A chapter-driven research report data acquisition and evidence pool construction method and system thereof

Abstract

The present application relates to the technical field of automatic report generation and data intelligence processing, and discloses a chapter-driven research report data acquisition and evidence pool construction method and system. It aims to solve the problems of insufficient chapter evidence support and low construction efficiency caused by the ambiguity of natural language requirement semantics and the blindness of multi-source retrieval. The technical scheme includes: receiving the original requirement text for standardized analysis and parameter completion, establishing a data requirement model to output a structured task object; decomposing the task into chapter-level retrieval targets and generating a retrieval plan; parallelly expanding the text and structured field collection tasks and performing hierarchical storage; calculating the sufficiency score based on the collection state and comparing it with the threshold to generate a dynamic supplementary collection decision; gathering the collection results to calculate the correlation score and construct candidate clusters, and outputting unified evidence entities after deduplication, normalization and conflict coordination; dividing the evidence pool based on the purpose of labeling and calculating the chapter adaptation degree to output the chapter adaptation mapping result.

Description

Technical Field

[0001] This invention relates to the field of automated report generation and intelligent data processing technology, specifically to a chapter-driven method and system for acquiring research report data and constructing evidence pools. Background Technology

[0002] With the development of large language models and automated content generation technologies, automatically generating research reports based on user needs has become an important application area. However, existing research report data acquisition and evidence construction processes typically suffer from the following technical shortcomings:

[0003] First, the requirements analysis is crude and lacks structured task modeling. Existing methods often directly convert user-input natural language requests into search keywords without standardizing colloquial expressions, omissions, vague references, cross-process supplementation, etc., or filling in missing parameters. This results in unclear task boundaries, ambiguous data requirements, and difficulty in accurately defining the key support points of different chapters for the main text or chart statistics.

[0004] Secondly, the retrieval strategy is simplistic and lacks chapter-level adaptation and dynamic broadening mechanisms. Existing systems typically use a single global retrieval approach for data collection, failing to break down the global task into local retrieval targets for different chapters. Furthermore, they lack pre-defined tiered broadening paths and primary / backup data source configurations for scenarios with insufficient results or missing fields, making it difficult to balance initial retrieval accuracy with subsequent scalability.

[0005] Third, the data acquisition channels are not differentiated, and the storage structure is flat. The existing data acquisition process does not distinguish between the data characteristics of text generation and chart generation, and adopts a single channel for mixed acquisition. This results in a mixture of high-information-density text records and aggregable structured fields. It lacks hierarchical storage based on usage and semantic tagging assistance, making it difficult to efficiently support the dual-modal generation of text writing and visualization statistics in the future.

[0006] Fourth, there is a lack of quantitative assessment and dynamic closed-loop mechanism for data sufficiency. Traditional data collection often adopts fixed procedures or is executed in one go, without establishing assessment models for the sufficiency of the main text and the sufficiency of charts and graphs oriented towards chapter objectives. When the collection results are insufficient, it is impossible to automatically trigger tiered supplementary collection, data source switching, or target downgrading based on preset rules, which can easily lead to reports with hollow content or weak evidence support.

[0007] Fifth, there is insufficient capacity for managing multi-source heterogeneous data, and a lack of unified evidentiary entities. The collected results often suffer from problems such as duplication from multiple sources, inconsistent field standards, and information conflicts. Existing methods lack systematic mechanisms for candidate association construction, deduplication and merging, field standardization, cross-source completion, and conflict coordination, making it difficult to transform multi-source heterogeneous data into high-quality unified evidentiary entities that can be directly invoked.

[0008] Sixth, the way evidence is organized is disconnected from its downstream uses. Traditional evidence databases are mostly organized by data form or source, without pooling them according to their subsequent uses such as text, statistics, relationships, and themes. They also lack a pre-adaptive mapping between evidence entities and chapter objectives, resulting in low efficiency in evidence allocation and difficulty in supporting accurate chapter-level citations and collaborative generation of text and graphics.

[0009] In summary, how to provide a research report data acquisition and evidence construction method that can accurately analyze user intent, formulate differentiated retrieval and collection strategies by chapter, dynamically evaluate data sufficiency and achieve closed-loop supplementary collection, and finally construct a multi-purpose evidence pool has become an urgent technical problem to be solved in this field. Summary of the Invention

[0010] The purpose of this invention is to solve the technical problems of insufficient chapter-level evidence support, difficulty in ensuring data sufficiency, and low efficiency in evidence pool construction caused by semantic ambiguity in natural language reporting requirements and blind retrieval of multi-source heterogeneous data.

[0011] To achieve the above objectives, the present invention adopts the following technical solution:

[0012] This invention provides a chapter-driven method for acquiring research report data and constructing an evidence pool, comprising the following steps:

[0013] Step S1: Receive the original requirement text input by the user, perform normalization processing to obtain standardized semantic units, perform requirement semantic parsing to extract core task elements, complete and coordinate missing parameters and constraints, establish a data requirement model to calculate the requirement intensity of the main text data and the requirement intensity of the structured field data, and output the structured task object.

[0014] Step S2: Read the structured task object, decompose the global task into multiple chapter-level search targets, generate an initial tight search strategy for each target and construct hierarchical relaxation rules, perform data source matching and master / slave configuration, and form a set of chapter-level search plan objects.

[0015] Step S3: Read the set of chapter-level retrieval plan objects, expand them into executable collection tasks, execute the text data collection channel and the structured field collection channel in parallel, generate semantic tags and store them in layers, and output the collection status objects and initial collection results corresponding to each chapter.

[0016] Step S4: Based on the collection status object, the hierarchically stored initial collection results and retrieval plan object, construct a chapter-level sufficiency judgment object, calculate the sufficiency score of the main text generation and the sufficiency score of the chart generation respectively, compare them with the preset threshold, and generate dynamic supplementary collection, rollback and downgrade trigger decisions according to the comparison results and the hierarchical relaxation rules, and execute the corresponding supplementary collection task according to the trigger decision to obtain the supplementary collection result.

[0017] Step S5: Aggregate the initial collection results and supplementary collection results, calculate the candidate association scores between records to construct candidate association clusters, remove duplicates from the candidate association clusters and form unified evidence entities, perform field standardization, cross-source completion and conflict coordination, conduct evidence quality assessment and usage labeling, and output a unified evidence entity set, quality score and usage-type usability labeling.

[0018] Step S6: Based on the usage-type availability label, the unified evidence entities are divided into different types of usage-type evidence pools. The degree of adaptation of each evidence entity to the chapter target is calculated to form a chapter adaptation mapping. The evidence pool set and the pre-adaptation mapping result are output.

[0019] In the above scheme, step S1 includes:

[0020] Step 1.1, Receiving and Normalizing Original Requirements: Perform normalization mapping on the original requirement text sequence to obtain a normalized requirement set. ,in , This indicates the input related to the task in the current or previous round. Represents the normalized mapping function. Represents standardized semantic units, ;

[0021] Step 1.2, Requirement Semantic Parsing and Task Element Extraction: Extract core task elements and organize them into task semantics:

[0022]

[0023] in, Indicates thematic elements, Representing domain elements, Indicates the report type. Indicates the scope of analysis. Indicates the target audience, This indicates the key points. Indicates the output format. To indicate the need for charts, Indicates a citation requirement, Indicates style preference;

[0024] Step 1.3, Missing Parameter Completion and Constraint Coordination: ... Each task parameter is represented as ,in Indicates the parameter value. Indicates the source identifier. This indicates the priority weight, and the conflicting parameters are coordinated based on the source identifier and priority weight;

[0025] Step 1.4, Data Requirements Modeling and Collection Direction Generation: Calculate the data requirement intensity for the main text. and the intensity of demand for structured field data The formula is:

[0026]

[0027]

[0028] in This indicates the factors influencing the demand for the main text of the report. Indicates the citation strength requirement. Indicates the required depth of content. This indicates the need for the main body of the text to elaborate on the key analytical content; Indicates the strength of the chart preference. This indicates a need for trend analysis. Indicates the main analysis needs, Indicates the statistical requirements for distribution; For preset weights or learnable weights; and based on and The size relationship triggers the corresponding data acquisition strategy.

