] Through these checking and optimization operations, the preliminary report draft is perfected and corrected, and a final report text that conforms to the framework specification is finally generated. The final report text is complete in content, clear in logic, and consistent in expression, and can meet the needs of clinical physicians for digital academic reports.

[0167] Step S151: calling the chapter checking agent in the multi-agent collaborative framework to parse the chapter structure of the preliminary report draft, comparing the parsed chapter title sequence with the chapter hierarchy relationship in the framework specification, and identifying missing chapter nodes and redundant chapter nodes.

[0168] The chapter checking agent in the multi-agent collaborative framework is called, which can parse the chapter structure of the preliminary report draft. During the parsing process, all chapter titles in the preliminary report draft are extracted to form a chapter title sequence.

[0169] The parsed chapter title sequence is compared with the chapter hierarchy relationship in the framework specification, and each chapter node that should exist is checked one by one to see if it exists in the preliminary report draft. If a chapter node specified in the framework specification does not have a corresponding title in the preliminary report draft, that node is a missing chapter node.

[0170] At the same time, it is checked whether there are chapter titles in the preliminary report draft that are not mentioned in the framework specification. These titles correspond to redundant chapter nodes. Through the above comparison, missing and redundant chapter nodes are identified.

[0171] Step S152: For the missing chapter node, based on the agent task assignment table, re-call the resource retrieval agent and the content extraction agent to generate the corresponding structured content unit, and supplement it to the corresponding position of the draft report; for the redundant chapter node, delete the corresponding structured content unit and chapter title.

[0172] For the identified missing chapter node, according to the task information corresponding to the chapter node in the agent task assignment table, re-call the resource retrieval agent to perform directional retrieval of relevant network resources and obtain resources related to the missing chapter node. The resource retrieval agent accurately retrieves the required content from the preset medical academic database, authoritative medical research website and other network resources according to the chapter code, core theme term and content attribute identifier in the task information, and generates a new associated resource subset.

[0173] Then, the content extraction agent is called to process the associated resource subset. The content extraction agent first extracts the metadata of each resource item in the resource subset to obtain resource title, content abstract and other information, and then performs semantic matching of these information with the task objective description to select resource items that meet the requirements. Then, the content of the selected resource items is subjected to paragraph division, core information extraction and other operations to generate structured content units with chapter codes, which are then supplemented to the position corresponding to the chapter node in the draft report to ensure the integrity of the report structure.

[0174] For the identified redundant chapter node, i.e. those that do not exist in the framework specification chapter hierarchy relationship but appear in the draft report, the corresponding structured content unit and chapter title need to be deleted. During the deletion process, it is necessary to check whether the content of the redundant chapter node is associated with other chapters, and if there is associated content, it is necessary to confirm whether these associated content has been reasonably reflected in other chapters to avoid the loss of useful information due to the deletion of redundant chapters. After deletion, the chapter order of the draft report is reorganized to ensure that the logical coherence between chapters is not affected.

[0175] Step S153: Call the content verification agent to perform content expression consistency verification on the draft report after the chapter structure is corrected, extract the core concept terms in the full text, and check whether the expression of the same core concept term in different chapters is consistent. If there is a difference in expression, it is uniformly replaced by the standard expression.

[0176] The content verification agent is called to start content expression consistency verification on the draft report after the chapter structure is corrected. First, the content verification agent will scan the draft report, process the text into independent word sequences through word segmentation, and then determine the candidate set of core concept terms based on the frequency of words in the full text, the number of times of appearance in the viewpoint expression sentence and the degree of association with the theme demand.

[0177] For each core concept term in the candidate set, the content verification agent will perform synonym and near-synonym identification, and construct a term group containing the main term and associated terms, where the main term is the term in the term group that appears most frequently in the full text and has the highest degree of match with the chapter theme description in the framework specification. For example, for the core concept "new immune regulation mechanism", its associated terms may include "new immune action mechanism", "new immune escape mode", etc., and after analysis, "new immune regulation mechanism" is determined as the main term.

