Education review thesis structured generation method, device and system and storage medium

Abstract

The invention discloses a method, a device and a system for structural generation of an education review thesis and a storage medium, which are characterized in that structural generation of the education review thesis is realized by adopting a multi-agent cooperation method of literature review, data analysis and logic verification. The device comprises a request analysis module, a task distribution module, an evidence processing module, a conflict resolution module and a structured writing module. The system comprises a configuration storage component, a data access component, a model service component and a shared workspace component. A computer program is stored in the storage medium, and when the computer program runs on the processor, the method for generating the structural education review thesis is executed. Compared with the prior art, the method has the advantages that the factual accuracy, demonstration coherence and reference credibility of review generation are remarkably improved, the structure drift and logic confusion risks during single large model long text generation are reduced, multi-source heterogeneous data are realized through evidence unitization and reference anchor point association, and the method has a good application prospect.

Description

Technical Field

[0001] This invention relates to the field of educational informatization and intelligent writing assistance technology, specifically to a method, apparatus, system and storage medium for the structured generation of educational review papers based on multi-agent collaboration. Background Technology

[0002] Review articles in education (including systematic reviews, scope reviews, and narrative reviews) are crucial for defining research questions, gathering evidence, identifying research gaps, and proposing research agendas in educational research. Currently, review writing in educational research typically requires searching, screening, summarizing, and comparing a large number of literature entries, forming a hierarchical structure within a structured writing framework, such as "research topic—research paradigm—methodological elements—context and sample—conclusions and controversies—research gaps and future directions." This process heavily relies on the researcher's experience and manual compilation, is time-consuming, lacks verifiability, and has high iteration costs. Meanwhile, systematic reviews and evidence gathering have relatively mature reporting standards and process requirements in academia. For example, PRISMA 2020 proposed a checklist and flowchart for systematic review reports to promote transparency and traceability in the retrieval, inclusion, and reporting processes. Regarding the "tooling / semi-automation of the review process," existing research has attempted to assist or automate several stages of systematic reviews using machine learning and natural language processing. For example, pipeline-style processing frameworks have been constructed around stages such as retrieval, deduplication, screening, information extraction, and evidence summarization, and methods and applicable boundaries for different stages have been provided. Related work indicates that while automating systematic reviews can reduce the human burden to some extent, challenges remain in terms of consistent implementation of inclusion and exclusion criteria, traceable citation of evidence extraction, and merging results across different sources, as well as stability and interpretability.

[0003] In recent years, pre-trained language models have demonstrated strong capabilities in text generation and summarization tasks, making it possible to generate long review articles in natural language. However, long text generation still faces significant challenges in terms of factual consistency, citation credibility, and logical coherence. Existing review studies have summarized the "illusion" phenomenon of large models, pointing out that models may generate seemingly fluent content that lacks source support or is inconsistent with evidence, thus posing risks to the authenticity, verifiability, and auditability of academic writing. The complexity of educational research data further amplifies these problems. Educational research often includes unstructured data such as classroom videos, interview recordings, and observation records. Researchers need to transcribe, encode, and thematically analyze the audio and video data before organizing it with documentary evidence into a review argument chain. Related research has discussed the capability boundaries and error sources of automatic speech recognition in dialogue transcription research, suggesting that when using audio transcription results for research analysis, traceable information such as source location and time range should be retained to support subsequent verification and citation. From an engineering perspective, single models are prone to problems such as task goal drift, loss of constraints between stages, and difficulty in aligning intermediate results when undertaking multi-stage tasks such as "retrieval, extraction, statistics, writing, verification, and revision." To alleviate the vulnerability of single-point generation, academia and engineering have proposed a strategy of breaking down complex tasks and collaborating with external resources through a "reasoning-action" interaction. By obtaining evidence through explicit action interfaces and then performing reasoning and generation, the controllability of the process and the interpretability of the output are improved. Multi-agent systems provide the theoretical and methodological foundation for the division of labor and collaboration in complex tasks.

[0004] Early research on intelligent agents systematically discussed the key issues of "autonomy, interaction, and cooperation," and regarded task allocation, cooperative control, and communication mechanisms as the core capabilities of multi-agent systems. Works such as the Contract Network protocol proposed task sharing and allocation mechanisms for solving distributed problems, enabling multiple agents to complete collaborative execution through bidding / tendering interactions. At the same time, multi-agent communication languages ​​and interaction protocols (such as FIPA-related specifications) emphasize message encapsulation, interaction semantics, and interoperability, providing a standardized foundation for communication and cooperation between heterogeneous intelligent agents.

[0005] Existing educational review paper writing processes suffer from several shortcomings, including: heavy reliance on human experience; difficulty in unifying the organization of cross-source evidence; structural drift and logical inconsistencies in long text generation; lack of traceability and verification of citations and evidence; and difficulty in forming a verifiable structured argument chain when dealing with unstructured data such as classroom videos and interview recordings. These shortcomings include: 1) Many review aids focus on retrieval and abstracting, lacking a multi-agent collaborative framework with clear role division, clear communication protocols, and traceable task documents and result packages, leading to inconsistent alignment of multi-stage outputs; 2) a lack of configurable conflict detection and resolution mechanisms for conflicts during the generation process, such as factual conflicts, citation conflicts, structural conflicts, and logical conflicts, making it difficult to output auditable arbitration records while retaining uncertainty warnings; and 3) a lack of evidence unitization and citation anchoring mechanisms deeply integrated with the review writing structure for unstructured data such as classroom videos and interview recordings, making it difficult for audio-visual evidence to form a structured argument chain together with documentary evidence. In summary, there is an urgent need for a structured generation technology for educational review papers that can simulate the collaborative model of real scientific research teams and enable task distribution, communication collaboration, and conflict resolution for multi-source data (especially unstructured data) in educational research. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings of existing technologies by providing a method, apparatus, system, and storage medium for generating structured review papers in education. It employs a multi-agent collaborative workflow with clearly defined roles, where different agents act as literature reviewers, data analysts, and logic checkers. Through communication protocols, task distribution logic in the form of vouchers, and conflict resolution mechanisms, the collaborative process among multiple agents can generate traceable tasks, results, and arbitration records within a shared workspace. Simultaneously, structured / semi-structured data such as paper entries, along with unstructured data such as classroom recordings and interview audio recordings, are uniformly incorporated into an evidence unitization processing framework. Evidence tracing is achieved through source and location identifiers, enabling key arguments in the review text to be mapped to verifiable sources and locations of evidence. This method designs a unified workflow encompassing "retrieval—evidence extraction—statistical summarization—structured writing—logic verification—conflict resolution—output and feedback." It not only focuses on generating the final review text but also emphasizes the explicitness, traceability, and alignment of intermediate products during the generation process. This enhances the structural stability, argumentation coherence, citation verifiability, and auditability of the review paper. This invention reduces the risk of target drift in single-point generation through multi-role division of labor and multi-stage alignment. The literature review agent organizes the review structure and writing expression; the data analysis agent generates structured intermediate results such as topic clustering, distribution statistics, methodological element summarization, and comparative induction; and the logic verification agent triggers consistency checks, citation backtracking checks, and conflict detection. The conflict resolution module outputs arbitration results and uncertainty warnings. This method ensures that "writing output" is supported by "statistics and evidence" and that "conflicts and uncertainties" are explicitly recorded and reflected in the output, thereby improving interpretability and verifiability. The method is simple, effective, and has promising application prospects.

[0007] The specific technical solution to achieve the purpose of this invention is: a method for generating structured literature review papers in education, characterized by employing a multi-agent collaborative approach involving literature review, data analysis, and logical verification to generate structured literature review papers in education. This method specifically includes the following steps:

[0008] S1. Constructing a multi-agent collaborative workflow

[0009] A multi-agent collaborative workflow is constructed based on agent configuration files. This workflow includes at least three agents: a literature reviewer agent, a data analyst agent, and a logic checker agent. Each agent is configured with its input field set, output structure constraints, tool access permissions, and communication endpoints. The workflow also includes a task distribution module, a shared workspace module, and a conflict resolution module. The shared workspace module stores shared objects such as request parameter sets, task ticket sets, candidate item sets, evidence unit sets, structured intermediate result sets, draft sets, and conflict record sets. It associates session identifiers, version identifiers, and timestamp identifiers with these shared objects to facilitate traceability, review, and auditing of the generation process and results.

[0010] S2. Obtain the review request and map it to a set of request parameters.

