A structured outline driven scientific report automatic generation method and system

Abstract

The application provides a structured outline driven scientific report automatic generation method and system, and relates to the technical field of natural language processing, the method comprises the following steps: performing semantic analysis and matching on a user input research topic to generate a first-level structure node, extracting and dividing research problems from multiple standard scientific literatures to construct a second-level structure node, performing opinion sentence extraction and refinement to generate a third-level structure node, generating a structured outline according to the first-level, second-level and third-level structure nodes and corresponding explicit binding relationships, performing task analysis, generating a task set according to the hierarchical order of the structured outline, generating a target scientific report and outputting the target scientific report by performing text generation, assembling and integrating. The technical problems of unstable structure, easy divergence of theme, difficulty in corresponding paragraphs and evidence, and easy repetition of chapter content in the existing scientific report automatic generation are solved. The technical effect of stable structure, focused theme and traceable content of the scientific report automatic generation is achieved.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, specifically to a method and system for automatically generating scientific and technological reports based on a structured outline. Background Technology

[0002] In scientific report generation tasks, the input is typically a collection of scientific literature revolving around the same research topic. The key technical challenge lies not in generating single sentences or short texts, but in abstracting a reasonable chapter structure, research sub-directions, and the core content to be discussed in each section from the literature collection. Traditional report generation methods, relying on manually defined templates, usually pre-set the report's chapter framework, such as research background and progress, and fill in the corresponding chapters with the titles, abstracts, or keywords of the literature. While these methods are structurally stable and simple to implement, they heavily depend on manually defined templates and struggle to automatically adjust the chapter structure based on changes in the semantic distribution of literature under different research topics. This makes it impossible to guarantee that the generated chapter structure accurately reflects the distribution of research topics within the literature collection.

[0003] Existing end-to-end generation methods based on large language models typically involve constructing prompt words, inputting the research topic or part of the literature content into the model, and then having the model directly generate the entire or large sections of report text. While these methods can generate fluent text, their generation process lacks stable intermediate structural representations and explicit modeling of the overall semantic distribution of the literature collection. When the number of documents is large or the research scope is complex, this can easily lead to instability in the generated results in terms of chapter structure, topic coverage, and logical consistency.

[0004] In summary, existing technologies for automatically generating scientific and technological reports suffer from several technical problems, including structural instability, divergent themes, difficulty in matching paragraphs with evidence, and repetition within chapters due to template-dependent or end-to-end generation. Summary of the Invention

[0005] The purpose of this application is to provide a method and system for automatically generating scientific and technological reports based on a structured outline, in order to solve the technical problems existing in the automatic generation technology of scientific and technological reports, such as unstable structure, divergent themes, difficulty in matching paragraphs with evidence, and easy repetition within chapters.

[0006] In view of the above problems, this application provides a method and system for automatically generating scientific and technological reports based on a structured outline.

[0007] The first aspect of this application provides a method for automatically generating scientific and technological reports based on a structured outline. The method includes: receiving multiple scientific and technological documents on the same research topic; preprocessing the multiple documents to obtain multiple standard scientific and technological documents; storing the multiple standard documents in a data storage server; performing semantic analysis and matching on a user-input research topic to generate first-level structural nodes; extracting and dividing research questions from the multiple standard documents according to the first-level structural nodes to construct second-level structural nodes; extracting and refining viewpoint sentences based on the second-level structural nodes to generate third-level structural nodes; collecting viewpoint sentences from the third-level structural nodes to generate core viewpoints and bind them with evidence, establishing explicit binding relationships; representing the first-level structural nodes, second-level structural nodes, third-level structural nodes, and the corresponding explicit binding relationships in a hierarchical data structure to generate a structured outline; performing task parsing and modeling on the structured outline to construct a set of section generation tasks; generating and assembling the main text of the set of section generation tasks according to the hierarchical order of the structured outline to generate and output the target scientific and technological report.

[0008] Optionally, multiple primary research topic category description texts are predefined; semantic analysis and matching quantification are performed on the user-input research topic based on the multiple primary research topic category description texts, and a semantic relevance score is calculated; primary research topics in the multiple primary research topic category description texts whose semantic relevance scores are higher than a preset threshold are retained to obtain research intent analysis results; the research intent analysis results are used as target research directions to generate primary structure nodes.

[0009] Optionally, the multiple standard scientific and technological documents are uniformly assigned and redistributed according to the first-level structural nodes to obtain multiple classified scientific and technological documents; research questions are extracted from the multiple classified scientific and technological documents belonging to the same first-level structural node to obtain research question text; semantic feature vectors are constructed by combining the research question text and document metadata, wherein the document metadata includes document title, abstract and keywords; semantic clustering is performed on the semantic feature vectors to construct second-level structural nodes.