[0178] Subsequently, all chapters of the preliminary draft report are traversed, and the associated terms belonging to each term group appearing in each chapter are recorded, and the expressions of the same term group in different chapters are checked one by one to see if they are unified as the main term. If it is found that an associated term is used instead of the main term in a chapter, such as using "new immune action mechanism" instead of "new immune regulation mechanism" in a chapter, the associated term is replaced with the main term.

[0179] If there is a case where the same term group corresponds to different main terms in different chapters, such as "new immune regulation mechanism" being used as the main term in the first chapter and "new immune action mechanism" being used as the main term in the third chapter, at this time the content verification agent will refer to the expression of the term in the chapter theme description in the framework specification, combined with the occurrence of the core concept term in the theme requirements, to determine a standard main term, and all expressions of the term group in the full text are replaced with the standard main term.

[0180] After all the replacement operations are completed, the content verification agent will generate a content expression consistency report, which details the term expressions before and after replacement, replacement location and replacement reason, as the basis for content expression consistency verification, and facilitates subsequent tracing and auditing of the report content.

[0181] For example, step S1531: performing full-text segmentation processing on the preliminary draft report after modifying the chapter structure to obtain a word sequence, and determining the core concept term candidate set based on the frequency of the word appearing in the full text and the number of times the word appears in the viewpoint expression sentence.

[0182] The segmentation tool used for full-text segmentation processing on the preliminary draft report after modifying the chapter structure can recognize professional terms in the medical research field, and split continuous text into independent word sequences according to semantics, while filtering out stop words in the text, such as "of", "in", "and" and other meaningless virtual words.

[0183] After obtaining the sequence of words, the frequency of each word in the full text is counted, and the higher the frequency, the more important the word in the text may be. At the same time, the number of times each word appears in the opinion expression sentence is counted, and the opinion expression sentence is the sentence expressing the core opinion and conclusion in the report. The more times a word appears in it, the closer it is to the core content of the report.

[0184] Considering the frequency of occurrence of the word and the number of times it appears in the opinion expression sentence, a corresponding screening threshold is set, and the words that meet the threshold condition are included in the core concept term candidate set. For example, for words that appear in the full text a certain number of times and appear in the opinion expression sentence more than a certain value, such as "immune checkpoint inhibitors", "advanced lung cancer", etc., they are listed as core concept term candidates.

[0185] Step S1532: Perform synonym and near-synonym recognition on each term in the core concept term candidate set to generate a term group containing a main term and associated terms, where the main term is the term with the highest frequency of occurrence in the term group.

[0186] For each term in the core concept term candidate set, the content verification agent will call the built-in synonym and near-synonym recognition library, which contains a large number of professional term synonym and near-synonym correspondence relationships in the field of medical research. By comparing, all synonyms and near-synonyms of each term are identified to form a term set.

[0187] Each term set is analyzed to count the frequency of each term in the full text of the initial draft of the report, and the term with the highest frequency of occurrence is determined as the main term, and the remaining terms are associated as associated terms, thereby generating a term group containing a main term and associated terms. For example, for the term set {“new immune regulation mechanism”, “new immune action mechanism”, “new immune escape mode”}, “new immune regulation mechanism” has the highest frequency of occurrence, so it is determined as the main term, and the main term group is {main term: new immune regulation mechanism; associated terms: new immune action mechanism, new immune escape mode}.

[0188] Step S1533: Traverse all chapters of the initial draft of the report, record the associated terms in each chapter of the term group, and check whether the associated terms of the same term group in different chapters are unified as the main term expression.

[0189] The content verification agent traverses all chapters of the initial draft of the report in the order of chapters, retrieves the associated terms belonging to each term group in each chapter, and records the position and number of times each associated term appears.

[0190] The associated terms of the same term group in different chapters are compared with the main term of the term group to check if there is a use of associated terms instead of the main term. For example, in the "tumor immune research case analysis" chapter, the associated term "new type of immune action mechanism" is found, and the corresponding main term is "new type of immune regulation mechanism". This situation is recorded for subsequent uniform replacement.