[0029] In the above scheme, step S2 includes:

[0030] Step 2.1, Expanding Search Targets and Mapping Chapters: Forming a set of chapter-level search targets. ,in , Indicates chapter direction or semantic tags. Indicates the data usage type. This indicates the set of fields that are preferentially depended upon. Indicates weight;

[0031] Step 2.2, Initial Search Strategy Generation: Constructing the initial search strategy ,in This represents the initial search expression. Represents a set of range constraints. Represents the target data source collection. This indicates the set of fields to be returned; and the core query is selected based on the candidate search expression adaptation score, with the scoring formula being:

[0032]

[0033] in Indicates semantic relevance. Indicates field coverage capability. Indicates the degree to which range constraints are satisfied. For the corresponding weights;

[0034] Step 2.3, Construction of Tiered Relaxation Rules: Construct a set of tiered relaxation rules. , of which Level relaxation rules , This indicates a relaxation operation on the expression. This indicates a range constraint relaxation operation. This indicates that the field requirements have been relaxed. Indicates a data source extension operation;

[0035] Step 2.4, Data Source Matching and Plan Formulation: Calculate the candidate data source matching score:

[0036]

[0037] in For domain matching degree, As to the degree of field support, To ensure the credibility of the source, To ensure timeliness, For the corresponding weights;

[0038] Output chapter-level search plan object:

[0039]

[0040] in, Indicates chapter-level search target. Indicates the initial search strategy. This represents the set of tiered relaxation rules. Represents the main data source collection. Represents a set of secondary data sources. This indicates that a collection of fields will be returned. This indicates the basic threshold configuration or expected data collection budget required for subsequent sufficiency assessments.

[0041] Step 2.5: Final output of the collection of chapter-level search plan objects:

[0042]

[0043] Each Each corresponds to a chapter-level search target and includes its initial search strategy, relaxed rules, data source configuration, and basic threshold configuration.

[0044] In the above scheme, step S3 includes:

[0045] Step 3.1, Task Deployment: Expand the chapter-level search plan object into a set of data collection tasks. Subtasks , Indicates the target data source. This represents a search expression. Indicates the required return fields. Indicates the channel type. Indicates the expected collection volume;

[0046] Step 3.2, Main Text Channel Execution: Extract Main Text Data:

[0047]

[0048] Among them, a single record , For recording identification, As the title, For the abstract, For key segments, This is the conclusion segment. For the source, For time;

[0049] Step 3.3, Structured Channel Execution: Extracting Structured Data:

[0050]

[0051] Among single records Each field represents the year, author, institution, country, keywords, relational fields, and source, respectively.

[0052] Step 3.4, Tag Generation and Status Output: Generate semantic tags:

[0053]

[0054]

[0055] Indicates the corresponding record identifier. Indicates topic tags, Indicates method label, Indicates the application tag, Labels indicating research or industry stage Indicates the confidence level of the label;

[0056] Step 3.5, and output the acquisition status. ;

[0057] in, Indicates the number of text records. Indicates the number of structured records. Indicates the completeness rate of key fields. This indicates the source coverage or the hit rate of primary and secondary sources.

[0058] In the above scheme, step S4 includes:

[0059] Step 4.1: Read the acquisition status object Main text evidence layer Structured metadata layer Semantic tag layer and the corresponding chapter-level search plan objects And construct a chapter-level sufficiency determination object for each chapter:

[0060] ;

[0061] in, This represents a subset of text-based candidate data associated with this chapter. This represents a subset of structured field data associated with this chapter. This represents a subset of semantic tags associated with this chapter. This indicates the preset threshold and budget parameters in the chapter-level search plan object;

[0062] Step 4.2, Text Sufficiency Assessment: Calculate the score:

[0063]

[0064] , ,

[0065] in, Indicates the satisfaction level of the number of text records. This indicates the percentage of high-quality main text. Indicates subtopic coverage. Indicates time coverage. Indicates writability, For the corresponding weights, This represents the minimum threshold required for the main text records. For high-quality record count, To prevent the division into zero minimum constants;

[0066] Will Compared with the preset text generation threshold Compare. When satisfied...

[0067]

[0068] If the current chapter is deemed ready for text generation, it is determined that the current chapter is ready for text generation; otherwise, the current chapter is deemed to have insufficient text evidence.

[0069] Step 4.3, Assessment of Chart Sufficiency: Calculate the score:

[0070] ,

[0071] ,

[0072] in, Indicates the completeness rate of key fields. Indicates the satisfaction of sample size. Indicates the distribution efficiency. Indicates relation density. Indicates chart generability. For the corresponding weights, For the core field set, For field completeness rate, For field weights, The minimum sample threshold for the chart;

[0073] Will With chart generation threshold Compare. When satisfied...

[0074]

[0075] If the current chapter is ready, determine that it is ready for chart generation; otherwise, determine that the current chapter has insufficient structured fields or sample distribution.

[0076] Step 4.4, Dynamic Trigger Decision: Generate trigger decision:

[0077] ,

[0078] Positioned based on the comparison result between the score and the threshold. and And perform the corresponding action;

[0079] in, This indicates the action of adding text to the main body. This indicates a structured field supplementary data collection action. This indicates a data source rollback or expansion action. This indicates a downgrade action for chapter or chart objectives.

[0080] In the above scheme, step S5 includes:

[0081] Step 5.1, Fusion and Association Construction: Calculate any two records Association score:

[0082]

[0083] Those exceeding the threshold are included in the candidate association set:

[0084] ,

[0085] in This indicates a hard identifier match. Indicates title similarity. Indicates the similarity of authors or subject combinations. Indicates proximity in time. Indicates the semantic similarity of topics. For the corresponding weights, Indicates the first One candidate association cluster;

[0086] Step 5.2, Deduplication and Merging: Execute the merge function to construct a unified evidence entity:

[0087] ;

[0088] Represents the merge function. This indicates a unified internal evidence identification system. This represents the merged collection of text content. This represents the merged set of structured fields. This represents the merged set of semantic labels. Represents the set of source trajectories;

[0089] Step 5.3, Standardization and Conflict Reconciliation: Perform field mapping:

[0090]

[0091] This represents the original set of fields before merging. This represents the standardized set of fields, merging field values. Score based on conflict resolution:

[0092]

[0093] in, Indicates the credibility of the source. Indicates the completeness of the field content. This indicates consistency with other fields of the current unified entity. Indicates timeliness, For the corresponding weights;

[0094] The final values ​​are determined, with each item representing source credibility, content completeness, entity consistency, and timeliness, respectively. For the corresponding weights;

[0095] Step 5.4, Quality Assessment and Labeling: Calculate the quality score:

[0096]

[0097] in, This indicates the relevance to the current task theme and chapter objectives. , Indicates the completeness of structured fields. Indicates the availability of references. This indicates consistency within a unified entity. For the corresponding weights;

[0098] And generate usage labels:

[0099]

[0100] in, Indicate whether it is suitable to be included in the main text evidence pool. This indicates whether the item is suitable for inclusion in the statistical evidence pool. This indicates whether the individual is suitable for inclusion in the relationship evidence pool. Indicate whether it is suitable to enter the subject evidence pool. Indicates whether it is suitable as a candidate for explicit citation.

[0101] In the above scheme, step S6 includes:

[0102] Step 6.1, based on unified evidence entity Usage label Unified evidence entities will be included in the main text evidence pool. Statistical evidence pool Relationship Evidence Pool Thematic evidence pool The set that constitutes:

[0103] ,

[0104] Step 6.2, Adaptation Mapping: Calculate Evidence Entities For the The degree of fit of the chapter objectives This generates adaptation information:

[0105] ,

[0106] Step 6.3, Final Output .

[0107] This invention also provides a chapter-driven research report data acquisition and evidence pool construction system, including a processor and a storage medium. When the processor executes the program in the storage medium, it implements the aforementioned chapter-driven research report data acquisition and evidence pool construction method.

[0108] Because the present invention employs the above-mentioned technical means, it has the following beneficial effects:

[0109] 1. This invention solves the technical problems of traditional single global search being unable to accurately match the differentiated data needs of each chapter of the report and lacking a contingency plan for insufficient search results by standardizing and parsing user natural language needs, completing missing parameters, and modeling data needs (for steps 1 to 4), and decomposing the global task into chapter-level search targets, generating an initial tight search strategy and preset hierarchical relaxation rules (corresponding to step 2). It achieves the effect of transforming fuzzy task intent into a calculable and executable chapter-level search plan, while taking into account both initial search accuracy and subsequent expansion capabilities.