[0010] Optionally, author viewpoint sentences are extracted from the collection of scientific and technological documents corresponding to the secondary structure nodes; the author viewpoint sentences and the research question text are used as semantic feature inputs to perform semantic clustering, while semantic information of the same-level secondary structure nodes is introduced as mutual exclusion constraints to generate tertiary structure nodes.

[0011] Optionally, viewpoint sentences belonging to the same third-level structural node in the three-level structural nodes are aggregated to obtain original viewpoint sentences; the original viewpoint sentences are input into a large language model for induction and summarization to generate core viewpoints of the nodes, and an explicit binding relationship is established between the core viewpoints of the nodes, the original viewpoint sentences, and the document identifiers.

[0012] Optionally, the structured outline is parsed and split to obtain multiple section generation units; a data structure field is defined for each section in the multiple section generation units, the data structure field including a unique identifier, paragraph type, title path and core viewpoint information; task modeling is performed based on the data structure fields of the multiple section generation units to construct a set of section generation tasks.

[0013] Optionally, based on the explicit binding relationship, the original text of the viewpoint sentences and the document metadata corresponding to the multiple section generation units are retrieved from the data storage server to form a paragraph-level evidence set; a generation prompt word template is constructed according to the type of the multiple section generation units; the paragraph-level evidence set is used as a generation constraint, and a large language model is called to generate the main text of the section generation task set based on the prompt word template and the generation constraint to obtain an initial section generation text set; the initial section generation text set is assembled and integrated according to the hierarchical order of the structured outline to generate the target scientific and technological report and output it.

[0014] Optionally, the citation format in the initial section generated text set is standardized to obtain a standard section generated text set; the standard section generated text set is then assembled and integrated according to the hierarchical order of the structured outline to generate and output the target scientific and technological report.

[0015] Optionally, regular expression matching and structured parsing are performed on the citation format in the initial section generated text set to convert the identified citation tags into a preset format, and rule-based conversion is performed on multiple citations in parallel and compound citation strings to obtain a standard section generated text set.

[0016] A second aspect of this application provides an automatic technology report generation system based on a structured outline, the system comprising: a document input module for receiving multiple scientific documents on the same research topic, preprocessing the multiple scientific documents to obtain multiple standard scientific documents, and storing the multiple standard scientific documents in a data storage server; a structure node generation module for performing semantic analysis and matching on the user-input research topic to generate primary structure nodes, extracting and dividing research questions from the multiple standard scientific documents according to the primary structure nodes, and constructing secondary structure nodes; and a viewpoint extraction module for extracting viewpoint sentences based on the secondary structure nodes. The system is refined to generate a three-level structure node. The viewpoint sentences from these three-level structure nodes are then aggregated to generate core viewpoints and bind them with evidence, establishing explicit binding relationships. A structured outline generation module is used to represent the first-level, second-level, and third-level structure nodes and their corresponding explicit binding relationships in a hierarchical data structure to generate a structured outline. A report integration and output module is used to perform task parsing and modeling on the structured outline, construct a set of section generation tasks, and generate, assemble, and integrate the main text of the section generation task set according to the hierarchical order of the structured outline to generate and output the target scientific and technological report.

[0017] One or more technical solutions provided in this application have at least the following technical effects or advantages: The method provided in this application embodiment receives multiple scientific and technological documents on the same research topic, preprocesses these documents to obtain multiple standard scientific and technological documents, and stores these documents on a data storage server; performs semantic analysis and matching on the user-input research topic to generate first-level structural nodes; extracts and divides the research questions from the multiple standard scientific and technological documents according to the first-level structural nodes to construct second-level structural nodes; extracts and refines viewpoint sentences based on the second-level structural nodes to generate third-level structural nodes; gathers the viewpoint sentences from the third-level structural nodes to generate core viewpoints and bind them with evidence, establishing explicit binding relationships; represents the first-level structural nodes, second-level structural nodes, third-level structural nodes, and corresponding explicit binding relationships in a hierarchical data structure to generate a structured outline; performs task parsing and modeling on the structured outline to construct a set of section generation tasks; generates and assembles the main text of the set of section generation tasks according to the hierarchical order of the structured outline to generate and output the target scientific and technological report. By introducing a structured outline as the core intermediate representation in the process of generating scientific and technological reports, the document collection is first organized hierarchically and semantically, and then the report text is generated under the constraints of this structure. This achieves the technical effect of automatically generating scientific and technological reports with stable structure, focused theme, and traceable content, thereby improving the controllability of automatic generation of scientific and technological reports and the quality of content generation.