[0011] The system acquires the user's review generation request and parses, validates, and maps the user input fields according to the agent configuration file to form a set of request parameters for subsequent processing. The user input fields include at least a topic field representing the review topic, a time range field representing the search and inclusion scope, a research type field representing the included research paradigm, and a structural constraint field representing the output structure and citation requirements. The agent configuration file determines the mandatory fields, default value strategies, default fallback strategies, field conflict handling strategies, and output structure constraints for each field. The set of request parameters is used to drive at least the generation of search queries, search filtering, execution of inclusion and exclusion rules, evidence timeliness screening, output chapter level constraints, and citation presentation rule execution, and is also used to constrain the input and output boundaries of subsequent multi-agent tasks to ensure that subsequent products maintain structural consistency and facilitate comparison.

[0012] S3, Task plan generation, breakdown, and ticket distribution

[0013] The task distribution module generates a task plan based on the request parameter set and decomposes the summary generation task into a task ticket set. The task ticket set includes at least any combination of the following: retrieval task ticket, candidate item organization task ticket, evidence extraction task ticket, evidence unitization task ticket, statistical summary task ticket, structured writing task ticket, and logic verification task ticket. Each task ticket includes at least a task type identifier, input parameter description, expected output description, dependency identifier, priority or deadline strategy identifier, and fallback strategy identifier. The task distribution module distributes the task tickets to the corresponding agents for execution according to the task type to role mapping rules and records the task status and task version in the shared workspace module. When task dependencies are not met, the task distribution module controls the triggering of subsequent tasks. When a task fails or the result does not meet the gating conditions, a retry, degradation, or alternative path is executed according to the fallback strategy. This task ticketing mechanism is used to transform multi-agent collaboration from "implicit dialogic collaboration" to "explicit traceable collaboration," making each step of input, output, and responsible party clearly traceable.

[0014] S4. Document retrieval, candidate item aggregation, and metadata standardization

[0015] The literature reviewer agent and its collaborating retrieval agents perform literature retrieval and candidate item aggregation, and standardize the metadata of the candidate items. The candidate items include at least paper entries, and may include any combination of project materials, conference materials, policy documents, course resources, research data descriptions, and practice reports. Metadata standardization includes at least field normalization, time field unification, source identifier writing, and duplicate item removal, forming a candidate item set that can be used for subsequent evidence extraction and statistical summarization. When the request parameter set contains unstructured data entries, classroom recordings, interview recordings, observation records, or their transcribed texts are further integrated as candidate evidence sources, and source identifiers and location index rules are written for these candidate evidence sources, enabling subsequent evidence units to be traced back to the audio / video time range or transcribed segment position.

[0016] By incorporating both unstructured data and entry data into the candidate set, the needs for generating reviews in educational and research scenarios characterized by "multi-source, multi-modal, and cross-evidence types" can be met.

[0017] S5. Evidence Extraction, Evidence Unitization, and Citation Anchor Point Association

[0018] Evidence extraction and unitization are performed on candidate entries and candidate evidence sources to form a set of evidence units. Each evidence unit includes at least any combination of factual description fragments, methodological description fragments, conclusive description fragments, limitation description fragments, and controversial description fragments. The unitization process includes at least: segmenting text sources into paragraphs, sentence groups, or semantic units; and segmenting audio and video sources into time windows or transcribed segments. Each evidence unit is associated with a source identifier, location identifier, time identifier, and citation anchor, enabling it to be referenced and traced back. Text processing of classroom recordings and interview recordings can employ automatic speech recognition or manual transcription, and the text processing results are indexed by time and location. The resulting evidence units can be traced back to the corresponding audio and video segment range via location identifiers. The set of evidence units is written to a shared workspace module and associated with candidate entries, research type fields, and time range fields, allowing subsequent statistical summarization and writing citations to be conducted on a unified data structure.

[0019] S6. Statistical summarization, topic clustering, and generation of structured intermediate results.

[0020] The data analyst agent statistically summarizes, structurally summarizes, and compares the evidence unit set and candidate item metadata to form a structured intermediate result set for writing. The structured intermediate result set includes at least any combination of topic clustering results, research paradigm distribution results, summary results of research questions or research object elements, summary results of method elements, summary results of variable and indicator elements, summary results of data source and sample elements, summary results of context elements, and summary results of research conclusions and points of contention. For each structured summary conclusion, it is associated with its supporting evidence unit citation set to form a support relationship of "conclusion - evidence unit - source location". This support relationship is used to ensure citation alignment in the subsequent writing stage and to check "whether the conclusion has evidence, whether the evidence is traceable, and whether the evidence is sufficient" in the verification stage, thereby improving the verifiability of the generated review.

[0021] S7, Consistency Check, Reference Backtracking Check, and Conflict Resolution

[0022] A logic checker agent performs consistency checks, citation traceability checks, and conflict detection on the structured intermediate result set and the draft. The conflict resolution module then categorizes, aggregates evidence, and arbitrates the detected conflicts to form a conflict record. These conflicts include at least any combination of factual conflicts, citation conflicts, structural conflicts, and logical conflicts. When arbitrating, the conflict resolution module considers at least source consistency, evidence coverage, timeliness consistency, and citation traceability availability. If a unique arbitration conclusion cannot be determined, an uncertainty warning is output. The conflict record includes at least the conflict type, conflict location, set of related evidence units, arbitration strategy identifier, arbitration result description, and uncertainty warning fields, and is written to the shared workspace module so that subsequent writing and output can be processed based on the conflict record. For factual conflicts, parallel viewpoints or retained disputed descriptions can be output; for citation conflicts, missing citations or failed traceability can be indicated; for structural conflicts, inconsistent chapter levels or paragraph classifications can be indicated; and for logical conflicts, broken inference chains or jumps in argumentation can be indicated.

[0023] S8, Structured Writing, Chapter Filling, and Citation Chain Generation

[0024] The literature reviewer agent generates structured text results of educational review papers based on a structured intermediate result set, conflict records, and output structure constraints. The structured text results include at least the paper title, abstract, introduction, hierarchical structure of the review topic, comparison of research paradigms and points of contention, research gaps and future directions, methodological suggestions, references, and citation marks. Key arguments in the structured text results are linked to source and location identifiers in the evidence unit set through citation anchors, thus forming a traceable citation chain. It can also output evidence summary fields and conflict summary fields corresponding to chapters, enabling the display and verification of the generation basis without limiting the specific front-end format or specific data format.

[0025] By employing the aforementioned structured writing and citation association, the "structure-content-evidence-source" structure of the review text is kept aligned, reducing the risk of structural and factual drift in the generation of long texts.

[0026] S9, Output and Feedback Closed-Loop Update

[0027] The system outputs structured text results and optionally receives user feedback to form a closed-loop update. This user feedback includes at least any combination of modification instructions, inclusion / exclusion of items, structural adjustment preferences, citation supplementation requirements, opinions on the selection or rejection of disputed points, and opinions on the adoption of unstructured evidence. The closed-loop update includes updating the task distribution strategy, conflict resolution parameters, evidence extraction rules, topic clustering strategy, output structure constraint template, and knowledge base management strategy, enabling subsequent generation to achieve adaptive optimization while maintaining consistency with the main methodology. For example, when users consistently prefer a certain research paradigm or type of evidence, the statistical summary dimensions and writing template can be adjusted; when users have stricter requirements for citation backtracking, the evidence unitization and citation anchor generation strategies can be adjusted; when users tend to be conservative in conflict handling, the arbitration output can be adjusted to include stronger uncertainty prompts and stricter evidence gating.

[0028] In addition to the methods described above, the present invention also provides apparatus and systems that are compatible with the methods to support the implementation and deployment of the methods in a computer environment.

[0029] This invention also provides a structured generation device for educational review papers based on multi-agent collaboration. The device is characterized by including at least: a configuration parsing module, a request parsing module, a task distribution module, an evidence processing module, a statistical summary module, a conflict resolution module, a structured writing module, and an output and feedback module. The modules are aligned and versioned through a shared workspace module, enabling the device to stably execute multi-stage collaborative processes.

[0030] The configuration parsing module is used to parse the agent configuration file and build a multi-agent collaborative workflow; the request parsing module is used to obtain the user's review generation request and form a request parameter set; the task distribution module is used to generate a task plan, form a task ticket set, and distribute it to the corresponding agent; the evidence processing module is used to perform textualization and evidence unitization processing on candidate items and unstructured data and establish citation anchor associations; the statistical summary module is used to perform topic clustering, distribution statistics, and comparative summary on the evidence unit set and output structured intermediate results; the conflict resolution module is used to trigger consistency verification, citation backtracking verification, and conflict detection and output conflict records and uncertainty prompts; the structured writing module is used to generate structured text results of the review paper under the output structure constraints and establish the association between text citations and evidence units; the output and feedback module is used to output results and optionally receive user feedback to form a closed-loop update.