[0110] 2. This invention solves the technical problem of mutual interference between subsequent text generation and chart generation and low field availability caused by mixed collection of multi-source heterogeneous data by dividing the collection links into dual channels based on the demand intensity of the main text and structured fields, and simultaneously executing semantic tag generation and hierarchical storage of the main text evidence layer, structured metadata layer and semantic tag layer (step 3). It achieves the effect of targeted diversion of collection results according to the downstream generation purpose and separately ensuring the integrity of text narrative evidence and information density and statistical aggregation fields.

[0111] 3. This invention solves the technical problem that fixed collection strategies easily lead to redundant evidence in some chapters and a lack of evidence in key chapters, and the collection link cannot be closed-loop and adaptive. It achieves the effect of driving collection on demand iteration with generation availability as the judgment benchmark, avoiding invalid retrieval and waste of computing resources.

[0112] Synergistic effect

[0113] The steps in this invention are not performed independently of each other in terms of requirement analysis, data retrieval, data collection, and evidence organization. Instead, they form a sequential and interconnected processing relationship during execution. The topics, scope, output formats, chart requirements, and citation requirements obtained after analyzing user requirements are directly used to determine the data types and search priorities required for different chapters. The search results for each chapter then serve as the basis for subsequent text data collection, structured field collection, and sufficiency assessment. Therefore, the results of the preceding analysis can continuously constrain the subsequent collection process, avoiding the problem of blindly searching based on a single keyword in existing technologies.

[0114] Meanwhile, after data collection, this invention does not directly use the results for report generation. Instead, it assesses the data based on the number of text records obtained for each chapter, field completeness, source coverage, and chart generation conditions. When a chapter lacks sufficient textual evidence or structured fields, the system can supplement data collection or switch data sources according to the relaxed rules in the original retrieval plan for that chapter. When the data already meets the generation requirements, the retrieval scope is no longer expanded. Thus, the relaxed retrieval rules and data sufficiency assessment work together to ensure that supplementary data collection is triggered only for chapters with weak evidence, solving the problems of insufficient evidence and redundant data in some chapters in traditional one-time data collection methods.

[0115] Furthermore, this invention performs deduplication, field standardization, and conflict resolution on both the initial data collection results and the supplementary data collection results. Then, it categorizes the evidence into different evidence pools based on evidence quality and subsequent uses, and establishes an adaptation relationship between the evidence and the chapters. In this way, the data obtained from the supplementary data collection does not form scattered records, but can be organized into unified evidence that can directly support text writing, statistical charts, relationship analysis, and thematic summarization. This processing method creates a continuous coordination between the data collection results, supplementary data collection results, and evidence pool construction results, solving the problems of multi-source data duplication, field conflicts, unclear uses, and difficulty in calling data by chapter.

[0116] Therefore, this invention constrains chapter retrieval through demand analysis results, triggers supplementary data collection actions through acquisition results, and supports evidence pool division and chapter allocation through governance results. This enables the report data acquisition process to dynamically adjust according to the evidence status of each chapter, ultimately forming an evidence set that can be directly used for subsequent report generation. The combined use of these technical means improves the relevance and completeness of research report data acquisition, reduces invalid searches and duplicate data processing, and enhances the reliability of chapter text generation, chart generation, and citation support. Attached Figure Description

[0117] Figure 1 This is a simplified flowchart of the present invention. Detailed Implementation

[0118] The embodiments of the present invention will be described in detail below. Although the present invention will be described and illustrated in conjunction with some specific embodiments, it should be noted that the present invention is not limited to these embodiments. On the contrary, any modifications or equivalent substitutions made to the present invention should be covered within the scope of the claims of the present invention.

[0119] Furthermore, to better illustrate the present invention, numerous specific details are set forth in the following detailed embodiments. Those skilled in the art will understand that the present invention can be practiced without these specific details.

[0120] This invention provides a chapter-driven method and system for data acquisition and evidence pool construction in research reports, including:

[0121] Module 1: User Requirement Reception, Semantic Parsing, and Data Requirement Modeling

[0122] This module is used to receive report generation requests from users and convert the raw requests in natural language into structured task objects and data requirement descriptions that can be directly invoked in subsequent data retrieval, collection, and storage processes.

[0123] Unlike directly converting user input into search keywords, this module treats user input as a report generation task, standardizing, structuring, completing missing parameters, and modeling data requirements to clarify the data types, field ranges, and key chapters to be collected subsequently.

[0124] 1.1 Receiving and Standardizing Original Requirements

[0125] The system receives the user's original requirement text. The requirement text can be a single-turn input or a task description accumulated from multiple turns of dialogue. Since the original input often contains colloquial expressions, omissions, vague references, synonym substitutions, and cross-turn additions, the system first performs standardization processing on it.

[0126] The normalization process includes: terminology cleaning, redundant expression removal, time representation standardization, synonym merging, pronoun recovery, cross-round condition merging, and invalid constraint filtering. After processing, a semantically clear and well-defined normalized task description is obtained, and several standardized semantic units are output for subsequent field extraction.

[0127] Let the original requirement text sequence be:

[0128]

[0129] in, This represents a segment of input relevant to the current task in the current or historical rounds. The system performs normalization mapping on the sequence of requirement texts to obtain a normalized requirement set:

[0130]

[0131] in, Represents the normalized mapping function. This represents the standardized semantic unit after processing.

[0132] 1.2 Requirement Semantic Parsing and Task Element Extraction

[0133] After standardization, the system performs semantic parsing of the requirements to extract the core task elements needed for report generation, including: report topic, target domain, report type, analysis scope, target audience, key content, output format, chart requirements, citation requirements, and style requirements.

[0134] The above information is organized into structured task parameters for subsequent retrieval strategy generation and data acquisition configuration. The task semantics can be represented as:

[0135]

[0136] in, Indicates thematic elements, Representing domain elements, Indicates the report type. Indicates the scope of analysis. Indicates the target audience, This indicates the key points. Indicates the output format. To indicate the need for charts, Indicates a citation requirement, Indicates style preference.

[0137] 1.3 Missing Parameter Completion and Constraint Coordination

[0138] Since user input is often incomplete, the system completes missing parameters that affect subsequent execution after extracting explicit task parameters. The completion is based on report type, domain priors, historical context, and preset templates, and the completion results are saved separately from the user's explicit input.

[0139] After completion, the system coordinates constraints that may conflict, overlap, or have unclear scope. For example, when the requirements contain both constraints such as "as comprehensive as possible" and "controlling the length," the system restates them uniformly to form an executable set of task constraints.

[0140] Let the first The task parameters are: Then it can be expressed as:

[0141]

[0142] in, Indicates the parameter value. Indicates the source identifier. This indicates the priority weight. The source identifier includes at least: explicit user input, context inheritance, template default value, and system inferred value. The system retains and coordinates conflicting parameters based on the priority weight.

[0143] 1.4 Data Requirements Modeling and Acquisition Direction Generation

[0144] After establishing stable task parameters, the system further establishes a data requirement model for data retrieval and collection to answer the following questions: What data is needed to generate this report, whether this data is used for text generation or chart generation, and what are the key dependencies of different chapters on the data?

[0145] This module categorizes data requirements into two types:

[0146] The main body data requirements are used to support the generation of the report body, with a focus on the title, abstract, key paragraphs, concluding sentences, source credibility, and citation availability;

[0147] The requirement for structured field data is to support chart generation, with a focus on year, institution, country, author, keywords, relational fields, field completeness, and statistical aggregability.

[0148] The system calculates the demand intensity of the main text data separately. and the intensity of demand for structured field data :

[0149]

[0150]

[0151] in, This indicates the factors influencing the demand for the main text of the report. Indicates the citation strength requirement. Indicates the required depth of content. This indicates the need for the main body of the text to elaborate on the key analytical content; Indicates the strength of the chart preference. This indicates a need for trend analysis. Indicates the main analysis needs, Indicates the statistical requirements for distribution; , These are preset weights or learnable weights.

[0152] when At that time, the system prioritizes the retrieval and collection of high-quality textual evidence; when When the values ​​are high, the system prioritizes ensuring the coverage and integrity of structured fields; when both are high, the system initiates a data acquisition strategy that combines the main text and structured fields.