[0018] The above description is merely an overview of the technical solution of this application. To better understand the technical means of this application and to facilitate its implementation according to the description, and to make the above and other objects, features, and advantages of this application more apparent, specific embodiments of this application are described below. It should be understood that the content described in this section is not intended to identify key or important features of the embodiments of this application, nor is it intended to limit the scope of this application. Other features of this application will become readily apparent through the following description. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0020] Figure 1 This application provides a flowchart illustrating an automatic technical report generation method based on a structured outline.

[0021] Figure 2 This is a schematic diagram of a structured outline-driven automatic technical report generation system provided in this application.

[0022] Figure labeling: Document input module 11, structure node generation module 12, viewpoint extraction module 13, structured outline generation module 14, report integration and output module 15. Detailed Implementation

[0023] This application provides a method and system for automatically generating scientific and technological reports based on a structured outline, addressing the technical problems in existing automatic scientific and technological report generation technologies, such as unstable structure, divergent themes, difficulty in matching paragraphs with evidence, and repetitive content within chapters. By introducing a structured outline as the core intermediate representation in the scientific and technological report generation process, the document set is first hierarchically semantically organized, and then the report text is generated under the constraints of this structure. This achieves the technical effect of automatically generating scientific and technological reports with stable structure, focused themes, and traceable content, thereby improving the controllability and content quality of automatic scientific and technological report generation.

[0024] The technical solutions of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. It should be understood that the present invention is not limited to the exemplary embodiments described herein. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention. It should also be noted that, for ease of description, only the parts related to the present invention are shown in the accompanying drawings, not all of them.

[0025] Example 1, as Figure 1 As shown, this application provides a method for automatically generating scientific and technological reports based on a structured outline, which includes: The system receives multiple scientific and technological documents on the same research topic, preprocesses these documents to obtain multiple standard scientific and technological documents, and stores these standard scientific and technological documents in a data storage server.

[0026] Specifically, relevant literature can be downloaded from online academic resources such as Google Scholar, IEEE Xplore, PubMed, CNKI and other databases, or literature can be searched using keywords, research fields or specified research questions to receive multiple scientific and technological documents on the same research topic. These multiple scientific and technological documents must include at least the title, abstract, author, publication year and full or part of the main text.

[0027] After receiving the documents, multiple scientific and technological documents undergo preprocessing to ensure the quality and consistency of the data. This preprocessing includes deduplication, format standardization, and metadata parsing. Deduplication is crucial to avoid duplicate content in subsequent processing, improving efficiency and accuracy. Deduplication is performed using unique identifiers such as DOI and document title to eliminate duplicate uploads or documents. Documents from various sources may have different formats and layouts. Format standardization is applied to multiple scientific and technological documents, including standardizing the font and layout of document titles and abstracts, regulating the order and format of author names (e.g., first name followed by last name), and standardizing date formats (e.g., converting 2021-05 to May 2021). This ensures that parsing errors do not occur during subsequent analysis due to inconsistent formats. Simultaneously, the metadata of the documents is parsed. Metadata describes basic information about the documents, such as title, abstract, authors, and publication year. Natural language processing techniques and regular expressions are used to parse the metadata. For the title, it is identified by looking for specific formats or keywords at the beginning of the document. The abstract is usually after the title and has specific paragraph characteristics. It is extracted by analyzing the paragraph structure and keywords. The author information is contained at the beginning or end of the document and is guided by keywords such as "author" or "authors". The author's name is obtained by matching keywords and their specific content using regular expressions. The publication year is determined by looking for time-related keywords in the document, such as year, month, and day, and combining them with contextual information.

[0028] Through preprocessing, multiple standard scientific and technological documents are obtained and stored on a data storage server. This server manages all preprocessed documents, and the storage process includes storing the preprocessed document data in standardized formats such as JSON, XML, or database formats. Simultaneously, to ensure convenient access to the stored document data, an API interface is designed to provide document retrieval, query, and update functions.

[0029] Semantic analysis and matching are performed on the research topic input by the user to generate primary structure nodes. Research questions are extracted and divided from the multiple standard scientific and technological documents according to the primary structure nodes to construct secondary structure nodes.

[0030] Furthermore, generating primary structure nodes includes: predefining multiple primary research topic category description texts; performing semantic analysis and matching quantification on the user-input research topic based on the multiple primary research topic category description texts, and calculating semantic relevance scores; retaining primary research topics in the multiple primary research topic category description texts whose semantic relevance scores are higher than a preset threshold, and obtaining research intent analysis results; using the research intent analysis results as target research directions to generate primary structure nodes.