[0031] This invention also provides a structured generation system for educational review papers based on multi-agent collaboration. The system comprises at least a server, a configuration storage component, a data access component, a model service component, a shared workspace component, and a client. The server hosts the device and provides interface services. The server communicates with each of the aforementioned components to execute the method and output structured results consistent with the agent configuration file. The configuration storage component stores the agent configuration file and output structure constraint templates. The data access component accesses unstructured data such as paper entry metadata, full-text, classroom recordings, interview recordings, observation records, or their transcribed texts, and writes source and time stamp identifiers. The model service component provides semantic representation computation, topic clustering and summary generation, evidence type annotation, logical consistency verification, and conflict detection reasoning capabilities. The shared workspace component stores request parameters, task tickets, evidence units, structured intermediate results, drafts, and conflict records, and supports versioning. The client receives user requests, displays structured text results and citation information, and submits user feedback.

[0032] The present invention also provides a computer-readable storage medium, characterized in that the storage medium stores a computer program, which, when run on a processor, causes the processor to execute the steps of the present invention's method for generating structured educational review papers based on multi-agent collaboration, or causes the processor to call various modules of the device to execute the method, and completes the following during execution: multi-agent collaborative workflow construction, request parsing and field mapping, task decomposition and task distribution, candidate item aggregation, evidence unitization processing, statistical summarization and structured induction, logical verification and conflict resolution, structured writing and citation association, and optional feedback closed-loop update.

[0033] Compared with the prior art, the present invention has the following beneficial technical effects and significant technical progress:

[0034] 1) It can simulate the division of labor in a real scientific research team through multi-agent collaboration, making the responsibility boundaries of each stage of retrieval, statistics, writing, and verification clear, thereby reducing the risk of structural drift and logical confusion in the generation of long texts by a single model; 2) It can accumulate intermediate products through task ticketing distribution and shared workspaces, making the generation process traceable, verifiable, and auditable, facilitating the verification of the generation basis; 3) It can integrate paper entries with unstructured data such as classroom videos and interview recordings into the evidence framework through evidence unitization and citation anchoring, and realize the backtracking of the source location of audio and video evidence. This enhances the credibility of educational research reviews under multi-source evidence conditions; 4) It can detect and arbitrate factual conflicts, citation conflicts, structural conflicts, and logical conflicts through logical verification and conflict resolution mechanisms, and output uncertainty prompts when a unique conclusion cannot be determined, thereby maintaining the coherence of the review argument while preserving the boundaries of disputes; 5) It can optionally receive user feedback to form a closed-loop update, and adaptively adjust the task distribution strategy, evidence extraction rules, conflict resolution parameters, and output structure constraints, so that the system maintains structural consistency and continuous optimization under different educational research scenarios and different user preferences. Attached Figure Description

[0035] Figure 1 This is a schematic diagram of the structured generation method for educational review papers according to the present invention.

[0036] Figure 2 This is a schematic diagram of the structure of the educational review paper generation system of the present invention. Detailed Implementation

[0037] To facilitate the description of the technical solutions of the embodiments of the present invention, some terms and symbols used in the present invention will be explained first. The values ​​of the following terms and symbols are all preset or configurable, and are not limited to fixed values.

[0038] 1) Multi-agent collaborative workflow: This refers to an execution process consisting of multiple agents with defined roles, along with support mechanisms such as task distribution, shared workspace, and conflict resolution. A multi-agent collaborative workflow can be represented as... .

[0039] 2) Literature Reviewer Agent: This refers to the agent used to generate structured text results for review papers, organize chapter levels, arrange paragraph descriptions, and organize citation tags in the text. It is denoted as... .

[0040] 3) Data Analyst Agent: This refers to an agent used to perform topic clustering, distribution statistics, comparative induction, and output structured intermediate results on the metadata of evidence units and candidate items. It is denoted as... .

[0041] 4) Logic Verifier Agent: This refers to the agent responsible for performing consistency checks, citation traceability checks, and conflict detection on the draft and structured intermediate results, and triggering conflict resolution output. It is denoted as... .

[0042] 5) Agent Configuration File: This refers to the configuration file used to define agent roles, user input fields, default value retrieval strategies, default fallback strategies, output structure constraints, tool invocation permissions, and communication endpoints, etc., denoted as... In one implementation, It can consist of multiple configuration files, which are used to define the configuration of sub-processes such as review topic input, search result organization, search question and answer, entry ID organization, knowledge base management and review writing.

[0043] 6) Request parameter set: This refers to the set of parameters generated from user input fields after parsing and mapping, used to drive subsequent task breakdown, retrieval, evidence extraction, statistical summarization, writing, and verification. It is denoted as... .

[0044] 7) Shared Workspace: This refers to the storage space used to store shared objects, version information, and state information during multi-agent collaboration, denoted as WS. WS can contain a collection of objects. With version collection .

[0045] 8) Candidate Item Set: This refers to the set of paper items obtained through retrieval and aggregation, along with other optional data items, denoted as... Candidate entries may contain a set of metadata fields. For example, fields such as title, author, year, venue, abstract, keyword, identifier, or subsets thereof.

[0046] 9) Evidence Unit: Refers to the smallest unit of evidence extracted and categorized from the textualized results of candidate entries, full text, abstracts, and unstructured data such as classroom recordings and interview recordings. It is denoted as... The set of evidence units is denoted as... .

[0047] 10) Task Ticket: Refers to a standardized task description object used for task breakdown and distribution, denoted as... The set of task tickets is denoted as... .

[0048] 11) Result Package: Refers to the standardized result object output by the agent after executing a task ticket, denoted as... The resulting set of packages is denoted as .

[0049] 12) Message Encapsulation: Refers to the outer encapsulation object of messages exchanged between multiple agents, denoted as... It is used to carry session identifier, version identifier, sender, receiver, message type, and payload content.

[0050] 13) Conflict Log: Refers to the traceable object of logic verification and conflict resolution output, denoted as... The set of conflict records is denoted as... Conflict types include at least factual conflicts, citation conflicts, structural conflicts, and logical conflicts.

[0051] 14) Uncertainty Warning: This field indicates a warning that is output when there is insufficient evidence, inconsistent sources, failure of retrospective investigation, or when arbitration cannot uniquely determine the issue. It is denoted as... This is a placeholder description and does not specify a specific value.

[0052] The present invention will be further described below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0053] Example 1

[0054] See Figure 1 A structured generation method for educational review papers based on multi-agent collaboration is proposed. The method includes steps S1 to S9 and forms a closed-loop processing link oriented towards user feedback. The inputs / outputs of each step are aligned in the shared workspace WS, and multi-agent collaboration is realized through task ticketing mechanism and communication protocol.

[0055] In this embodiment, user input includes at least the overview topic field. Time range field With research type field In another implementation, user input may also include structural constraint fields. Reference style fields Include exclusion rule fields And access fields for unstructured data such as classroom recordings and interview recordings. The required fields, default value selection strategy, and default fallback strategy are determined by the agent configuration file. definition.

[0056] Step S1: Construct a multi-agent collaborative workflow

[0057] Building a multi-agent collaborative workflow based on agent configuration files specifically includes:

[0058] Step 1-1: Load configuration and parse character set

[0059] The system reads from the storage component The set of characters is obtained through parsing. , which includes at least , and The set of roles It can be expressed as follows:

[0060] .

[0061] in, Suitable for structured writing and citation organization in review articles; Suitable for evidence summarization, statistical analysis, and topic clustering; It is suitable for consistency checks, reference backtracking checks, and conflict detection.

[0062] Step 1-2: Parse the user input field set and field constraints Define a set of user input fields , At least includes: for characterization

[0063] Theme fields Time range field used to characterize the scope of retrieval and inclusion. and .

[0064] The system The system reads the required field flags (req_{flag}), default values ​​(def_{val}), maximum length (len_{max}), default fallback policy (fb_{policy}), and hint information (hint_{txt}) item by item. For missing required fields, it triggers either the default fallback policy or interactive completion. For missing non-required fields, it writes a default value or an empty placeholder and records the source of the default value.