[0191] Step S1534: If there is a use of associated terms instead of the main term in a chapter, replace the associated term with the main term; if there is a case where the same term group corresponds to different main terms in different chapters, determine the standard main term based on the occurrence of the term in the topic description in the framework specification chapter, and uniformly replace it with the standard main term.

[0192] When it is found that an associated term is used instead of the main term in a chapter, the content verification intelligent agent will directly replace the associated term with the main term to ensure that the same term group is expressed consistently with the main term in the chapter. For example, replace "new type of immune action mechanism" in the "tumor immune research case analysis" chapter with "new type of immune regulation mechanism".

[0193] If the same term group corresponds to different main terms in different chapters, such as "new type of immune regulation mechanism" being used as the main term in the "research status" chapter and "new type of immune escape mechanism" being used as the main term in the "molecular mechanism review" chapter, the content verification intelligent agent will consult the chapter topic description in the framework specification to count the number of occurrences of each term in the topic description. The term group determines the standard main term with the most occurrences. If the number of occurrences is the same, the matching degree of the term and the topic demand is judged to finally determine a standard main term, and all expressions of the term group in the whole text are uniformly replaced with the standard main term.

[0194] Step S1535: Generate a core concept term consistency report to record all replacement operations and term expressions before and after replacement as the basis for content expression consistency verification.

[0195] After completing the uniform replacement of all core concept terms, the content verification intelligent agent generates a core concept term consistency report. The core concept term consistency report details the main term and associated term of each term group, the chapter location where the replacement operation occurs, the term expression before replacement, the standard expression after replacement, and the replacement reason and other information.

[0196] For example, the report will have the above record: the main term of the term group is "new immune regulation mechanism", in the third paragraph of the "tumor immune research case analysis" chapter, "new immune mechanism" is replaced by "new immune regulation mechanism", and the replacement reason is that the expression is a related term and needs to be unified as the main term. This report can serve as an important basis for content consistency verification, facilitating subsequent review and modification of the report content, and ensuring the consistency of the core concept term expression in the report.

[0197] Step S154: Call the cohesion verification intelligent agent to perform cohesion checking on the transition paragraphs between modules in the preliminary draft of the report, calculate the semantic correlation degree of the transition paragraphs with the content of the previous and next modules, and regenerate the transition paragraphs if the semantic correlation degree is lower than the preset value.

[0198] Call the cohesion verification intelligent agent to perform cohesion checking on the transition paragraphs between modules in the preliminary draft of the report. The cohesion verification intelligent agent first extracts the content of each transition paragraph and its previous and next modules, converts the transition paragraph text into a transition semantic vector, converts the core viewpoint sentences of the previous module content into a previous module semantic vector, and converts the core viewpoint sentences of the next module content into a next module semantic vector.

[0199] Then, calculate the cosine similarity of the transition semantic vector and the previous module semantic vector to obtain the forward semantic correlation degree; calculate the cosine similarity of the transition semantic vector and the next module semantic vector to obtain the backward semantic correlation degree. Average the forward semantic correlation degree and the backward semantic correlation degree to obtain the comprehensive semantic correlation degree of the transition paragraph and the content of the previous and next modules.

[0200] If the comprehensive semantic correlation degree is higher than the preset value, it means that the transition paragraph can effectively connect the content of the previous and next modules and does not need to be modified; if the comprehensive semantic correlation degree is lower than the preset value, it means that the correlation between the transition paragraph and the content of the previous and next modules is weak, and the transition paragraph needs to be regenerated.

[0201] When regenerating the transition paragraph, the cohesion verification intelligent agent will re-extract the core viewpoint sentences of the previous and next module content, analyze the logical relationship between them, call the corresponding template in the transition sentence template library, fill in the core concept terms, generate a new transition paragraph, and again calculate the semantic correlation degree of the new transition paragraph and the content of the previous and next modules until it reaches the preset value or above.

[0202] Step S155: Perform format standardization processing on the preliminary draft of the report that has completed verification and optimization to generate a final report text that conforms to the framework specification.