[0153] 1.5 Structured Output

[0154] This module ultimately outputs a structured task object. The task object includes at least: task identifier, standardized topic expression, extended topic expression, domain label, report type, time range, spatial range, target reader attributes, key analysis dimensions, output format, style requirements, citation requirements, chart preferences, parameter source identifiers, parameter priority, as well as the intensity of text data requirements, the intensity of structured field data requirements, candidate chapter directions, and high-value field hints.

[0155] This structured task object serves as a unified input for subsequent modules, used to generate retrieval strategies, configure acquisition channels, set sufficiency judgment conditions, and support subsequent dynamic supplementary acquisition and evidence pool construction.

[0156] 1.6 Technical Effects

[0157] This module transforms natural language reporting requirements into computable task objects and executable data requirement models, achieving a stable connection from user input to data acquisition. This approach reduces the impact of ambiguity in the original expression on subsequent retrieval and acquisition, improves the completeness, consistency, and executability of task parsing, and provides unified input for subsequent chapter-level retrieval, dual-channel acquisition, and evidence organization.

[0158] Module 2: Initial Search Strategy Generation, Hierarchical Relaxation Rule Construction, and Chapter-Level Search Plan Formation

[0159] After receiving user requirements, semantic parsing, and data requirement modeling, the system enters the initial retrieval strategy generation, hierarchical relaxation rule construction, and chapter-level retrieval plan formation module. This module is used to convert the structured task objects and data requirement objects output by the previous module into retrieval plans that can be directly executed by the subsequent data acquisition module.

[0160] This module does not employ a single global search approach, but instead constructs a chapter-level search plan based on chapter objectives. The system first generates a high-precision initial search strategy, then pre-sets progressively wider paths, and combines data source capabilities to complete primary and backup configurations, thereby balancing search quality with subsequent supplementary data acquisition and expansion capabilities.

[0161] 2.1 Search Target Expansion and Chapter Mapping

[0162] The system reads the structured task object, the intensity of text data requirements, the intensity of structured field data requirements, and candidate chapter direction suggestions, and based on this, breaks down the global task into multiple chapter-level search targets, forming a set of chapter-level search targets:

[0163]

[0164] in, Indicates the first Chapter-level search targets. Each It should include at least chapter direction, data usage type, field preferences, and priority information, and can be represented as:

[0165]

[0166] in, Indicates chapter direction or chapter semantic tags. Indicates the data usage type. This indicates the set of fields that this chapter depends on first. This indicates the weight of the search target at the chapter level.

[0167] The system can call a large language model to break down the global task into chapters, outputting a chapter-level candidate set of search targets; then, combined with the rules module, it completes filtering, deduplication, and weight normalization to form the final chapter-level search target set. .

[0168] Through the above processing, the system transforms the data requirements of the entire report into local search targets for different chapters, providing input for subsequent differentiated retrieval and dual-channel data collection.

[0169] 2.2 Initial Tight Search Strategy Generation

[0170] For each chapter-level search target The system constructs the corresponding initial retrieval strategy. This strategy prioritizes ensuring that search results are highly relevant to the current chapter's objectives and meet the corresponding data usage and field requirements.

[0171] The initial search strategy can be expressed as:

[0172]

[0173] in, This represents the initial search expression. Represents a set of range constraints. Represents the target data source collection. This indicates that the returned fields require a set.

[0174] The range constraint set This can include time range, geographical range, object range, etc.; field collection Configure according to the purpose of the data: for text purposes, prioritize titles, abstracts, key paragraphs, concluding sentences, and source fields; for charts and graphs, prioritize year, institution, country, keywords, and relationship fields.

[0175] The system can invoke a large language model to generate several candidate search expressions around the current chapter's target and label them with applicable scenarios. Subsequently, the system scores the candidate expressions based on semantic relevance, field coverage, and range matching, and selects one or more expressions with the highest scores as the core query part of the initial search strategy.

[0176] The fit score of the candidate search expression can be expressed as:

[0177]

[0178] in, This indicates the semantic relevance between the search expression and the chapter-level search target. This indicates the coverage capability of the target field requirement. Indicates the degree to which the range constraint is satisfied. For the corresponding weights.

[0179] This step results in a set of high-precision initial retrieval strategies for different chapter objectives, rather than a unified global query.

[0180] 2.3 Construction of Tiered Relaxation Rules

[0181] To address situations where some chapters result in insufficient results, incomplete fields, or insufficient sample distribution after the initial search, the system pre-sets tiered relaxation rules for each chapter-level search target while generating the initial search strategy.

[0182] The relaxation rules include at least the following categories: relaxation of retrieval expressions, relaxation of range constraints, relaxation of field completeness, expansion of data source range, and downgrading of target expressions when necessary.

[0183] Regarding the first For each chapter-level search target, the system constructs a corresponding set of hierarchical relaxation rules:

[0184]

[0185] in, Indicates the first The level of relaxation rule can be expressed as:

[0186]

[0187] in, This indicates a relaxation operation for the search expression. This indicates a range constraint relaxation operation. This indicates that the field requirements have been relaxed. This indicates a data source extension operation.

[0188] The system can call the large language model to generate relaxation suggestions based on chapter objectives, initial retrieval strategies, scope constraints, and preset relaxation level frameworks, and then combine them with the rule base to solidify them into an executable set of relaxation rules.

[0189] This set of rules only defines "how to relax" the rules, and the specific triggering time is determined by the data sufficiency judgment results of subsequent modules.

[0190] 2.4 Data Source Matching and Chapter-Level Search Plan Object Formation

[0191] After generating the initial retrieval strategy and tiered relaxation rules, the system further completes data source matching and primary / backup configuration.

[0192] The system maintains data source registration information, which includes at least the source type, domain coverage, field support capabilities, text completeness, structure level, update timeliness, credibility level, and access method. The system searches based on chapter-level targets. Initial search strategy and field requirement set The candidate data sources are matched and sorted.

[0193] The matching score of the candidate data source can be expressed as:

[0194]

[0195] in, This indicates the degree of domain relevance between the data source and the chapter's target. This indicates the degree of support for the target field set. Indicates the credibility of the source. Indicates the timeliness of the update. For the corresponding weights.

[0196] The system can call a large language model to perform semantic analysis on the adaptation relationship of candidate data sources, and combine the above scoring results to determine the primary data source, secondary data source and backup data source.

[0197] Based on this, the system outputs a corresponding chapter-level search plan object for each chapter-level search target. :

[0198]

[0199] in, Indicates chapter-level search target. Indicates the initial search strategy. This represents the set of tiered relaxation rules. Represents the main data source collection. Represents a set of secondary data sources. This indicates that a collection of fields will be returned. This indicates the basic threshold configuration or expected data collection budget required for subsequent sufficiency assessments.

[0200] The Parameters may include the budget for the number of records in the main text, the budget for the number of structured records, the threshold for the completeness of key fields, and the minimum sample requirement for visualization.

[0201] 2.5 Module Output

[0202] This module ultimately outputs a collection of chapter-level search plan objects:

[0203]

[0204] Each of them Each item corresponds to a chapter-level search target and includes its initial search strategy, relaxation rules, data source configuration, and basic threshold configuration. Subsequent modules can directly perform multi-source data collection, sufficiency assessment, and dynamic supplementary data collection based on the chapter-level search plan object.

[0205] Through this module, the system transforms global tasks into chapter-oriented retrieval execution units, ensuring initial retrieval accuracy while providing a unified entry point for subsequent broadening, supplementary retrieval, and rollback.

[0206] Module 3: Dual-channel multi-source heterogeneous data acquisition and hierarchical storage

[0207] After generating the initial retrieval strategy, constructing tiered relaxation rules, and forming a chapter-level retrieval plan, the system enters the dual-channel multi-source heterogeneous data acquisition and hierarchical storage module. This module is used to perform targeted acquisition from multiple heterogeneous data sources based on the chapter-level retrieval plan object, and to distribute and store the acquisition results according to their subsequent uses to support text generation, chart generation, topic classification, and subsequent sufficiency assessment.

[0208] This module divides the data acquisition process into a text data acquisition channel and a structured field acquisition channel, and can add a semantic tag generation channel as needed. The text channel is mainly for subsequent text generation, the structured field channel is mainly for subsequent statistical analysis and chart generation, and the semantic tag channel is used to generate high-level semantic tags to provide auxiliary information for subsequent evidence pool construction and chapter-level evidence allocation.