[0065] Steps 1-3: Parse the output structure constraints and writing template

[0066] The `Cfg_{agent}` class defines the output structure constraint `out_{schema}`. `out_{schema()*` is used to ensure that the review output includes at least the following paragraph slots: title, abstract, introduction, hierarchical review body, points of contention, research gaps and future directions, methodological suggestions, and a list of references. It also constrains the slot order, slot selectability, and default description method. The system writes `out_{schema}` to the shared workspace `WS` as a hard constraint input for subsequent writing stages.

[0067] Steps 1-4: Initialize shared workspace and version identifier

[0068] The system initializes the shared workspace WS, creating a session identifier sid, a version identifier ver_{0}, and a timestamp t_{0}. WS contains pre-defined sets of shared objects, including: request parameter set Req_{set}, task ticket set Task_{set}, candidate entry set Cand_{set}, evidence unit set EU_{set}, structured intermediate result set Mid_{set}, draft set Draft_{set}, conflict record set Conf_{set}, and result package set Result_{set}.

[0069] In one implementation, the collection of objects WS_{obj} in the shared workspace can be described as follows:

[0070] WS_{obj} = { Req_{set}, Task_{set}, Cand_{set}, EU_{set}, Mid_{set},Draft_{set}, Conf_{set}, Result_{set}}.

[0071] Each shared object is associated with a sid and ver_{k} to record the evolution of the artifacts at each stage within the same session.

[0072] Steps 1-5: Initialize the communication protocol and endpoints

[0073] The `Cfg_{agent}` class defines the communication endpoint `Endpoint_{set}` and the `Message_{env}` field set for the communication message encapsulation. `Message_{env}` must contain at least: `sid`, version identifier `ver`, timestamp `t`, sender identifier `sender_{id}`, receiver identifier `receiver_{id}`, message type `msg_{type}`, and payload content. The `Message_{env}` field set can be described in the following form:

[0074] Message_{env} = ⟨ sid, ver, t, sender_{id}, receiver_{id}, msg_{type}, payload >.

[0075] The payload is used to carry a Task_{ticket} or a Result_{pack} or other shared object reference.

[0076] The system registers the sender identifier sender_{id} for each agent and opens its receiving endpoint, enabling the literature reviewer agent A_{review}, the data analyst agent A_{data}, and the logic checker agent A_{logic} to exchange tasks and results in a unified encapsulation format.

[0077] Steps 1-6: Initialize the task distribution module and conflict resolution module

[0078] The task distribution module is used to generate task plans, break down Task_{set}, schedule task dependencies, and perform retries and degradation. The conflict resolution module is used to classify, aggregate, and arbitrate conflict records Conf_{rec}. All of the above modules write the state_{flag} and log_{txt} fields into the WS to support auditing and review.

[0079] Through step S1, the system completes the construction of a multi-agent collaborative workflow, enabling subsequent steps to be executed under unified field constraints, unified message protocols, and a unified shared space.

[0080] Step S2: Obtain the review request and map it to a set of request parameters.

[0081] Obtain the user's summary generation request and parse and map the user input fields according to the agent configuration file, specifically including:

[0082] Step 2-1: Receive user input and form the original input object

[0083] Users input the review topic (review_{scope}) into the system, and can also input the time range (time_{window}) and study type (study_{types}). The system aggregates the user input into a raw input object (User_{in}) represented by the following formula:

[0084] User_{in} = { review_{scope}, time_{window}, study_{types}, …}.

[0085] The “…” indicates optional fields, such as the output structure constraint field out_{schema}, the inclusion exclusion rule field inc_{exc}, the citation style field cite_{style}, and the access field data_{source} for unstructured data.

[0086] Step 2-2: Field Validation and Default Rollback

[0087] The system performs field-by-field validation on the original input object User_{in}. If review_{scope} is empty or does not meet the maximum length len_{max} constraint, the default fallback policy fb_{policy} is executed. If time_{window} is empty, a default time window is written or it is marked as "undefined". If study_{types} is empty, a default set of study types is written or it is marked as "any study type". The system writes the default source to the field metadata meta_{in}, for example, meta_{in}[review_{scope}].source = "user" or "default".

[0088] Steps 2-3: Field Mapping and Parameter Derivation

[0089] The system maps input fields to a request parameter set Req_{set} based on Cfg_{agent}. Req_{set} includes at least: query generation parameter q_{gen}, filtering parameter f_{filter}, inclusion / exclusion parameter r_{inc}, evidence expiration parameter t_{evi}, output structure parameter o_{schema}, and reference parameter c_{rule}. Req_{set} can be expressed as follows:

[0090] Req_{set} = Map( User_{in}, Cfg_{agent} ).

[0091] Where Map() is a field mapping function, the output of which satisfies: — q_{gen} is derived from review_{scope} and its synonym extension set syn_{set}; — f_{filter} is derived from time_{window}; — r_{inc} is derived from study_{types}; — o_{schema} is defined by Cfg_{agent} or overridden by user input; — c_{rule} is defined by Cfg_{agent} or overridden by user input.

[0092] Steps 2-4: Write to the shared workspace and generate session-level parameter snapshots

[0093] The system writes the request parameter set Req_{set} into WS and generates a version snapshot ver_{1}, recording the field values ​​of Req_{set}, the default fallback situation, and the mapping strategy identifier, which is convenient for subsequent writing output explaining "why a certain range and a certain structure are used".

[0094] In step S2, the system transforms the user input from a "natural language request" into an "executable set of request parameters" and provides a unified entry point for subsequent task breakdown, distribution, and alignment.

[0095] Step S3: Task plan generation, breakdown, and ticket-based distribution

[0096] The task distribution module generates a task plan based on the request parameter set, and decomposes the summary-generated task into a set of task tickets, specifically including:

[0097] Step 3-1: Generate a set of task types and determine the execution role mapping.

[0098] The system generates a task type set `Type_{set}` based on `Req_{set}`. `Type_{set}` includes at least: `search`, `organize`, `extract`, `unit`, `summ`, `write`, and `check`. The system then maps task types to roles: — `search` and `organize` are preferentially assigned to `A_{review}` or the retrieval sub-agent; — `extract` and `unit` can be assigned to `A_{review}` or the evidence processing sub-agent; — `summ` is assigned to `A_{data}`; — `check` is assigned to `A_{logic}`; — `write` is assigned to `A_{review}`.

[0099] Step 3-2: Generate task tickets and write dependencies.

[0100] The system generates a Task_{ticket} for each task type. The Task_{ticket} contains at least: task_{id}, task_{type}, input_{param}, output_{spec}, dep_{id}, retry_{policy}, and fb_{policy}, and is described by the following formula:

[0101] Task_{ticket} = ⟨ task_{id}, task_{type}, input_{param}, output_{spec}, dep_{id}, retry_{policy}, fb_{policy} >

[0102] Where: task_{id} is the unique identifier of the task; dep_{id} is used to describe the set of dependent tasks, such as statistical summary depending on evidence unitization, and writing depending on statistical summary and conflict record; retry_{policy} defines the timeout retries and the maximum number of retries; fb_{policy} defines the degradation strategy, such as relaxing the query or switching the search source when the retrieval fails.

[0103] Step 3-3: Create a task plan and schedule and prioritize tasks.

[0104] The system generates a task plan Plan_{task}, which contains a directed dependency graph G_{dep}. The system performs topological sorting on G_{dep} to obtain a task execution sequence Seq_{task} and a set of parallelizable tasks Par_{set}. For parallelizable tasks, the system controls the concurrency based on resource constraints res_{cap} to avoid write conflicts in the shared workspace and congestion in the model service.

[0105] Steps 3-4: Distribute task tickets using communication protocols

[0106] The system encapsulates Task_{ticket} into Message_{env} and sends it to the corresponding agent. For A_{review}, A_{data}, and A_{logic}, the message type msg_{type} can be "task". For example:

[0107] Message_{env} = ⟨ sid, ver_{1}, t_{1}, sender_{id}=Dispatcher, receiver_{id}=A_{data}, msg_{type}=task, payload=Task_{ticket} >

[0108] After receiving the data, the agent writes a "received" status to the WS and writes back the result packet after execution.

[0109] Steps 3-5: Task Failure Detection and Rollback Trigger

[0110] When a task times out or its output does not meet the `output_{spec}` constraint, the task distribution module retryes according to `retry_{policy}`. If the retry fails, a fallback path is executed according to `fb_{policy}`. For example, when a retrieval task fails, the query is broadened, synonyms are expanded, the retrieval source is switched, or a reference chain is used to trace the retrieval; when a statistical summary task fails, it is changed to rule-based summary or segmented summary before merging. All of the above fallback processes are written to the WS log fields for later explanation.