[0031] Specifically, for specific application areas, several primary research topics are predefined based on the specific characteristics of the field. Each primary research topic corresponds to a clear research scope or focus. The primary research topic not only describes the key areas of the field but also uses a set of descriptive texts to describe the research content covered in that direction, forming multiple primary research topic category description texts. For example, in the field of Dunhuang studies, primary research topics include, but are not limited to, mural research, grotto conservation, earthen site conservation, management and promotion, etc. By predefining multiple primary research topic category description texts, the system can determine the research direction to which the user's input research topic belongs and make a reasonable classification. The primary research topics and their category descriptions can be configured and expanded according to specific fields or user needs to ensure adaptability to scientific research work in different fields.

[0032] When a user inputs a research topic, semantic analysis is performed. Natural language processing (NLP) techniques are used to understand the specific meaning of the input. First, the text of the input research topic is analyzed, and the part-of-speech tagging of each word is performed. Then, word vector models such as Word2Vec, GloVe, and BERT are used to convert the input research topic into a vector representation. The input research topic is then matched with multiple primary research topic category description texts, and semantic similarity is calculated between them to quantify the relevance between the research topic and each primary research topic. During the semantic matching process, a semantic relevance score is calculated to quantify the similarity between the input research topic and the description texts of each primary research topic category. For example, cosine similarity is used, which calculates the angle between two vectors to obtain the similarity between them. A higher cosine similarity indicates that the input topic and the predefined primary research topics are semantically closer. Through calculation, a semantic relevance score is generated for each primary research topic, representing the degree of matching between the input research topic and that primary research topic.

[0033] A preset threshold, such as 0.7, is set. The semantic relevance scores of multiple primary research topic category description texts are compared with this threshold. Primary research topics with semantic relevance scores greater than the threshold are selected and retained as research intent analysis results. This serves to characterize the core research area corresponding to the current scientific report, thus preventing the generated research topics from deviating from the user's needs. For example, if a user inputs "mural protection," matching analysis prioritizes retaining relevant literature under the themes of "mural research" and "management and promotion." The preset threshold can be based on a validation set constructed from historical sample data. Statistical analysis is performed on the semantic relevance scores between the user-input research topic and the primary research topic category description texts. The balance between recall and precision under different thresholds is analyzed by plotting score distribution curves or ROC curves. When historical data is lacking, an empirical threshold, such as 0.6-0.8, can be used as an initial value, and manually calibrated in conjunction with domain expert annotation results. The score that yields the best classification effect or the highest F1 score is selected as the baseline threshold.

[0034] The results of the research intent analysis are used as the target research direction to generate the first-level structural nodes of the structured outline. The structured outline, before the scientific report is generated, is a systematic plan and organization of the report's chapter hierarchy, research topic distribution, and core content to be covered in each section, based on the set of literature to be analyzed. Unlike a simple table of contents used only for formatting or writing prompts, a structured outline typically includes not only chapter titles but also implicitly the hierarchical relationships between chapters, topic boundaries, and content organization logic. The first-level structural nodes represent the report's main research direction or chapter theme.

[0035] Through semantic analysis and matching, the determination of primary research topics no longer relies solely on the static classification labels of documents, but is constrained by the semantic understanding of research objectives. This limits the semantic organization space of the document set during the structured outline construction stage, avoiding problems such as divergent research topics or unclear chapter boundaries in the subsequent structural division process.

[0036] Furthermore, constructing a secondary structure node includes: uniformly classifying and reassigning the multiple standard scientific and technological documents according to the primary structure node to obtain multiple categorized scientific and technological documents; extracting research questions from the multiple categorized scientific and technological documents belonging to the same primary structure node to obtain research question text; constructing a semantic feature vector by combining the research question text and document metadata, wherein the document metadata includes document title, abstract, and keywords; and performing semantic clustering on the semantic feature vector to construct a secondary structure node.

[0037] Specifically, based on the primary structural nodes, multiple standard scientific and technological documents are uniformly assigned and redistributed. This means that, according to the research direction corresponding to the primary structural nodes, the original classification results of the standard scientific and technological documents are assigned and redistributed, remapping the documents to the primary structural node that best matches their content, and correcting the original primary classification results when necessary. Specifically, when standard scientific and technological documents already have primary classification labels during data preprocessing, these labels are used as the initial classification basis. When the original classification of standard scientific and technological documents is inconsistent with the target research direction, the semantic attribution result of the research topic is given priority, and the standard scientific and technological documents are uniformly assigned to the primary structural node corresponding to the target research direction, resulting in multiple categorized scientific and technological documents. For example, if a document was originally labeled as belonging to the category of earthen site protection, but the topic is more relevant to mural research in the user-input research direction, then the document is classified under the mural research node, ensuring that the document is accurately classified into the field that matches the current research topic.