[0203] The first step is to standardize the format of the report draft after the chapter structure is revised, the content expression consistency is checked, and the transition paragraph coherence is optimized. First, according to the requirements of font, font size and line spacing in the framework specification, the text format of the entire report is unified, such as using specific font and font size for main chapter titles, using corresponding font and font size for section titles and text content, to ensure the standardization of text layout.

[0204] Secondly, the format of the charts in the report is adjusted. If the report contains clinical data charts, the format of the chart title, coordinate axis label and legend should be set according to the requirements of the framework specification to ensure the clarity and standardization of the charts. At the same time, the reference relationship between the charts and the text content is checked to ensure that the chart number is consistent with the reference number in the text.

[0205] Then, the page number, header and footer of the report are set according to the requirements of the framework specification to determine the position and format of the page number, and the header and footer can contain the theme name of the report, chapter title and other information to make the overall structure of the report more complete.

[0206] Finally, the overall layout of the report is checked to ensure that the page division between chapters is reasonable and to avoid chapter content breaking across pages, making the report more visually neat and easy to read. After all the format standardization is completed, the final report text that meets the framework specification is generated.

[0207] Figure 2 A report generation system 100 based on a multi-agent architecture is shown, which includes a processor 1001, a memory 1003, and program code stored in the memory 1003. The processor 1001 executes the above-mentioned program code to realize the steps of the report generation method based on the multi-agent architecture.

[0208] Figure 2 The report generation system 100 based on the multi-agent architecture includes a processor 1001 and a memory 1003. The processor 1001 and the memory 1003 are connected, such as through a bus 1002. Optionally, the report generation system 100 based on the multi-agent architecture can also include a transceiver 1004, which can be used for data interaction between the report generation system based on the multi-agent architecture and other report generation systems based on the multi-agent architecture, such as data transmission and / or data reception. It should be noted that the transceiver 1004 is not limited to one in actual scheduling, and the structure of the report generation system 100 based on the multi-agent architecture does not constitute a limitation on the embodiments of the present application.

[0209] The processor 1001 can be a CPU (Central Processing Unit), a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array) or other programmable logic device, transistor logic device, hardware component, or any combination thereof. The processor 1001 can implement or execute the various exemplary logical blocks, modules, and circuits described in connection with the disclosure. The processor 1001 can also be a combination of computing functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, and the like.

[0210] The bus 1002 can include a path that transmits information between the above-mentioned components. The bus 1002 can be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, or the like. The bus 1002 can be divided into an address bus, a data bus, a control bus, and the like.

[0211] The memory 1003 can be a ROM (Read Only Memory) or other type of static storage device that can store static information and instructions, a RAM (Random Access Memory) or other type of dynamic storage device that can store information and instructions, an EEPROM (Electrically Erasable Programmable Read Only Memory), a CD-ROM (Compact Disc Read Only Memory) or other optical disk storage, a magnetic disk storage medium, other magnetic storage device, or any other medium that can be used to store program code in connection with the disclosure, and that can be read by a computer, without limitation.

[0212] The memory 1003 is used to store program code for implementing the embodiments of the disclosure, and is controlled by the processor 1001 to execute. The processor 1001 is used to execute the program code stored in the memory 1003 to implement the steps shown in the foregoing method embodiments.

[0213] The embodiment of the present application provides a computer readable storage medium, which stores program codes, and the program codes are executed by a processor to implement the steps and corresponding contents of the foregoing method embodiment.

[0214] It should be understood that, although the flowcharts of the embodiments of the present application indicate the respective operation steps by arrows, the implementation order of the steps is not limited to the order indicated by the arrows. Unless explicitly stated herein, in some implementation scenarios of the embodiments of the present application, the implementation steps in each flowchart can be executed in other orders based on requirements. In addition, part or all of the steps in each flowchart can include multiple sub-steps or multiple stages, and part or all of the sub-steps or stages can be executed at the same time, or each of the sub-steps or stages can be executed at different times. In the scenario where the execution times are different, the execution order of the sub-steps or stages can be flexibly configured based on requirements, and the embodiments of the present application do not limit this.