[0111] Through step S3, the system transforms the review generation from "one-time long text generation" to "multi-stage traceable task execution," enabling each step's output to be independently inspected and reused.

[0112] Step S4: Document retrieval, candidate item aggregation, and metadata standardization

[0113] Literature retrieval and candidate item aggregation are performed by a literature reviewer agent and / or retrieval agents that collaborate with it, specifically including:

[0114] Step 4-1: Generate search query and execute search.

[0115] A_{review} generates a search query_{txt} based on q_{gen} in Req_{set}. The query_{txt} can consist of core keywords and extended keywords extracted from review_{scope}, and can be combined with time_{window} and study_{types} to form filtering conditions. The system executes the search and obtains an initial candidate item set Cand_{raw}. Cand_{raw} can contain paper entries and other resource entries.

[0116] Step 4-2: Item deduplication and field normalization

[0117] The system performs deduplication on Cand_{raw}, with deduplication rules based on the identifier field (e.g., doi or other identifiers) or on title similarity. Title similarity can be determined by semantic vector similarity s_{title} or string similarity sim_{str}. For each entry, the system standardizes field names and formats, forming a standardized candidate entry set Cand_{set}.

[0118] Step 4-3: Write candidate entry metadata records and source identifiers

[0119] For each item_j} in Cand_{set}, write the source identifier source_{id} and the retrieval timestamp t_{retr}. The metadata of item_j} is represented by the following formula:

[0120] Item_{j}.Meta = { title_{j}, author_{j}, year_{j}, venue_{j}, abstract_{j}, keyword_{j}, identifier_{j}, source_{id}, t_{retr}}.

[0121] The system writes Cand_{set} into WS and generates version ver_{2} for subsequent evidence extraction and citation organization.

[0122] Step 4-4: Unstructured Data Access

[0123] When Req_{set} contains a data_{source} field, the system accesses classroom recordings, interview recordings, observation records, or their transcribed text, forming an unstructured candidate evidence set U_{set}. The source identifier source_{u} and the location index rule index_{rule} are written to U_{set}. The location index rule is used to subsequently trace the evidence units back to the time range of the audio / video recordings or the paragraph range of the transcribed text.

[0124] Through step S4, the system obtains a set of structured candidate entries and a set of optional unstructured candidate evidence, providing a unified input for evidence extraction and unitization.

[0125] Step S5: Evidence extraction, evidence unitization, and association with citation anchors

[0126] Evidence extraction and evidence unitization are performed on candidate entries and candidate evidence sources to form an evidence unit set EU_{set}, which specifically includes:

[0127] Step 5-1: Entry Text Acquisition and Preprocessing

[0128] For each item Item_j in Cand_{set}, the system obtains its abstract_j and optional full text_j. The system performs text preprocessing, including paragraph segmentation, sentence group segmentation, noise removal, and encoding standardization. For unstructured data U_{set}, the system obtains the transcribed text trans_u or performs automatic speech recognition to obtain asr_u, while preserving the timeline alignment information time_align.

[0129] Step 5-2: Extraction and Type Labeling of Evidence Fragments

[0130] The system extracts evidence fragments (seg_k) from abstract_{j}, full_{j}, and trans_{u}. Evidence fragments can be extracted according to rules or models. Each seg_{k} is labeled with an evidence type (type_evi), which includes at least the following: fact_{seg}, method_{seg}, result_{seg}, limit_{seg}, and dispute_{seg}. These type labels are used for subsequent summarization and writing slot mapping.

[0131] Step 5-3: Evidence Unitization and Location Marker Generation

[0132] The system organizes seg_{k} into evidence units EU_{i}. Each EU_{i} contains at least text content text_{i}, evidence type type_{i}, source identifier source_{id}, location identifier pos_{id}, time identifier time_{id}, and citation anchor_{anchor}. It can be described in the following form:

[0133] EU_{i} = ⟨ text_{i}, type_{i}, source_{id}, pos_{id}, time_{id},cite_{anchor} >.

[0134] For paper entries, pos_{id} can represent a paragraph number or page number range as a placeholder; time_{id} can be empty or take t_{retr}; for audio and video transcription, pos_{id} can represent the transcription paragraph number; time_{id} can represent the audio and video time range as a placeholder, such as [ t_{start}, t_{end} ]; cite_{anchor} is the citation anchor mark, used to associate arguments with evidence units during the writing stage.

[0135] Step 5-4: Research Feature Labeling and Slot Mapping Label Generation

[0136] The system further generates research element tags (tag_{elem}) from EU_{i}. These tags cover at least any combination of elements such as research question, research subject, research method, variable indicators, sample context, and research conclusions and limitations. Research element tags are used to map evidence units to fixed slots in the writing structure, such as "methodological comparison paragraph," "contextual difference paragraph," and "points of contention paragraph."

[0137] Step 5-5: Write the evidence unit set to the shared workspace and create an index.

[0138] The system writes EU_{set} into WS and creates an index Index_{EU}. Index_{EU} supports at least the following: searching by topic tag, filtering by research type, filtering by time range, filtering by evidence type, and tracing back by source identifier. The system generates version ver_{3}, ensuring that subsequent statistical summaries, logical verifications, and writing all use EU_{set} as a unified evidence base.

[0139] Through step S5, this invention transforms "literature and unstructured data" into a unified set of evidence units that are "citationable, traceable, and statistically verifiable," enabling review writing to no longer rely solely on model generation but to be evidence-driven and traceable.

[0140] Step S6: Statistical summary, topic clustering, and generation of structured intermediate results

[0141] The data analyst agent performs statistical summarization and structured generalization of the evidence unit set and candidate item metadata, specifically including:

[0142] Step 6-1: Topic Representation and Similarity Calculation

[0143] A_{data} generates a semantic representation emb_{i} for EU_{set} and calculates the similarity s_{i,j} between evidence units. In one implementation, cosine similarity can be expressed as follows:

[0144] s_{i,j} = cos( emb_{i}, emb_{j} ).

[0145] Here, emb_{i} and emb_{j} are vector representations, and cos(·) is the cosine similarity function. Similarity is used for topic clustering and merging duplicate evidence.

[0146] Step 6-2: Topic Clustering and Topic Cluster Naming

[0147] Data A clusters the evidence units according to s_i,j, forming topic clusters Cluster_k. Each Cluster_k contains a subset of evidence units EU_set^(k)}. Data A generates a topic name topic_k and a keyword set key_k for each Cluster_k. The topic_k is used to generate chapter and section titles during the writing phase.

[0148] Step 6-3: Statistical analysis of the distribution of research paradigms, research types, and methodological elements.

[0149] A_{data} statistically analyzes the distribution of research paradigms (Dist_{type}) and methodological elements (Dist_{method}) based on study_{types} and evidence tags (tag_{elem}). Dist_{type} can be represented as a "placeholder description of the proportion of each research type" without specifying a concrete value; Dist_{method} is used to summarize common methodological paths, common data sources, and common analytical frameworks.

[0150] Step 6-4: Summary of Variable Indicators, Sample Context, and Conclusion Disputes

[0151] A_{data} generates an indicator summary table Ind_{summ} for evidence units labeled with variable indicators in EU_{set}, a situation summary table Ctx_{summ} for evidence units labeled with sample situation, and a dispute point set Dispute_{set} for evidence units labeled with conclusion and controversy. The dispute point set includes at least the controversial topic, a summary of opposing viewpoints, a corresponding set of evidence citations ref_{evi}, and an uncertainty hint U_{hint}.

[0152] Step 6-5: Form a structured set of intermediate results and establish evidence associations.

[0153] A_{data} outputs a structured intermediate results set Mid_{set}, which includes at least: topic clustering results, research paradigm distribution results, summary results of methodological elements, summary results of variable indicators, summary results of sample context, and summary results of conclusions and points of contention. Each conclusion object Mid_{m} in Mid_{set} is associated with a supporting evidence set Ref_{m}.

[0154] Mid_{m}.Ref = { cite_{anchor} | EU_{i} supports Mid_{m}}.

[0155] Among them, supports() represents "evidence support relationship", which is determined by tag matching, topic cluster affiliation and content coverage rules.