[0038] After the literature is classified, rule matching based on keywords is used, such as identifying questions such as "the problem is," "the key challenge is," "urgently needs to be solved," and "this paper studies." Combined with deep learning sequence labeling models, such as the BiLSTM-CRF model, research questions are extracted from multiple classified scientific and technological documents belonging to the same primary structure node. That is, key information such as explicitly stated research questions, research hypotheses, technical difficulties, or unresolved problems are identified from the main text or abstract of the documents to form research question text. Among them, the research question usually directly reflects the scientific or technical problem that the document attempts to solve and is an effective semantic feature for characterizing the research sub-direction. The research question text refers to sentences or paragraphs that can directly reflect the core research focus of the document.

[0039] The research question texts of all categorized scientific and technological literature under the first-level node are extracted and combined with the literature metadata to form semantic feature vectors through vector concatenation. The literature metadata includes the literature title, abstract, and keywords. The semantic feature vectors are used to characterize the research content features of the literature under the current first-level structural node. Then, semantic clustering is performed on the semantic feature vectors. The constructed semantic feature vectors are input into a large language model. Leveraging the powerful semantic understanding and text generation capabilities of the large language model, automatic merging and clustering analysis of the literature research content is achieved. During the clustering process, for all literature under the same first-level structural node, its semantic feature vectors and their corresponding research question texts and literature metadata are used as input. Specialized prompts are designed to guide the large language model to identify the semantic similarity and research topic relevance between the literatures. For example, prompts could be such as, "Please perform topic clustering analysis on the following multiple literatures belonging to the same research field. The analysis requirements are: based on the core research questions and technical methods of the multiple literatures, identify their semantic similarities and divide them into several clusters, each cluster representing a relatively independent research sub-direction." The large language model, based on a comprehensive understanding of the deep semantic features of the research question's core focus, technical approach, and research object, automatically aggregates semantically similar and research-question-similar documents into the same semantic cluster. The goal of the clustering process is not to generate an arbitrary number of topics, but rather, based on a thorough understanding of the document content, to aggregate semantically similar and research-question-similar documents into the same cluster, forming several semantic clusters covering the main research subfields of that research direction. Each semantic cluster corresponds to a second-level node in a structured outline, representing a relatively stable research sub-direction, thus forming a second-level structural node. This second-level structural node not only contains clustering topic labels but also associates the corresponding document set and research question set.

[0040] By using clustering based on research questions, the secondary structure of the structured outline can directly reflect the distribution of different research routes or technical directions within the field, rather than being determined by manual templates or random generation, thereby achieving a refined organization and structured expression of research questions.

[0041] Based on the second-level structural nodes, the viewpoint sentences are extracted and refined to generate third-level structural nodes. The viewpoint sentences of the third-level structural nodes are then collected to generate core viewpoints and bind evidence, establishing explicit binding relationships.

[0042] Furthermore, generating a third-level structure node includes: extracting author viewpoint sentences from the collection of scientific and technological documents corresponding to the second-level structure node; performing semantic clustering by using the author viewpoint sentences and the research question text as semantic feature inputs, while introducing semantic information from the same-level second-level structure nodes as mutual exclusion constraints to generate a third-level structure node.

[0043] Specifically, for each secondary structural node, author opinion sentences are extracted from the collection of scientific literature. This involves identifying sentences in the text that embody the author's core stance, methodological summary, or experimental conclusions. These sentences typically contain obvious opinion-expressing keywords, such as "this paper believes," "experiments show," "results show," "we propose," and "research findings." These keywords are used as rule triggers, and a large language model is used to identify opinion sentences. Sentences are categorized based on their viewpoints, such as: the author's viewpoint revealing the practical significance and application value of the research problem; the author's consensus on existing research related to the research problem; and the author's viewpoint on the limitations of current research. Specific prompts, such as "Please extract author opinion sentences from the following literature paragraphs and classify them according to the specified categories. Requirements: Only extract sentences that clearly express the author's stance, judgment, or conclusion (i.e., opinion sentences); assign only one most suitable category to each sentence," guide the large language model to automatically classify opinion sentences while identifying them. This constructs a structured set of opinion sentences, allowing for the use of different types of opinion sentences when generating text for different types of paragraphs.