[0215] The above is only an optional implementation of some implementation scenarios of the present application, and it should be pointed out that, for those skilled in the art, other similar implementation manners according to the technical idea of the present application without departing from the technical concept of the present application also belong to the protection scope of the embodiments of the present application.

Claims

1. A report generation method based on multi-agent architecture, characterized in that, The method comprises: receiving a report generation instruction, parsing the report generation instruction to obtain a topic requirement and a corresponding framework specification, obtaining initial reference materials associated with the topic requirement, and the framework specification including chapter hierarchical relationship and content module division standard; starting a multi-agent collaborative framework, calling a resource retrieval agent, a content extraction agent and a report integration agent, generating an agent task allocation table based on the chapter hierarchical relationship in the framework specification, the agent task allocation table including the task identification, chapter code and task execution requirement corresponding to each agent; performing a network resource directional retrieval operation according to the task identification and chapter code in the agent task allocation table through the resource retrieval agent, generating an associated resource set matching the chapter hierarchical relationship, calling the content extraction agent to perform content structured conversion processing on the associated resource set and the initial reference materials, and generating a structured content unit with chapter code; calling the report integration agent to receive the structured content unit, classifying and logically concatenating the structured content unit according to the content module division standard in the framework specification, and generating a report draft conforming to the chapter hierarchical relationship; performing content checking and optimization operations on the report draft through the multi-agent collaborative framework to generate a final report text conforming to the framework specification; the calling of the content extraction agent to perform content structured conversion processing on the associated resource set and the initial reference materials to generate a structured content unit with chapter code comprises: merging the resource entry content corresponding to each chapter code in the associated resource set and the material content of the corresponding chapter topic in the initial reference materials to obtain a comprehensive content pool classified by chapter code; calling the paragraph division module of the content extraction agent to perform paragraph boundary identification processing on the text content in the comprehensive content pool to obtain a plurality of independent paragraphs, each independent paragraph including a paragraph start position identifier and a paragraph end position identifier; performing core information extraction processing on each independent paragraph to identify viewpoint expression sentences, argument support sentences and case reference sentences in the independent paragraph, wherein the viewpoint expression sentence is a sentence containing a conclusion expression, the argument support sentence is a sentence containing data or fact description, and the case reference sentence is a sentence containing specific instance description; arranging the extracted viewpoint expression sentences, argument support sentences and case reference sentences in logical order, supplementing inter-sentence conjunctions to form structured paragraphs, and the structured paragraphs including three components: viewpoint items, argument items and case items; adding the corresponding chapter code as the chapter identifier to each structured paragraph, grouping all structured paragraphs under the same chapter code, and generating a structured content unit with chapter code.

2. The report generation method based on multi-agent architecture according to claim 1, wherein, the starting of the multi-agent collaborative framework, the calling of the resource retrieval agent, the content extraction agent and the report integration agent, and the generation of the agent task allocation table based on the chapter hierarchical relationship in the framework specification comprise: The chapter hierarchical relationship includes a main chapter node, a sub chapter node and a sub chapter node, each node has a unique chapter code and a chapter theme description; Extract the chapter theme description of each node in the chapter hierarchical relationship, and determine the content attribute identifier corresponding to each node by content attribute determination on the chapter theme description. The content attribute identifier includes theory type, data type and case type; Based on the content attribute identifier, the intelligent agent capability matching rule is queried to determine the intelligent agent type corresponding to each content attribute identifier. The resource retrieval intelligent agent corresponds to the theory type, the content extraction intelligent agent corresponds to the data type and the case type, and the report integration intelligent agent corresponds to the integration of all type nodes; For each chapter node, a task description item is generated combining its chapter code, content attribute identifier and corresponding intelligent agent type. The task description item includes chapter code, intelligent agent type identifier, content attribute identifier and task target description; Based on the node order in the chapter hierarchical relationship, the task description item is sorted to generate an intelligent agent task allocation table indexed by chapter code, including intelligent agent type identifier and task target description.