[0156] Step 6-6: Write to the shared workspace and output the result package

[0157] A_{data} writes Mid_{set} to WS and returns it as Result_{pack}:

[0158] Result_{pack} = ⟨ task_{id}=type_{summ}, status, content=Mid_{set},evidence_{refs}, conflict_{flags}, U_{hint} >

[0159] Among them, evidence_{refs} must contain at least the set of Ref_{m} corresponding to the key Mid_{m}; conflict_{flags} can be empty or contain statistical anomaly prompts, and the system generates ver_{4} version based on this, providing input for logic verification and writing.

[0160] Through step S6, the system obtains structured intermediate results that can be directly used for writing structure filling and comparative argumentation, and each intermediate conclusion comes with a set of evidence citations, enabling subsequent writing to "first evidence and then narration".

[0161] Step S7: Consistency check, reference backtracking check and conflict resolution

[0162] The logic checker agent performs consistency checks and conflict detection on the structured intermediate results and the draft, and the conflict resolution module arbitrates and outputs conflict records, specifically including:

[0163] Step 7-1: Consistency check triggering and check input preparation

[0164] A_{logic} reads Mid_{set} and Draft_{set} from WS (if Draft_{set} has not yet been generated, A_{logic} can first perform a "structural consistency pre-check" on Mid_{set}). The structural consistency pre-check checks whether Mid_{set} meets the chapter slot coverage requirements of out_{schema}, such as whether objects like "set of points of contention" or "future direction suggestions" exist. If the coverage is insufficient, a structural gap hint is output and written to U_{hint}.

[0165] Step 7-2: Refer to traceability verification

[0166] A_{logic} performs reference backtracking checks on each Mid_{m}.Ref in Mid_{set}, verifying whether cite_{anchor} can locate EU_{i}, and whether EU_{i} can locate source_{id} and pos_{id} or time_{id}. When backtracking fails, a reference conflict Conf_{rec}^{cite} is generated, and the conflict location and the reason for backtracking failure are recorded.

[0167] Step 7-3: Detection of Conflicts and Disputes

[0168] Within the same topic cluster (Cluster_k), A_{logic} performs a consistency check on the conclusion summaries. If contradictory or mutually exclusive conclusion fragments exist, a fact conflict Conf_{rec}^{fact} is generated. A fact conflict record includes at least: the conflicting topic_{k}, opposing conclusion summaries, their respective evidence citation sets Ref_{a} and Ref_{b}, and an arbitration recommendation.

[0169] Step 7-4: Detection of structural and logical conflicts

[0170] When Draft_{set} exists, A_{logic} compares the draft chapter hierarchy with out_{schema}. If chapter order, hierarchy depth, or paragraph slots are missing, a structural conflict Conf_{rec}^{struct} is generated. A_{logic} performs a coherence check on the reasoning chains between draft paragraphs. If missing arguments, jumps in reasoning, or inconsistencies in definitions occur, a logical conflict Conf_{rec}^{logic} is generated.

[0171] Step 7-5: Gathering Conflicting Evidence and Arbitration Output

[0172] The conflict resolution module performs classification and evidence aggregation on Conf_{rec}. Evidence aggregation refers to pulling the set of evidence units EU_{conf} associated with the conflict from EU_{set} and calculating the source coverage summary cov_{src} and the timeliness consistency summary cov_{time}. The arbitration includes the following strategies: 1) Source consistency priority: If a party's evidence has consistent sources and more comprehensive coverage, then that party's conclusion is retained first; 2) Evidence coverage gating: If the evidence coverage of both parties is insufficient, then parallel viewpoints are output and "insufficient evidence" is indicated; 3) Timeliness consistency filtering: If the conflict originates from different time periods, then "time period conclusions" are output and the applicable boundaries are indicated; 4) Citation traceability gating: If a party's citation traceability fails, then it is downgraded or marked as uncertain.

[0173] The arbitration output forms a conflict record Conf_{rec}, which contains at least: conf_{type}, conf_{loc}, Ref_{conf}, rule_{id}, decision_{txt}, and U_{hint}.

[0174] Step 7-6: Write conflict logs to the shared workspace and generate versions.

[0175] The system writes Conf_{set} into WS and generates version ver_{5}. The subsequent writing stage will read Conf_{set} to determine how to express the controversy, how to indicate the uncertainty, and how to point out the research opportunities where "the controversy has not yet been resolved" in the review text.

[0176] Through step S7, the present invention explicitly transforms problems such as "inconsistency between different parts, untraceable citations, and conflicting conclusions" in long text generation into conflict records, enabling the system to retain boundaries in the output and support auditing.

[0177] Step S8: Structured Writing, Chapter Filling, and Citation Chain Generation

[0178] The literature reviewer agent generates structured text results of the review paper based on a set of structured intermediate results, conflict records, and output structure constraints, specifically including:

[0179] Step 8-1: Generate the structural skeleton of the review paper

[0180] A_{review} reads out_{schema} and generates the chapter skeleton Outline_{doc}. Outline_{doc} contains at least a title paragraph, abstract paragraph, introduction paragraph, body structure paragraph, controversy and comparison paragraph, research gaps and future directions paragraph, methodological suggestions paragraph, and references paragraph. The chapter skeleton also includes input slot descriptions (slot_{desc}) for each paragraph, specifying which Mid_{m} objects and evidence types should be cited.

[0181] Step 8-2: Generation of the main hierarchical structure and mapping of topic clusters

[0182] A_{review} maps the topic clusters_{k} in Mid_{set} to sections of the main chapters. For each topic_{k}, a section title is generated, and the section content is populated, including an overview of the research question, an overview of commonly used methods, a summary of key conclusions, a description of contextual differences, and a summary of limitations. During the population process, the evidence set with more comprehensive evidence coverage in Mid_{m}.Ref is prioritized, and for topics involving conflict, decision descriptions and U_{hint} hints from Conf_{rec} are inserted.

[0183] Step 8-3: Generation of Comparison and Controversial Paragraphs

[0184] A_{review} reads the Dispute_{set} and Conf_{set} to generate comparison and controversy paragraphs. The comparison paragraph should include at least a summary of opposing viewpoints, an overview of the sources of evidence for each, and a description of the reasons for the differences. The controversy paragraph should include at least a description of the unresolved boundaries of the controversy and possible research design suggestions, enabling readers to understand "why the controversy exists" and "how subsequent research can bridge it."

[0185] Step 8-4: Research Gaps and Future Directions

[0186] A_{review} generates a research gap description Gap_{txt} and future direction suggestions Future_{txt} based on information such as "limitations description fragments," "weak evidence coverage topics," and "unresolved conflict topics" in Mid_{set}. The research gap description must at least include the gap type (insufficient evidence, context gap, methodological gap, indicator gap, etc.) and a set of supporting evidence anchors. Future direction suggestions must at least include feasible research path recommendations, such as what type of unstructured data to collect, what research paradigm to adopt, or in what educational context to validate the research.

[0187] Step 8-5: Citation Chains and Reference Organization

[0188] For each key argument, A_{review} inserts a cite_{anchor} and maps the cite_{anchor} to the source identifier in Item_{j}.Meta or U_{set}, forming a citation chain Link_{cite}. The reference section outputs at least a list of entries and can also output a mapping of "citation anchor - source identifier - location identifier" for backtracking without limiting the specific front-end format. The citation chain mapping relationship is expressed by the following formula:

[0189] Link_{cite}( cite_{anchor} ) → ⟨ source_{id}, pos_{id}, time_{id} >.

[0190] Among them, pos_{id} and time_{id} are placeholder range descriptions and do not require a fixed format.

[0191] Step 8-6: Generate structured text results and write them to the shared workspace

[0192] A_{review} outputs a structured text result Doc_{out}, which includes at least the main text and reference chain fields. A_{review} writes Doc_{out} to WS and generates version ver_{6}, while also sending back Result_{pack} for task closure.

[0193] Through step S8, the present invention achieves a review writing output that is "structure-first, evidence-driven, conflict-visible, and citation-backtrackable", so that the generated result is no longer an uninterpretable long text, but a structured result with a chain of evidence.

[0194] Step S9: Output and Feedback Closed-Loop Update

[0195] Output structured text results, and optionally receive user feedback to form a closed-loop update, specifically including:

[0196] Step 9-1: Output Results and Uncertainty Hints

[0197] The system outputs Doc_{out} to the user and U_{hint} at appropriate locations. When a conflict record Conf_{rec} exists, the system displays "Dispute exists" in the corresponding paragraph and provides evidence backtracking anchors to facilitate user verification.

[0198] Step 9-2: Receive user feedback and structure it.