[0209] 3.1 Chapter-level retrieval plan object reading and collection task deployment

[0210] The system reads the collection of chapter-level search plan objects output by the previous module:

[0211]

[0212] Each chapter-level retrieval plan object is further expanded into a set of executable data collection tasks:

[0213]

[0214] in, Indicates surrounding the first The first chapter target generated A single data collection subtask can be represented as:

[0215]

[0216] in, Indicates the target data source. This represents a search expression. Indicates the required return fields. Indicates the channel type. This indicates the expected data collection volume or data collection budget.

[0217] The system can invoke a large language model to break down data collection tasks based on chapter-level retrieval plan objects, data source capability descriptions, field requirements, and channel constraints. It then determines the execution method for each task: entering the main text channel, structured field channel, or semantic tag channel. Finally, by combining the rule engine and data source interface constraints, an executable set of data collection tasks is formed. .

[0218] 3.2 Execution of the text data acquisition channel

[0219] The main text data acquisition channel is geared towards the generation of subsequent chapters' main text, prioritizing the acquisition of text-type records with high information density, clear expression, and reliable sources.

[0220] This channel extracts the title, abstract, key excerpts, conclusion excerpts, source information, and necessary author and time information from the target data source. For data sources that support standard interfaces, the system directly returns the corresponding fields; for web-based data sources, the system extracts the required text content through page crawling and content parsing.

[0221] The results of the main text data collection are represented as follows:

[0222]

[0223] A single text-based data collection record can be represented as:

[0224]

[0225] in, Indicates record identifier, This indicates the title field. This represents the summary field. Indicates key text segments, This indicates a conclusion segment. Indicates source information, Indicates time information.

[0226] During the execution of the main text channel, the system can invoke a large language model to perform semantic filtering and key segment extraction on the returned text, identify records suitable as evidence for the main text of chapters, and extract summary segments, conclusion segments, or method overview segments suitable for subsequent generation from long texts. The filtered results are written to the main text-type temporary storage area.

[0227] 3.3 Execution of Structured Field Acquisition Channel

[0228] The structured field acquisition channel is geared towards subsequent statistical analysis and chart generation, and its goal is to obtain field-type data that can be aggregated, sorted, and statistically analyzed.

[0229] This channel prioritizes extracting year, author, institution, country, keywords, source, citation information, collaboration, category tags, and necessary unique identifiers. For data sources that return complete fields, the system directly receives the structured results; for data sources that return only partial fields, the system first collects the core fields and then supplements the missing fields later using supplementary sources.

[0230] The structured field collection results are represented as follows:

[0231]

[0232] A single structured data collection record can be represented as:

[0233]

[0234] in, Indicates record identifier, Represents the year field. This indicates the author field. Indicates the organization field. Represents the country field. This indicates a keyword field. Represents relational fields. Indicates the source field.

[0235] During the execution of the structured field channel, the system can call the large language model to perform field recognition, field mapping, and field attribution determination on the semi-structured results, extracting fields that can be mapped to year, organization, keywords, relationship edges, or entity categories, and providing record-level field completeness descriptions. After rule validation, the results are written to the structured field temporary storage area.

[0236] 3.4 Semantic Tag Generation and Hierarchical Storage

[0237] After collecting the main text data and structured field data, the system further generates a semantic tag layer. Semantic tags can come from existing tags in the data source, or they can be generated by the system based on the title, abstract, keywords, or text fragments.

[0238] The semantic tag set is represented as:

[0239]

[0240] A single tag record can be represented as:

[0241]

[0242] in, Indicates the corresponding record identifier. Indicates topic tags, Indicates method label, Indicates the application tag, Labels indicating research or industry stage This indicates the confidence level of the label.

[0243] The system can call a large language model to generate record-level semantic tags and their confidence scores based on the title and abstract fragments output from the main text channel, the keywords and entity fields output from the structured channel, the preset tag system, and the current task topic boundary.

[0244] After data collection and annotation are completed, the system stores the results in a hierarchical manner, including at least a text evidence layer, a structured metadata layer, and a semantic tag layer. The text evidence layer stores the title, abstract, key excerpts, conclusion excerpts, and source information; the structured metadata layer stores the year, institution, author, keywords, relational fields, and their sources; and the semantic tag layer stores information such as topic tags, method tags, application tags, and stage tags.

[0245] 3.5. Data Acquisition Log and Data Acquisition Status Output

[0246] After completing dual-channel data acquisition and hierarchical storage, the system generates acquisition logs and acquisition status objects. The acquisition logs record the source of the acquisition task, execution time, target data source, search expression, number of returned records, field coverage, and any abnormal states. The acquisition status object describes the data acquisition results for the current chapter, providing input for subsequent sufficiency assessments.

[0247] For the The collection status of a chapter-level retrieval plan object can be represented as follows:

[0248]

[0249] in, Indicates the number of text records. Indicates the number of structured records. Indicates the completeness rate of key fields. This indicates the source coverage or the hit rate of primary and secondary sources.

[0250] The output of this module includes at least: the main text evidence layer. Structured metadata layer Semantic tag layer Collection of logs and collection status objects for each chapter Subsequent modules can then perform data sufficiency checks, trigger supplementary data collection, and implement subsequent governance processes based on the above results.

[0251] Module 4: Data Sufficiency Assessment and Dynamic Supplementation / Backoff Triggering

[0252] After completing the dual-channel, multi-source, heterogeneous data acquisition and hierarchical storage, the system enters the data sufficiency assessment and dynamic supplementary acquisition / rollback triggering module. This module is used to evaluate whether the current acquisition results are sufficient to support the generation of the main text and charts for each chapter; when the requirements are not met, supplementary acquisition, data source switching, rollback, or target downgrading are triggered according to preset hierarchical relaxation rules.

[0253] This module assesses sufficiency by chapter, distinguishing between two categories: sufficiency of text generation and sufficiency of chart generation. Sufficiency of text generation focuses on the evidence required for text writing, while sufficiency of chart generation focuses on the fields and samples required for statistical analysis and visualization.

[0254] 4.1 Construction of Chapter-Level Sufficiency Determination Objects

[0255] The system reads the acquisition status object output by the previous module. Main text evidence layer Structured metadata layer Semantic tag layer and the corresponding chapter-level search plan objects And construct a chapter-level sufficiency determination object for each chapter:

[0256]

[0257] in, This represents a subset of text-based candidate data associated with this chapter. This represents a subset of structured field data associated with this chapter. This represents a subset of semantic tags associated with this chapter. This indicates the preset threshold and budget parameters in the chapter-level search plan object.

[0258] The system can call upon a large language model to filter and aggregate data related to the current chapter based on the chapter objectives, the collected text record summaries, structured field statistics, and semantic tag information, forming chapter-level judgment objects. .

[0259] 4.2. Sufficiency judgment of text generation

[0260] The system assesses the sufficiency of text generation based on chapter-level judgment objects. This assessment comprehensively considers the number of text records, record quality, topic coverage, time coverage, and writability.

[0261] No. The sufficiency score of the main text of each chapter can be expressed as:

[0262]

[0263] in, Indicates the satisfaction level of the number of text records. This indicates the percentage of high-quality main text. Indicates subtopic coverage. Indicates time coverage. Indicates writability, For the corresponding weights.

[0264] The satisfaction level of the number of text records can be expressed as:

[0265]

[0266] in, Indicates the current and the first The number of valid text records related to each chapter. This indicates the minimum threshold for text records required by this chapter.

[0267] The percentage of high-quality main text can be expressed as:

[0268]

[0269] in, This indicates the number of high-quality text records. To prevent extremely small constants with a denominator of zero.

[0270] The system can invoke a large language model, combining chapter objectives, candidate records in the main text, key segments, semantic tag distribution, and paragraph structure requirements, to assess the writability of the current evidence and output subtopic missing information and weakness descriptions. The system then quantifies and incorporates this result into... and In the middle, the sufficiency score of the main text is formed. .

[0271] Then the system will Compared with the preset text generation threshold Compare. When satisfied...

[0272]

[0273] If the current chapter is deemed ready for text generation, it is determined that the current chapter is ready for text generation; otherwise, the current chapter is deemed to have insufficient text evidence.