[0044] After extracting the author's viewpoint sentence, the extracted author's viewpoint sentence and the corresponding research question text under the second-level structural node are used as semantic input elements. The research question text represents the core focus of the subfield, while the author's viewpoint sentence represents the specific literature's solution path and conclusion. Both are vectorized using the same semantic encoding model, and vector concatenation is used to form the semantic feature input. Semantic clustering analysis is then performed on the semantic feature input, such as using a large language model for grouping. Mutual exclusion constraints are introduced during semantic clustering. These constraints use the semantic information of other second-level nodes under the same first-level structural node as constraints, requiring that the generated third-level structural nodes avoid semantic overlap as much as possible. This improves the distinguishability between third-level structural nodes and reduces the problem of content repetition in different paragraphs of the report.

[0045] Through clustering, several semantically more refined clustering units are formed. Each clustering unit corresponds to a tertiary structure node, enabling the formation of several tertiary structure nodes under each secondary structure node. Each tertiary structure node corresponds to a specific paragraph in the report. By combining viewpoint extraction with constrained semantic clustering, the research content under the secondary nodes is further subdivided into more specific viewpoint hierarchies, thereby enhancing the granularity and logical expression of the structured outline.

[0046] Furthermore, establishing explicit binding relationships includes: aggregating viewpoint sentences belonging to the same third-level structural node among the three-level structural nodes to obtain original viewpoint sentences; inputting the original viewpoint sentences into a large language model for induction and summarization to generate node core viewpoints, and establishing explicit binding relationships between the node core viewpoints, the original viewpoint sentences, and document identifiers.

[0047] Specifically, for each third-level structural node, all related viewpoint sentences within that node are aggregated to form an original viewpoint sentence set. These viewpoint sentences originate from the core authorial expressions extracted from the main text of the literature and contain characteristic words indicating research conclusions or methodological evaluations, such as "indicates," "verifies," "explains," "results show," "experimental results prove," and "further findings," to enhance the semantic reliability of viewpoint sentence identification. During the aggregation process, the document identifier corresponding to each viewpoint sentence and its position index in the original text are recorded simultaneously. Document identifiers include DOI, document ID, or database unique identifiers.

[0048] The original viewpoint sentences under the same three-level structural node are used as input. Semantic induction and information compression are performed by calling a large language model. The large language model is optional, such as DeepSeek, LLAMA, Qwen, and GPT. The large language model explicitly requires it to comprehensively analyze multiple viewpoint sentences in the prompt words and extract one or more core viewpoints of the node that can summarize the research content of the node. The core viewpoint of the node refers to the high-level semantic expression formed by summarizing the common content of multiple documents in the semantic cluster, rather than a simple restatement of the conclusion of a single document.

[0049] To ensure the interpretability of the generated content, the correspondence between the core viewpoint and the original viewpoint sentence is preserved during the core viewpoint generation process. Each core viewpoint is associated with the viewpoint sentence identifier on which it is based, establishing an explicit binding relationship between the node core viewpoint, the original viewpoint sentence, and the document identifier. The explicit binding relationship can be stored using a graph structure or a triple structure, in the form of: core viewpoint - viewpoint sentence (evidence) - document, to achieve a searchable and traceable structured association.

[0050] By summarizing and integrating viewpoints and information scattered across multiple documents, and establishing a strong correlation between core viewpoints and sources of evidence through an explicit structured binding mechanism, the generated scientific and technological reports possess clear evidence support and traceability. At the same time, it enhances the credibility and knowledge organization capabilities of the three-level nodes in the structured outline, thereby improving the accuracy and reliability of the report content.

[0051] The first-level structure nodes, the second-level structure nodes, the third-level structure nodes, and the corresponding explicit binding relationships are represented in a hierarchical data structure form to generate a structured outline.

[0052] Specifically, after completing the construction of first-level, second-level, and third-level structural nodes and establishing explicit binding relationships, the first-level, second-level, and third-level structural nodes and their corresponding explicit binding relationships are uniformly integrated in a structured manner, organized and expressed in a hierarchical data structure form, and a structured outline is generated.

[0053] When generating a structured outline, the first-level structural nodes are used as top-level nodes, with their subordinate second-level structural nodes as child nodes, and then third-level structural nodes as finer-grained leaf nodes. This process is recursively organized according to the tree-like hierarchical relationship of first-level, second-level, and third-level structural nodes. Explicit binding relationships established in previous steps are linked within each node, i.e., the mapping relationship between core viewpoints and original viewpoint sentences and document identifiers. This forms a complete structured knowledge network as the structured outline. Each node contains its semantic description information, and the finest-grained nodes contain the core viewpoints and their evidence index. The semantic description information includes node number, node title, node type, hierarchical path, a summary of the core viewpoints, and a list of binding evidence.