[0199] Users can submit feedback (Feedback_{u}), which includes at least any combination of the following: modification instructions, inclusion / exclusion items, structural adjustment preferences, supplementary citation requests, opinions on disputed choices, and opinions on the adoption of unstructured evidence. The system will parse the feedback into a structured feedback object (FB_{set}) and write it to the WS.

[0200] Step 9-3: Update task distribution strategy and parameters

[0201] The system updates task distribution parameters based on FB_{set}, such as concurrency level res_{cap}, retry count n_{retry}, and task priority rule prio_{rule}. When a user explicitly requests "strengthening unstructured evidence," the priority of evidence unitization tasks is increased or the scope of audio and video transcription is expanded; when a user explicitly requests "strict citation backtracking," the citation backtracking gating strength is increased, triggering a secondary search of failed backtracking entries.

[0202] Step 9-4: Update conflict resolution parameters and evidence gating parameters

[0203] The system updates the conflict arbitration strategy weights or gating rules based on FB_{set}. For example, when the user tends to be conservative, the conflict arbitration output emphasizes parallel viewpoints and U_{hint}; when the user tends to be certain, the system increases the trigger probability of the source consistency priority rule, but at the same time retains boundary hints in the output.

[0204] Step 9-5: Update the knowledge base management strategy

[0205] When the system includes a knowledge base management process, the knowledge base writing strategy can be adjusted based on FB_{set}, such as adding new entries, marking entry status, and revising evidence unit tags. The entry weights can be updated using the following placeholder method:

[0206] w_{item_{new}} = λ · w_{item_{old}} + (1-λ) · Δw_{item}.

[0207] Where w_{item_{old}} is the item weight before the update, w_{item_{new}} is the item weight after the update, λ is the placeholder parameter for time decay or stabilization factor, and Δw_{item} is the placeholder amount for gain or decay triggered by user feedback. The above expression is used to illustrate the form of the update rule and does not limit the specific values.

[0208] Through step S9, this invention supports a closed-loop user feedback mechanism without altering the main workflow, enabling the system to gradually adapt to user preferences and research scenario constraints during repeated use. To ensure the feasibility and traceability of the multi-agent collaboration process, this embodiment further provides exemplary descriptions of communication protocols and task objects. It should be understood that the following description is one implementation method and does not limit the unique form.

[0209] 1) Example of the Message_{env} field in the message encapsulation object: Message_{env} = ⟨ sid, ver, t, sender_{id}, receiver_{id}, msg_{type}, payload >. Here, msg_{type} can be task, result, ack, error, etc.; the payload carries Task_{ticket} under the task type and Result_{pack} under the result type.

[0210] 2) Example of the Task_{ticket} field in the Task Ticket object: Task_{ticket} = ⟨ task_{id}, task_{type}, input_{param}, output_{spec}, dep_{id}, retry_{policy}, fb_{policy} >. The output_{spec} field can contain a set of output fields and structural constraints, such as "must return a list of candidate entries" or "must return a set of evidence units with source identifiers," etc.

[0211] 3) Example of the Result_{pack} field in the Result Pack object:

[0212] Result_{pack} = ⟨ task_{id}, status, content, evidence_{refs}, conflict_{flags}, U_{hint} >. Here, evidence_{refs} carries the set of reference anchors, conflict_{flags} carries conflict warning flags, and U_{hint} carries uncertainty warnings.

[0213] Through the above object definition, the collaboration between multiple agents does not rely on implicit dialogue memory, but on persistent task and result objects, which facilitates auditing and review.

[0214] Example 2

[0215] A structured generation device for educational review papers based on multi-agent collaboration is disclosed. The device can be deployed on a server or terminal device. The device includes a processor and a computer-readable storage medium. The processor is used to execute instructions in the storage medium to implement the above-described method steps.

[0216] The device includes: a configuration parsing module, a request parsing module, a task distribution module, an evidence processing module, a statistical summary module, a conflict resolution module, a structured writing module, and an output and feedback module. The above modules exchange objects and version identifiers through a shared workspace WS, thereby ensuring the consistency and traceability of the data flow within the device.

[0217] The configuration parsing module loads Cfg_{agent}, constructs a multi-agent collaborative workflow Flow_{MA}, and initializes the Message_{env} field set and shared workspace WS. The request parsing module obtains user input review_{scope}, time_{window}, and study_{types} and forms a Req_{set}, and performs field validation and default fallback. The task distribution module generates a Task_{set} and writes it into the task dependency graph G_{dep}, and distributes the task tickets to A_{review}, A_{data}, and A_{logic} for execution according to the mapping rules. The evidence processing module extracts evidence fragments from Cand_{set} and U_{set} and forms an EU_{set}, and writes source to EU_{i}. The system includes modules for classifying and arbitrating Conf_rec, generating Conf_set and U_hint. The system also includes modules for generating Outline_doc based on out_schema and Doc_out based on Mid_set and Conf_set, and establishing a Link_cite citation chain. The output and feedback module outputs Doc_out and receives Feedback_u, mapping the feedback to parameter updates to form a closed loop.

[0218] Example 3

[0219] See Figure 2 A structured generation system for educational review papers based on multi-agent collaboration is disclosed, comprising a server, a configuration storage component, a data access component, a model service component, a shared workspace component, and a client. The server hosts the device and provides interface services, receiving user requests and returning Doc_{out}. The server communicates with the configuration storage component, data access component, model service component, shared workspace component, and client to execute the aforementioned methods and output a result structure consistent with Cfg_{agent}. The configuration storage component stores structural constraint templates such as Cfg_{agent} and out_{schema}. The data access component accesses paper entry metadata, full-text text, classroom recordings, interview recordings, or other data. The system transcribes unstructured data such as text and writes source_{id} and timeliness identifier; the model service component provides semantic representation emb_{i} calculation, topic clustering, summary generation, logical consistency verification, and conflict detection reasoning capabilities, and returns corresponding reasoning results and placeholder confidence information to the server; the shared workspace component stores Req_{set}, Task_{set}, Cand_{set}, EU_{set}, Mid_{set}, Draft_{set}, and Conf_{set} and supports versioning and rollback, enabling the system to review the evidence and arbitration process corresponding to a certain version output; the client receives user input, displays structured text results and citation information, and submits user feedback.

[0220] Example 4

[0221] A computer-readable storage medium storing a computer program, which, when executed on a processor, causes the processor to execute the steps of the present invention's method for generating structured educational review papers based on multi-agent collaboration, or causes the processor to call various modules of the device to execute the method, and in the process of execution, completes the construction of multi-agent collaborative workflow, request parsing and field mapping, task decomposition and task distribution, candidate item aggregation, evidence unitization processing, statistical summarization and structured induction, logical verification and conflict resolution, structured writing and citation association, and optional feedback closed-loop update.

[0222] The above embodiments are methods, apparatus, systems, and computer program products (computer-readable storage media). Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-readable storage media containing computer-readable program code, including but not limited to disk storage, optical storage, etc. In the description of the present invention, it should be noted that the terms are used for descriptive purposes only and should not be construed as indicating or implying relative importance; the terms "comprising" or "including" are intended to indicate the presence of the listed elements or steps, but do not exclude the presence of other elements or steps; the term "connection" can be either a direct connection or an indirect connection.

[0223] The above are merely preferred embodiments of the present invention and do not limit the patent scope of the present invention. Those skilled in the art should understand that any equivalent structural or procedural transformations made based on the contents of the present invention specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the protection scope of the present invention.