[0274] 4.3. Assessment of the sufficiency of chart generation

[0275] The system also assesses the sufficiency of chart generation to evaluate whether the structured field data of the current chapter is sufficient to support subsequent chart generation.

[0276] No. The chart sufficiency score for each chapter can be expressed as:

[0277]

[0278] in, Indicates the completeness rate of key fields. Indicates the satisfaction of sample size. Indicates the distribution efficiency. Indicates relation density. Indicates chart generability. For the corresponding weights.

[0279] The completeness rate of key fields can be expressed as:

[0280]

[0281] in, This indicates the set of core fields required to generate the charts in this section. Representation field Completeness rate Indicates the field weight.

[0282] The sample size satisfaction can be expressed as:

[0283]

[0284] in, This indicates the number of structured records currently available for chart generation. This indicates the minimum sample threshold required to generate chapter-level charts.

[0285] The system can invoke a large language model to evaluate the chart generation feasibility of the current data based on chapter objectives, structured field statistics, field completeness rate, category distribution, and preset chart types. It then outputs a field insufficiency warning, chart type adaptation suggestions, and, if necessary, chart degradation suggestions. The system quantifies and incorporates these results. , and In the middle, a chart sufficiency score is generated. .

[0286] Then the system will With chart generation threshold Compare. When satisfied...

[0287]

[0288] If the current chapter is ready, it is determined that the current chapter is ready for chart generation; otherwise, it is determined that the current chapter has insufficient structured fields or sample distribution.

[0289] 4.4 Dynamic replenishment, rollback, and downgrade trigger decisions

[0290] To obtain a score for sufficiency of the text Chart sufficiency score Then, the system combines the preset grading relaxation rules of the corresponding chapters. It generates dynamic replenishment, rollback, and downgrade decisions.

[0291] No. The triggering decision result for each chapter can be represented as:

[0292]

[0293] in, This indicates the action of adding text to the main body. This indicates a structured field supplementary data collection action. This indicates a data source rollback or expansion action. This indicates a downgrade action for chapter or chart objectives.

[0294] The trigger condition for supplementary text collection can be expressed as:

[0295]

[0296] The trigger condition for supplementary sampling of structured fields can be expressed as:

[0297]

[0298] when When the system prioritizes applying the relaxation rules related to the text channel, it will prioritize applying them. When the system prioritizes the use of relaxed rules related to structured field channels, the system can generate dual-channel supplementary acquisition tasks in parallel or determine the priority order based on chapter weight and target type.

[0299] The system can invoke a large language model to generate supplementary data collection path suggestions, rollback priority suggestions, and downgrade suggestions based on text sufficiency scores, figure and graph sufficiency scores, reasons for missing data, tiered relaxation rules, chapter objectives, and historical supplementary data collection status. The system then combines these suggestions with a rule base and historical supplementary data collection counts to form a final action set. .

[0300] The system can also be set to the maximum number of supplementary mining rounds or the maximum relaxation level. When a chapter reaches the maximum number of supplementary collection rounds but still fails to meet the text threshold or chart threshold, the system stops further expanding the search scope and switches to a downgraded path.

[0301] 4.5 Module Output

[0302] This module's output should include at least: chapter-level text sufficiency score. Chapter-level chart sufficiency score A collection of trigger actions for each chapter The current status of the supplementary sampling round and the updated execution path are explained.

[0303] For chapters that meet the threshold, the system marks them as generateable and proceeds to the subsequent data governance and evidence pool construction module; for chapters that do not meet the threshold, the system... New supplementary collection tasks or rollback tasks are generated and re-entered into the collection link for execution, thus forming a closed-loop processing flow oriented towards chapter objectives.

[0304] Module 5: Fusion of Supplementary Collection Results, Standardization of Governance, and Construction of Unified Evidence Entities

[0305] After completing the data sufficiency assessment and dynamic supplementary collection / rollback triggering, the system enters the supplementary collection result fusion, governance standardization, and unified evidence entity construction module. This module is used to perform unified governance on the initial collection results, supplementary collection results, and text records, structured field records, and semantic tag records generated from different channels, transforming multi-source heterogeneous, duplicate, or conflicting data into a unified set of evidence entities that can be directly invoked for subsequent chapter-level evidence allocation, text generation, chart generation, and citation generation.

[0306] This module mainly includes the fusion of supplementary sampling results and construction of candidate associations, deduplication and merging, field standardization, cross-source completion, conflict coordination, quality assessment and usage labeling.

[0307] 5.1 Fusion of Supplementary Data Collection Results and Construction of Candidate Associations

[0308] The system reads the main evidence layer output by the preceding module. Structured metadata layer Semantic tag layer The results of supplementary data collection for each chapter, along with the corresponding source information and collection logs, are then aggregated in a unified manner.

[0309] After aggregation, the system constructs a candidate association set based on information such as internal record identifier, source link identifier, title, author, time, and topic tags to identify records that may correspond to the same object.

[0310] For any two records and The system calculates the candidate association score:

[0311]

[0312] This indicates whether a unified identifier, link identifier, number identifier, or other directly matching hard identifier exists. Indicates title similarity. Indicates the similarity of authors or subject combinations. Indicates proximity in time. Indicates the semantic similarity of topics. For the corresponding weights.

[0313] when When the threshold is exceeded, the system will include the two records in the same candidate association cluster, forming a candidate association set:

[0314]

[0315] in, Indicates the first 1 candidate association cluster.

[0316] For record pairs that are difficult to adjudicate using rules, the system can call a large language model to combine title, summary, structured field fragments, source description and semantic tag information to make a semantic judgment on whether they belong to the same object or whether they can be attributed and used, and output association suggestions and association confidence.

[0317] 5.2 Deduplication and Integration with Unified Evidence Entities

[0318] After forming a set of candidate associations, the system performs deduplication and merging on each candidate association cluster to construct a unified evidence entity.

[0319] For candidate association clusters The system constructs a unified evidence entity:

[0320]

[0321] in, This represents the merge function. A unified evidentiary entity can be represented as:

[0322]

[0323] in, This indicates a unified internal evidence identification system. This represents the merged collection of text content. This represents the merged set of structured fields. This represents the merged set of semantic labels. This represents the set of source trajectories.

[0324] The source trajectory set is used to record which original sources, collection rounds, and supplementary collection actions formed the unified evidence entity.

[0325] The system preferably employs a merging strategy that combines hard matching, soft matching, and semantic adjudication. Hard matching prioritizes merging based on unique identifiers such as DOI, patent number, unified link identifier, and official number; soft matching uses fields such as title, author, institution, time, and keywords for approximate merging; when a determination is still not possible, a large language model is invoked to perform semantic judgment on the merging boundaries.

[0326] After consolidation, the system forms a unified set of evidence entities:

[0327]

[0328] 5.3 Field standardization, cross-source completion, and conflict resolution

[0329] After completing the deduplication and merging, the system further performs field standardization, cross-source completion, and conflict reconciliation on the unified evidence entity.

[0330] Field standardization is used to unify synonymous fields from different sources into a unified representation space. For unified evidence entities... The original set of fields, system-defined field standardization mapping:

[0331]

[0332] in, Represents the normalized mapping function, This represents the original set of fields before merging. This represents the standardized set of fields.

[0333] Taking the organization field as an example, it can be represented as:

[0334]

[0335] in, This represents the normalization function for the organization name. The time, country, keyword, and author fields can be processed using the corresponding normalized mapping functions.

[0336] Based on standardization, the system merges candidate field values ​​from different sources. This is applied to fields within a unified evidentiary entity. final value , can be represented as:

[0337]

[0338] in, Representation field Candidate values ​​from different sources, This represents the fusion function. The fusion function determines the final value by comprehensively considering source credibility, field completeness, time recentity, and contextual consistency.

[0339] When field values ​​from different sources conflict, the system further calculates the conflict resolution score for the candidate values:

[0340]

[0341] in, Indicates the credibility of the source. Indicates the completeness of the field content. This indicates consistency with other fields of the current unified entity. Indicates timeliness, For the corresponding weights.

[0342] The system selects the candidate value with the highest score as the main version, while retaining other candidate values ​​as information for future reference.