3. The report generation method based on multi-agent architecture according to claim 1, wherein, The network resource directional retrieval operation is performed by the resource retrieval intelligent agent according to the task identifier and chapter code in the intelligent agent task allocation table to generate an associated resource set matched with the chapter hierarchical relationship, including: The task identifier corresponding to the chapter code and the task target description is obtained by analyzing the intelligent agent task allocation table, and the core theme term in the task target description is extracted. The core theme term is the key concept expression in the chapter theme description; Call the retrieval word expansion module of the resource retrieval intelligent agent to perform synonym expansion processing on the core theme term to generate a retrieval word set containing the core theme term and synonyms; Based on the retrieval word set and chapter code, a retrieval request is constructed, which includes retrieval word set, chapter code and content attribute identifier. The retrieval request is submitted to the resource database through a preset network resource interface; Receive the candidate resource list returned by the resource database, extract the metadata of the resource item in the candidate resource list, and obtain the resource title, publication time, source identifier and content abstract; The resource title and content abstract are semantically matched with the task target description, and the resource items whose semantic matching results meet the preset conditions are screened out. The screened resource items are stored according to the chapter code to generate an associated resource set matched with the chapter hierarchical relationship.

4. The report generation method based on multi-agent architecture according to claim 3, wherein, The resource title and content abstract are semantically matched with the task target description, and the resource items whose semantic matching results meet the preset conditions are screened out, including: Extract the theme keyword sequence in the task target description, which is composed of the core theme term and the feature word corresponding to the content attribute identifier; Extract the title keyword sequence in the resource title and the abstract keyword sequence in the content abstract. The title keyword sequence and the abstract keyword sequence are obtained by word segmentation processing and stop word filtering; Calculate the number of coincident words between the topic keyword sequence and the title keyword sequence, denoted as the title matching number, and the number of coincident words between the topic keyword sequence and the abstract keyword sequence, denoted as the abstract matching number; Perform semantic vector conversion on the task target description, the resource title and the content abstract respectively to obtain a target semantic vector, a title semantic vector and an abstract semantic vector, calculate the cosine similarity between the target semantic vector and the title semantic vector as a title similarity value, and calculate the cosine similarity between the target semantic vector and the abstract semantic vector as an abstract similarity value; Determine a comprehensive matching index based on the title matching number, the abstract matching number, the title similarity value and the abstract similarity value, the comprehensive matching index being obtained by multiplying the sum of the title matching number and the abstract matching number by the average of the title similarity value and the abstract similarity value, and screen resource entries with a comprehensive matching index greater than a preset threshold.

5. The multi-agent architecture based report generation method of claim 1, wherein, The core information extraction processing is performed on each independent paragraph, and viewpoint expression sentences, argument support sentences and case reference sentences in the independent paragraph are identified, including: Perform part-of-speech tagging on each sentence in the independent paragraph to identify predicate verbs and conjunctions in the sentence, match sentences containing viewpoint feature words based on a preset viewpoint feature word library, and mark the matched sentences as candidate viewpoint expression sentences; Perform semantic role labeling on the candidate viewpoint expression sentences to extract subject components, predicate components and object components in the candidate viewpoint expression sentences, and if the object component is a conclusion concept expression, the candidate viewpoint expression sentence is confirmed as a viewpoint expression sentence; Locate the position of the viewpoint expression sentence in the independent paragraph, extract the sentences adjacent to the viewpoint expression sentence as argument candidate sentences, and perform data feature recognition on the argument candidate sentences, and if the argument candidate sentence contains numerical description, experimental results or factual statements, it is marked as an argument support sentence; Scan all unmarked sentences in the independent paragraph, identify case guide sentences based on a preset case feature word, and mark the case guide sentences and their subsequent explanatory sentences as case reference sentences; Check whether the viewpoint expression sentence has a corresponding argument support sentence, if not, find an associated argument support sentence from the adjacent paragraph and supplement it to the current paragraph, so that each viewpoint expression sentence corresponds to at least one argument support sentence.