Claims

1. A method for generating structured review papers in education, characterized in that, A multi-agent collaborative approach, incorporating literature review, data analysis, and logical verification, is employed to generate structured review papers in education. This approach specifically includes the following steps: S1. Constructing a multi-agent collaborative workflow A multi-agent collaborative workflow is constructed based on agent configuration files. Each agent is configured with its input field set, output structure constraints, tool call permissions, and communication endpoints. The multi-agents include: a literature reviewer agent, a data analyst agent, and a logic checker agent. The multi-agent collaborative workflow includes: a task distribution module, a shared workspace module, and a conflict resolution module. The shared workspace module is a shared object that stores a set of request parameters, a set of task tickets, a set of candidate entries, a set of evidence units, a set of structured intermediate results, a set of writing drafts, and a set of conflict records. Session identifiers, version identifiers, and timestamp identifiers are associated with the shared objects. S2. Obtain the review request and map it to a set of request parameters. The user's review generation request is parsed and validated according to the agent configuration file, and mapped to a set of request parameters. The user input fields include: a topic field representing the review topic, a time range field representing the search and inclusion scope, a research type field representing the included research paradigm, and a structural constraint field representing the output structure and citation requirements. The agent configuration file determines the mandatory fields, default value strategies, default fallback strategies, field conflict handling strategies, and output structure constraints for each field. The set of request parameters drives the generation of search queries, search filtering, execution of inclusion and exclusion rules, evidence timeliness screening, output chapter level constraints, and execution of citation presentation rules. S3, Task plan generation, breakdown, and ticket distribution The task distribution module generates a task plan based on the request parameter set and decomposes the summary generation task into a task ticket set. The task ticket set is any combination of retrieval task tickets, candidate item organization task tickets, evidence extraction task tickets, evidence unitization task tickets, statistical summary task tickets, structured writing task tickets, and logic verification task tickets. Each task ticket contains at least a task type identifier, input parameter description, expected output description, dependency identifier, priority or deadline strategy identifier, and fallback strategy identifier. S4. Document retrieval, candidate item aggregation, and metadata standardization The literature reviewer agent and its collaborating retrieval agents perform literature retrieval and candidate item aggregation, and standardize the metadata of the candidate items. The candidate items are any combination of paper items, project materials, conference materials, policy documents, course resources, research data description information, and practice reports. The standardization of the metadata includes: field normalization, time field unification, source identifier writing, and duplicate item removal, to form a candidate item set. S5. Evidence Extraction, Evidence Unitization, and Citation Anchor Point Association Evidence extraction and unitization are performed on candidate entries and candidate evidence sources to form a set of evidence units. This set of evidence units can be any combination of factual description fragments, methodological description fragments, conclusive description fragments, limitation description fragments, and controversial description fragments. These units are written into a shared workspace module and associated with candidate entries, research type fields, and time range fields. The unitization process performs paragraph segmentation, sentence group segmentation, or semantic unit segmentation on text sources and time window segmentation or transcription fragment segmentation on audio and video sources. Each evidence unit is associated with a source identifier, location identifier, time identifier, and citation anchor. The audio and video sources include classroom recordings and interview recordings. Their text processing uses automatic speech recognition or manual transcription, and the text processing results are indexed by time and location to generate evidence units that can be traced back to the corresponding audio and video segment range through the location identifier. S6. Statistical summarization, topic clustering, and generation of structured intermediate results. The data analyst agent performs statistical summarization, structured induction, and comparative organization of the evidence unit set and candidate item metadata to generate a structured intermediate result set. The structured intermediate result set is any combination of topic clustering results, research paradigm distribution results, summary results of research questions or research object elements, summary results of method elements, summary results of variable and indicator elements, summary results of data source and sample elements, summary results of context elements, and summary results of research conclusions and points of contention. For each structured summary conclusion, it is associated with its supporting evidence unit citation set to form a support relationship of "conclusion - evidence unit - source location". S7, Consistency Check, Reference Backtracking Check, and Conflict Resolution The logic checker agent performs consistency checks, citation traceability checks, and conflict detection on the structured intermediate result set and draft. The conflict resolution module then classifies the detected conflicts, aggregates evidence, and arbitrates the results, forming a conflict record. The conflicts can be any combination of factual conflicts, citation conflicts, structural conflicts, and logical conflicts. When arbitrating, the conflict resolution module considers source consistency, evidence coverage, timeliness consistency, and citation traceability availability. When a unique arbitration conclusion cannot be determined, an uncertainty warning is output. The conflict record includes: conflict type, conflict location, related evidence unit citation set, arbitration strategy identifier, arbitration result description, and uncertainty warning fields, and is written to the shared workspace module. S8, Structured Writing, Chapter Filling, and Citation Chain Generation The literature reviewer agent generates structured text results for educational review papers based on a structured intermediate result set, conflict records, and output structure constraints. The structured text results include: paper title, abstract, introduction, hierarchical structure of the review topic, comparison of research paradigms and points of contention, research gaps and future directions, methodological suggestions, references, and citation markers. Key arguments in the structured text results are linked through citation anchors and source and location identifiers in the evidence unit set to form a traceable citation chain. The outputs evidence summary field and conflict summary field corresponding to the chapters, ensuring that the "structure-content-evidence-source" of the review text remains consistent and aligned. S9, Output and Feedback Closed-Loop Update The output structured text results receive user feedback, forming a closed-loop update. The user feedback can be any combination of modification instructions, inclusion / exclusion of items, structural adjustment preferences, supplementary citation requirements, opinions on the selection or rejection of controversial points, and opinions on the adoption of unstructured evidence. The closed-loop update includes updating the task distribution strategy, conflict resolution parameters, evidence extraction rules, topic clustering strategy, output structure constraint template, and knowledge base management strategy, so that subsequent generation can achieve adaptive optimization while maintaining consistency with the main methodological line.

2. The method for generating structured review papers in education according to claim 1, characterized in that, In step S3, the task distribution module distributes task tickets to the corresponding intelligent agents for execution according to the mapping rules from task type to role, and records the task status and task version in the shared workspace module. When task dependencies are not met, the task distribution module controls the triggering of subsequent tasks. When a task fails or the result does not meet the gating conditions, a retry, downgrade, or alternative path is executed according to the fallback strategy.

3. The method for generating structured review papers in education according to claim 1, characterized in that, When the request parameter set contains unstructured data entries, the candidate item set in step S4 includes classroom videos, interview recordings, observation records, or their transcribed texts as candidate evidence sources, and source identifiers and location index rules are written for the candidate evidence sources.

4. The method for generating structured review papers in education according to claim 1, characterized in that, The conflict types in step S7 are factual conflict, citation conflict, structural conflict, and logical conflict. For factual conflict, parallel viewpoints are output or disputed descriptions are retained; for citation conflict, a missing citation or failure to backtrack is indicated; for structural conflict, inconsistent chapter levels or inconsistent paragraph classification is indicated; for logical conflict, a broken chain of inferences or a jump in the argument is indicated.

5. A structured educational review paper generation device constructed using the structured generation method for educational review papers according to claim 1, characterized in that, The device consists of a configuration parsing module, a request parsing module, a task distribution module, an evidence processing module, a statistical summary module, a conflict resolution module, a structured writing module, and an output and feedback module. These modules communicate via a shared workspace module for data alignment and version management, enabling multi-agent collaboration at each stage and achieving structured generation of educational review papers. The configuration parsing module parses the agent configuration file and constructs a multi-agent collaborative workflow; the request parsing module obtains the user summary generation request and forms a request parameter set; the task distribution module generates a task plan and distributes the formed task ticket set to the corresponding agents; the evidence processing module performs textualization and evidence unitization processing on candidate items and unstructured data and establishes reference anchor associations; the statistical summary module performs topic clustering, distribution statistics, and comparative summary on the evidence unit set and outputs structured intermediate results; the conflict resolution module triggers consistency verification, reference backtracking verification, and conflict detection, and outputs conflict records and uncertainty prompts. The structured writing module generates structured text results for review papers under output structure constraints and establishes the association between text citations and evidence units; the output and feedback module outputs results and receives user feedback to form a closed-loop update.

6. A structured generation system for educational review papers constructed using the structured generation method for educational review papers as described in claim 1, characterized in that, The system includes a server, a configuration storage component, a data access component, a model service component, a shared workspace component, and a client. The server communicates with each component to execute a method for generating structured review papers in education and outputs structured results consistent with the agent's configuration file. The configuration storage component stores the agent's configuration file and the output structure constraint template. The data access component accesses unstructured data such as paper entry metadata, full-text text, classroom recordings, interview recordings, observation records, or their transcribed texts, and writes them with source and time stamp identifiers. The model service component provides semantic representation computation, topic clustering and summary generation, evidence type labeling, logical consistency verification, and conflict detection reasoning capabilities. The shared workspace component stores request parameters, task tickets, evidence units, structured intermediate results, writing drafts, and conflict records, and supports versioning. The client receives user requests, displays structured text results and citation information, and submits user feedback.

7. A computer-readable storage medium designed according to the structured generation method for educational review papers as described in claim 1, characterized in that, The storage medium stores a computer program that, when run on a processor, causes the processor to call various modules in the educational review paper structure generation device or system, execute the steps of the educational review paper structure generation method based on multi-agent collaboration, and complete the following during execution: multi-agent collaborative workflow construction, request parsing and field mapping, task decomposition and task distribution, candidate item aggregation, evidence unitization processing, statistical summary and structured induction, logical verification and conflict resolution, structured writing, citation association and feedback closed-loop update.

Images