[0343] For synonym unification, field completion, and field conflict that are difficult to adjudicate directly by rules, the system can call the large language model to perform semantic judgment by combining the unified entity context and source information, and output field unification suggestions, completion suggestions, and conflict adjudication suggestions.

[0344] 5.4 Evidence Quality Assessment and Usage Labeling

[0345] After the unified evidence entity is formed and the fields are standardized, the system performs a quality assessment on each unified evidence entity and generates a usage label.

[0346] No. The overall quality score of a unified evidentiary entity can be expressed as:

[0347]

[0348] in, This indicates the relevance to the current task theme and chapter objectives. Indicates the availability of the main text content. Indicates the completeness of structured fields. Indicates the availability of references. This indicates consistency within a unified entity. For the corresponding weights.

[0349] In addition to the overall quality score, the system further generates usage-specific usability labels:

[0350]

[0351] in, Indicate whether it is suitable to be included in the main text evidence pool. This indicates whether the item is suitable for inclusion in the statistical evidence pool. This indicates whether the individual is suitable for inclusion in the relationship evidence pool. Indicate whether it is suitable to enter the subject evidence pool. Indicates whether it is suitable as a candidate for explicit citation.

[0352] The purpose label can be represented using binary labels or a hierarchical scoring method.

[0353] The system can call upon a large language model to determine the most suitable subsequent use of a unified evidence entity based on its standardized fields, text fragments, tag information, source trajectory, and semantic description of the current task or chapter. It can then output usage adaptation suggestions, usability level judgments, and risk warnings.

[0354] 5.5 Module Output

[0355] After completing the fusion, deduplication and merging, field standardization, cross-source completion, conflict reconciliation, quality assessment and usage labeling of the supplementary data collection results, this module outputs a unified set of evidence entities. Quality rating set Usability label set and governance log collection .

[0356] The module output can be represented as:

[0357]

[0358] in, Represents a unified set of evidentiary entities. Represents the set of quality scores. Represents a set of usage-based usability labels. This represents the set of governance logs.

[0359] The governance log at least records which original records were merged from the unified evidence entity, which fields were standardized, which fields came from supplementary collection, which fields had conflicts and their adjudication results, as well as the end-use label and quality level. Subsequent modules can directly base it on... Construct evidence pools for main text, statistics, relationships, and themes, and perform chapter-level evidence allocation. Unify evidence entities. It can be used directly to build an evidence pool to connect with Module Six.

[0360] Module Six: Construction of Purpose-Specific Evidence Pools and Preparation for Chapter-Level Evidence Allocation

[0361] After completing the fusion of supplementary data collection results, governance standardization, and the construction of unified evidence entities, the system enters the module for constructing purpose-based evidence pools and preparing chapter-level evidence allocation. This module is used to reorganize the unified evidence entity set, quality scoring results, and purpose-based usability annotation results into evidence pool objects that can be directly called by subsequent chapter generation, chart generation, citation generation, and graphic collaboration modules.

[0362] The evidence pools constructed in this module are no longer organized by data format, but by their subsequent use. Based on the usage-type labeling results of unified evidence entities, the system categorizes them into different evidence pools for direct use by subsequent modules.

[0363] 6.1 Rules for the Classification and Placement of Evidence Pools

[0364] After completing the fusion of supplementary collection results, standardization of governance, and construction of unified evidence entities, the system reads the governance results output by the fifth module. .

[0365] The system should at least construct a textual evidence pool, a statistical evidence pool, a relational evidence pool, and a topical evidence pool, denoted as:

[0366]

[0367] in, This indicates the evidence pool in the main text. Indicates the statistical evidence pool. Represents a pool of evidence related to relationships. This indicates the subject evidence pool.

[0368] Let the unified set of evidence entities be:

[0369]

[0370] For a single unified entity of evidence Its classification rules are based on the purpose-type label. and quality rating Jointly determined:

[0371]

[0372] in, This indicates whether the evidence entity has entered the main evidence pool. Indicates whether to include in the statistical evidence pool. Indicates whether to include it in the relationship evidence pool. Indicate whether to include it in the subject evidence pool. Indicates whether it is suitable as a candidate for explicit citation.

[0373] when and When the quality threshold of the evidence pool is not lower than that of the main text, Included in the main text evidence pool; when and When the quality threshold of the statistical evidence pool is not lower than the threshold, Included in the statistical evidence pool; when and When the quality threshold of the relational evidence pool is not lower than the threshold, Included in the relational evidence pool; when and When the quality threshold of the subject evidence pool is not lower than the threshold, It is included in the thematic evidence pool.

[0374] The same evidence entity can be entered into multiple evidence pools simultaneously based on its usage type label, in order to support its reuse in different modules later.

[0375] 6.2 Chapter Adaptation Marks and Allocation Preparation

[0376] After completing the evidence pool division, the system further performs chapter-level adaptation preparation to establish a pre-adaptation mapping between evidence entities and chapter targets.

[0377] For the For each piece of evidence entity, its chapter adaptation information can be represented as follows:

[0378]

[0379] in, Indicates the first The evidence entity for the first The degree of fit between the goals of each chapter.

[0380] The system combines chapter objective semantics, evidence topic tags, evidence usage type, and quality score to analyze... Calculations are performed to provide a pre-sorting basis for subsequent chapter-level evidence retrieval and allocation.

[0381] The system can invoke a large language model to determine the appropriate chapter direction for a given evidence entity based on its purpose-based annotation, topic tags, text content summary, structured field summary, and chapter objective description. It then outputs chapter adaptation suggestions and explanations. The system combines this result with the results of rule-based calculations to form a set of chapter adaptation mappings. .

[0382] 6.3 Module Output

[0383] After completing the construction of the purpose-specific evidence pool and the chapter-level adaptation preparation, this module outputs the main text evidence pool. Statistical evidence pool Relationship Evidence Pool Thematic evidence pool , and the pre-adapted mapping results between evidence entities and chapter objectives.

[0384] The module output can be represented as:

[0385]

[0386] in, Represents the evidence pool set. This represents the set of chapter adaptation maps.

[0387] Subsequent modules can be directly based on Perform chapter-level evidence allocation, text generation, chart generation, and citation generation.

Claims

1. A chapter-driven method for data acquisition and evidence pool construction in research reports, characterized in that, Includes the following steps: Step S1: Receive the original requirement text input by the user, perform normalization processing to obtain standardized semantic units, perform requirement semantic parsing to extract core task elements, complete and coordinate missing parameters and constraints, establish a data requirement model to calculate the requirement intensity of the main text data and the requirement intensity of the structured field data, and output the structured task object. Step S2: Read the structured task object, decompose the global task into multiple chapter-level search targets, generate an initial tight search strategy for each target and construct hierarchical relaxation rules, perform data source matching and master / slave configuration, and form a set of chapter-level search plan objects. Step S3: Read the set of chapter-level retrieval plan objects, expand them into executable collection tasks, execute the text data collection channel and the structured field collection channel in parallel, generate semantic tags and store them in layers, and output the collection status objects and initial collection results corresponding to each chapter. Step S4: Based on the collection status object, the hierarchically stored initial collection results and retrieval plan object, construct a chapter-level sufficiency judgment object, calculate the sufficiency score of the main text generation and the sufficiency score of the chart generation respectively, compare them with the preset threshold, and generate dynamic supplementary collection, rollback and downgrade trigger decisions according to the comparison results and the hierarchical relaxation rules, and execute the corresponding supplementary collection task according to the trigger decision to obtain the supplementary collection result. Step S5: Aggregate the initial collection results and supplementary collection results, calculate the candidate association scores between records to construct candidate association clusters, remove duplicates from the candidate association clusters and form unified evidence entities, perform field standardization, cross-source completion and conflict coordination, conduct evidence quality assessment and usage labeling, and output a unified evidence entity set, quality score and usage-type usability labeling. Step S6: Based on the usage-type availability label, the unified evidence entities are divided into different types of usage-type evidence pools. The degree of adaptation of each evidence entity to the chapter target is calculated to form a chapter adaptation mapping. The evidence pool set and the pre-adaptation mapping result are output.