[0054] Using the generated structured outline as an intermediate representation allows it to be directly used for manual review and structural adjustment, while also serving as input for the subsequent text generation stage. This imposes clear structural and content constraints on the generation process, enabling a complete structural logic system to be formed between the research topic, sub-field division, and evidence sources. Consequently, the generated scientific report has a clear structure, with each part unfolding layer by layer around the research topic in a logically coherent manner, which helps readers better understand the report content.

[0055] The structured outline is parsed and modeled to generate a set of tasks for each section. The task set for each section is generated, assembled, and integrated according to the hierarchical order of the structured outline to generate and output the target scientific and technological report.

[0056] Furthermore, constructing a set of section generation tasks includes: parsing and splitting the structured outline to obtain multiple section generation units; defining data structure fields for each section in the multiple section generation units, the data structure fields including a unique identifier, paragraph type, title path, and core viewpoint information; and performing task modeling based on the data structure fields of the multiple section generation units to construct a set of section generation tasks.

[0057] Specifically, the generated hierarchical outline undergoes structural analysis and node decomposition. Each node in the hierarchical outline, including first-level, second-level, and third-level nodes, is mapped to an independently executable section generation unit according to the chapter hierarchy. Each section generation unit corresponds to a generateable paragraph or sub-chapter in the report. The division is based on node hierarchy attributes, node title information, and the core viewpoints and supporting evidence associated with the node. During the decomposition process, a tree traversal algorithm can be used to perform a depth-first or breadth-first traversal of the structured outline, expanding the nodes layer by layer and generating standardized task units. At the same time, nodes with logical continuity in structure are merged. Logical continuity can be determined through feature word recognition, such as nodes with sequential or summarizing words like "firstly," "secondly," "furthermore," "in addition," "finally," and "in conclusion," identifying their semantic connection attributes and maintaining chapter coherence during task decomposition to avoid excessive fragmentation.

[0058] After obtaining multiple section generation units, a unified data structure field is defined for each section in the multiple section generation units. The data structure field includes a unique identifier, paragraph type, title path, and core viewpoint information. The unique identifier is used to uniquely identify the section and support result write-back and tracking. The paragraph type is used to identify the functional position of the section in the report, including at least three types: intro, main, and conclusion. Intro corresponds to the beginning of a second-level heading, main corresponds to the main paragraph of a third-level subsection, and conclusion corresponds to the summary at the end of a second-level heading. The title path is used to provide thematic constraints for paragraph generation. The core viewpoint information is used to provide the core argument points that the paragraph needs to cover and is associated with the evidence materials through a list of supporting sentence IDs.

[0059] Based on the data structure fields, each section generation unit is converted into an executable generation task object, and a set of section generation tasks is constructed. This gives the main text generation clear input boundaries and output positioning, improves the controllability and reusability of the automatic generation of the overall content of the scientific and technological report, effectively avoids structural chaos and content deviation caused by manual intervention, and thus improves the structural consistency and execution efficiency of the automatic generation of scientific and technological reports.

[0060] Furthermore, generating and outputting the target scientific and technological report includes: retrieving the original text of the viewpoint sentences and document metadata corresponding to the multiple section generation units from the data storage server according to the explicit binding relationship, forming a paragraph-level evidence set; constructing a generation prompt word template according to the type of the multiple section generation units; using the paragraph-level evidence set as a generation constraint, calling a large language model to generate the main text of the section generation task set based on the prompt word template and the generation constraint, obtaining an initial section generation text set; assembling and integrating the initial section generation text set according to the hierarchical order of the structured outline, generating and outputting the target scientific and technological report.

[0061] Specifically, based on the explicit binding relationship, the original text of the viewpoint sentence and the literature metadata corresponding to multiple section generation units are retrieved from the data storage server. During the retrieval process, based on the supporting sentence ID and literature ID carried in the core viewpoint information, the corresponding viewpoint sentence original text, literature title, author, publication year, abstract and other metadata information are obtained from the data storage server. The retrieval results are summarized and deduplicated to form a paragraph-level evidence set corresponding to each section. The paragraph-level evidence set is equivalent to the evidence context of paragraph generation. Its role is to provide the generation model with citationable factual basis and source information, and to limit the generation range of the model at the input level, reducing the risk of cross-paragraph content interference and unfounded expansion.