6. The multi-agent architecture based report generation method of claim 1, wherein, The report integration agent receives the structured content unit, classifies the structured content unit into modules and performs logical concatenation processing according to the content module division standard in the framework specification, and generates a report draft conforming to the chapter hierarchical relationship, including: Analyze the content module division standard in the framework specification to determine the standard module structure of the report, the standard module structure including an introduction module, a main body module and a conclusion module, each module corresponding to a specific main chapter node in the chapter hierarchical relationship; Call the module mapping component of the report integration agent to match the chapter code in the structured content unit with the modules in the standard module structure, and determine the subset of structured content units corresponding to each module; sequentially processing each module in order of standard module structure, arranging a subset of structured content units corresponding to each module in order of division nodes in the section hierarchical relationship, and generating a module content sequence; performing logical concatenation processing on the structured content units in the module content sequence, inserting transition paragraphs between adjacent structured content units, and generating the transition paragraphs based on viewpoint expression sentences of two adjacent structured content units, for linking the logical relationship of different content units; combining the concatenated content of all modules, adding module titles and section titles, and generating a complete text containing the content of the introduction module, the main body module, and the conclusion module, as a draft report conforming to the section hierarchical relationship.

7. The report generation method based on multi-agent architecture according to claim 6, wherein, The logical concatenation processing on the structured content units in the module content sequence includes: extracting viewpoint expression sentences of two adjacent structured content units in the module content sequence, and denoting the two viewpoint expression sentences as a preceding viewpoint sentence and a following viewpoint sentence, respectively; performing semantic relationship analysis on the preceding viewpoint sentence and the following viewpoint sentence, determining the logical relationship type between the preceding viewpoint sentence and the following viewpoint sentence, and the logical relationship type including a progressive relationship, a parallel relationship, a cause-and-effect relationship, and a transition relationship; calling a transition sentence template library based on the logical relationship type, selecting a transition sentence template matching the logical relationship type, and the transition sentence template including a progressive template, a parallel template, a cause-and-effect template, and a transition template; filling core concepts in the preceding viewpoint sentence and core concepts in the following viewpoint sentence into corresponding positions of the transition sentence template to generate a preliminary transition paragraph; performing sentence fluency adjustment on the preliminary transition paragraph, replacing placeholders in the transition sentence template with specific concept expressions, so that the transition paragraph is consistent with the expression style of the adjacent structured content units, and inserting the adjusted transition paragraph between the adjacent structured content units.

8. The multi-agent architecture based report generation method of claim 1, wherein, The content verification and optimization operation on the draft report in the multi-agent collaborative framework includes: calling a section verification agent in the multi-agent collaborative framework, analyzing the section structure of the draft report, comparing the section title sequence obtained by the analysis with the section hierarchical relationship in the framework specification, and identifying missing section nodes and redundant section nodes; for the missing section nodes, re-calling the resource retrieval agent and the content extraction agent based on the agent task allocation table to generate corresponding structured content units, and supplementing the corresponding positions of the draft report; for the redundant section nodes, deleting the corresponding structured content units and section titles; calling a content verification agent to perform content expression consistency verification on the draft report after the section structure is corrected, extracting core concept terms in the full text, checking whether the expressions of the same core concept term in different sections are consistent, and uniformly replacing the expressions if there are differences; calling a linking verification agent to perform linking checking on the transition paragraphs between modules in the draft report, calculating the semantic association degree of the transition paragraphs and the content of the adjacent modules, and re-generating the transition paragraphs if the semantic association degree is lower than a preset value. The report draft completed with the check and the optimization is subjected to format standardization processing, and a final report text conforming to the framework specification is generated.

9. A report generation system based on multi-agent architecture, characterized in that, The application also provides a computer readable storage medium storing machine executable instructions, wherein the machine executable instructions are executed by a processor to implement the report generation method based on the multi-agent architecture according to any one of claims 1-8.

Images