2. The method according to claim 1, characterized in that, Step S1 includes: Step 1.1, Receiving and Normalizing Original Requirements: Perform normalization mapping on the original requirement text sequence to obtain a normalized requirement set. ,in , This indicates the input related to the task in the current or previous round. Represents the normalized mapping function. Represents standardized semantic units, ; Step 1.2, Requirement Semantic Parsing and Task Element Extraction: Extract core task elements and organize them into task semantics: in, Indicates thematic elements, Representing domain elements, Indicates the report type. Indicates the scope of analysis. Indicates the target audience, This indicates the key points. Indicates the output format. To indicate the need for charts, Indicates a citation requirement, Indicates style preference; Step 1.3, Missing Parameter Completion and Constraint Coordination: ... Each task parameter is represented as ,in Indicates the parameter value. Indicates the source identifier. This indicates the priority weight, and the conflicting parameters are coordinated based on the source identifier and priority weight; Step 1.4, Data Requirements Modeling and Collection Direction Generation: Calculate the data requirement intensity for the main text. and the intensity of demand for structured field data The formula is: in This indicates the factors influencing the demand for the main text of the report. Indicates the citation strength requirement. Indicates the required depth of content. This indicates the need for the main body of the text to elaborate on the key analytical content; Indicates the strength of the chart preference. This indicates a need for trend analysis. Indicates the main analysis needs, Indicates the statistical requirements for distribution; For preset weights or learnable weights; and based on and The size relationship triggers the corresponding data acquisition strategy.

3. The method according to claim 1, characterized in that, Step S2 includes: Step 2.1, Expanding Search Targets and Mapping Chapters: Forming a set of chapter-level search targets. ,in , Indicates chapter direction or semantic tags. Indicates the data usage type. This indicates the set of fields that are preferentially depended upon. Indicates weight; Step 2.2, Initial Search Strategy Generation: Constructing the initial search strategy ,in This represents the initial search expression. Represents a set of range constraints. Represents the target data source collection. This indicates the set of fields to be returned; and the core query is selected based on the candidate search expression adaptation score, with the scoring formula being: in Indicates semantic relevance. Indicates field coverage capability. Indicates the degree to which range constraints are satisfied. For the corresponding weights; Step 2.3, Construction of Tiered Relaxation Rules: Construct a set of tiered relaxation rules. , of which Level relaxation rules , This indicates a relaxation operation on the expression. This indicates a range constraint relaxation operation. This indicates that the field requirements have been relaxed. Indicates a data source extension operation; Step 2.4, Data Source Matching and Plan Formulation: Calculate the candidate data source matching score: in For domain matching degree, As to the degree of field support, To ensure the credibility of the source, To ensure timeliness, For the corresponding weights; Output chapter-level search plan object: in, Indicates chapter-level search target. Indicates the initial search strategy. This represents the set of tiered relaxation rules. Represents the main data source collection. Represents a set of secondary data sources. This indicates that a collection of fields will be returned. This indicates the basic threshold configuration or expected data collection budget required for subsequent sufficiency assessments. Step 2.5: Final output of the collection of chapter-level search plan objects: Each Each corresponds to a chapter-level search target and includes its initial search strategy, relaxed rules, data source configuration, and basic threshold configuration.

4. The method according to claim 1, characterized in that, Step S3 includes: Step 3.1, Task Deployment: Expand the chapter-level search plan object into a set of data collection tasks. Subtasks , Indicates the target data source. This represents a search expression. Indicates the required return fields. Indicates the channel type. Indicates the expected collection volume; Step 3.2, Main Text Channel Execution: Extract Main Text Data: Among them, a single record , For recording identification, As the title, For the abstract, For key segments, This is the conclusion segment. For the source, For time; Step 3.3, Structured Channel Execution: Extracting Structured Data: Among them, a single record Each field represents the year, author, institution, country, keywords, relational fields, and source, respectively. Step 3.4, Tag Generation and Status Output: Generate semantic tags: Indicates the corresponding record identifier. Indicates topic tags, Indicates method label, Indicates the application tag, Labels indicating research or industry stage Indicates the confidence level of the label; Step 3.5, and output the acquisition status. ; in, Indicates the number of text records. Indicates the number of structured records. Indicates the completeness rate of key fields. This indicates the source coverage or the hit rate of primary and secondary sources.

5. The method according to claim 1, characterized in that, Step S4 includes: Step 4.1: Read the acquisition status object Main text evidence layer Structured metadata layer Semantic tag layer and the corresponding chapter-level search plan objects And construct a chapter-level sufficiency determination object for each chapter: ； in, This represents a subset of text-based candidate data associated with this chapter. This represents a subset of structured field data associated with this chapter. This represents a subset of semantic tags associated with this chapter. This indicates the preset threshold and budget parameters in the chapter-level search plan object; Step 4.2, Text Sufficiency Assessment: Calculate the score: ， ， in, Indicates the satisfaction level of the number of text records. This indicates the percentage of high-quality main text. Indicates subtopic coverage. Indicates time coverage. Indicates writability, For the corresponding weights, This represents the minimum threshold required for the main text records. For high-quality record count, To prevent the division into zero minimum constants; Will Compared with the preset text generation threshold Compare. When satisfied... If the current chapter is deemed ready for text generation, it is determined that the current chapter is ready for text generation; otherwise, the current chapter is deemed to have insufficient text evidence. Step 4.3, Assessment of Chart Sufficiency: Calculate the score: ， ， in, Indicates the completeness rate of key fields. Indicates the satisfaction of sample size. Indicates the distribution efficiency. Indicates relation density. Indicates chart generability. For the corresponding weights, For the core field set, For field completeness rate, For field weights, The minimum sample threshold for the chart; Will With chart generation threshold Compare. When satisfied... If the current chapter is ready, determine that it is ready for chart generation; otherwise, determine that the current chapter has insufficient structured fields or sample distribution. Step 4.4, Dynamic Trigger Decision: Generate trigger decision: ， Positioned based on the comparison result between the score and the threshold. and And perform the corresponding action; in, This indicates the action of adding text to the main body. This indicates a structured field supplementary sampling action. This indicates a data source rollback or expansion action. This indicates a downgrade action for chapter or chart objectives.

6. The method according to claim 1, characterized in that, Step S5 includes: Step 5.1, Fusion and Association Construction: Calculate any two records Association score: Those exceeding the threshold are included in the candidate association set: ， in This indicates a hard identifier match. Indicates title similarity. Indicates the similarity of authors or subject combinations. Indicates proximity in time. Indicates the semantic similarity of topics. For the corresponding weights, Indicates the first One candidate association cluster; Step 5.2, Deduplication and Merging: Execute the merge function to construct a unified evidence entity: ； Represents the merge function. This indicates a unified internal evidence identification system. This represents the merged collection of text content. This represents the merged set of structured fields. This represents the merged set of semantic labels. Represents the set of source trajectories; Step 5.3, Standardization and Conflict Reconciliation: Perform field mapping: This represents the original set of fields before merging. This represents the standardized set of fields, merging field values. Score based on conflict resolution: in, Indicates the credibility of the source. Indicates the completeness of the field content. This indicates consistency with other fields of the current unified entity. Indicates timeliness, For the corresponding weights; The final values ​​are determined, with each item representing source credibility, content completeness, entity consistency, and timeliness, respectively. For the corresponding weights; Step 5.4, Quality Assessment and Labeling: Calculate the quality score: in, This indicates the relevance to the current task theme and chapter objectives. Indicates the availability of the main text content. Indicates the completeness of structured fields. Indicates the availability of references. This indicates consistency within a unified entity. For the corresponding weights; And generate usage labels: in, Indicate whether it is suitable to be included in the main text evidence pool. This indicates whether the item is suitable for inclusion in the statistical evidence pool. This indicates whether the individual is suitable for inclusion in the relationship evidence pool. Indicate whether it is suitable to enter the subject evidence pool. Indicates whether it is suitable as a candidate for explicit citation.

7. The method according to claim 1, characterized in that, Step S6 includes: Step 6.1, based on unified evidence entity Usage label Unified evidence entities will be included in the main text evidence pool. Statistical evidence pool Relationship Evidence Pool Thematic evidence pool The set that constitutes: ， Step 6.2, Adaptation Mapping: Calculate Evidence Entities For the The degree of fit of the chapter objectives This generates adaptation information: ， Step 6.3, Final Output .

8. A chapter-driven research report data acquisition and evidence pool construction system, comprising a processor and a storage medium, characterized in that, When the processor executes a program in the storage medium, it implements the method as described in any one of claims 1-7.

Images