[0062] Based on the types of multiple section generation units, such as introduction, main, and conclusion, differentiated generation prompt templates are constructed for each. Each prompt template includes at least a paragraph location description, title and topic constraints, a list of core viewpoints, detailed supporting evidence sentences for each viewpoint, and a set of citationable literature. For the introduction type, the prompt constraints emphasize an overview of the research background, significance, and chapter scope, avoiding overly detailed technical aspects. For the main type, the prompt constraints require the output to elaborate on each core viewpoint, integrating key information from supporting evidence sentences into the paragraph narrative and providing appropriate citations. For the conclusion type, the prompt constraints summarize the main research progress within the chapter and provide limitations and future research directions. Simultaneously, the prompts constrain the output length range, expression style (academic language), and citation insertion rules to ensure consistent writing conventions and structural functionality across different paragraph types.

[0063] During the generation phase, paragraph-level evidence sets are used as generation constraints. A large language model is invoked, and the main text is generated from the section generation task set based on prompt word templates and generation constraints. Specifically, the section generation task set is constructed as a task queue, and multiple sections are simultaneously invoked using a thread pool or equivalent parallel execution mechanism. Each worker thread independently completes the prompt construction, material assembly, and generation request for one section. The parallelism is configurable to adapt to different server resource conditions. The parallel mechanism breaks down the generation of the entire long text into multiple independent local generation tasks, thereby significantly reducing the overall generation time and improving throughput under batch document scale conditions, provided that computing resources allow. Through main text content reasoning and text generation, an initial set of section-generated text that meets the evidence support requirements is output. The generated text of each section is sorted, spliced, and formatted according to the hierarchical order determined by the structured outline. Chapter number mapping and title hierarchy alignment are automatically completed, and logical connection optimization is performed on cross-chapter content. Finally, a complete target scientific and technological report is assembled and output.

[0064] The aforementioned target scientific report is not generated end-to-end directly from the original literature collection. Instead, it is driven by a structured outline as an intermediate representation. The structured outline already contains a hierarchical structure, the core viewpoints corresponding to each node, and the evidence indexes tied to those viewpoints. Therefore, by using the structured outline as a constrained input, the problem of long-text writing is decomposed into several paragraph generation tasks with clear boundaries. During the generation process, the consistency between paragraph content and evidence sources, the standardization of citations, and the controllability of overall generation efficiency are maintained. This improves the stability of the report structure and the rationality of the topic coverage, enhances the consistency between the report and the user's research objectives, and thus improves the quality of content generation, reduces the risk of topic divergence, and avoids the structural drift caused by relying solely on templates or end-to-end generation.

[0065] Furthermore, the initial section-generated text set is assembled and integrated according to the hierarchical order of the structured outline to generate and output the target scientific and technological report. This includes: standardizing the citation format in the initial section-generated text set to obtain a standard section-generated text set; and assembling and integrating the standard section-generated text set according to the hierarchical order of the structured outline to generate and output the target scientific and technological report.

[0066] Furthermore, a standard section-generated text set is obtained, including: performing regular expression matching and structured parsing on the citation format in the initial section-generated text set, uniformly converting the identified citation tags into a preset format, and performing rule-based conversion on multiple parallel citations and compound citation strings to obtain a standard section-generated text set.

[0067] Specifically, since the large language model may output various citation formats such as (author, year), [1], and (ref), after receiving the initial section generated text set, regular expression matching and structured parsing are performed on the citation formats in the initial section generated text set. The citation format refers to the text feature symbols used to indicate the source of the literature, such as [1], (Smith, 2020), [3–5], [2,4,7], etc. These symbols usually appear after the opinion sentence or conclusion sentence to indicate the source of evidence for the argument. In the text recognition stage, regular expressions are used to perform pattern recognition and extraction on different forms of citation marks and parse them into structured fields, including document number, author information and year, etc. The recognition results are uniformly converted according to the pre-set citation standard template, such as uniformly converting them into the [document ID] format, thereby achieving consistency of citation style across literature sources. Simultaneously, multiple parallel or compound citation strings appearing in the initial section-generated text set are subjected to regularized transformation. For example, [1][2][3] is converted to [1–3], or [2,3][5] is standardized to [2,3,5]. This process is implemented through sequence parsing and interval merging algorithms to improve the standardization and readability of citation expressions. After completing the citation format standardization, a standard section-generated text set is obtained.

[0068] Following the hierarchical order of the structured outline—namely, the chapter path relationships between first-level, second-level, and third-level structural nodes—the main text in the standard section-generated text set is automatically sorted and assembled to ensure the output report maintains consistency with the structured outline in terms of chapter order. During the integration process, corresponding heading levels are automatically inserted using chapter title mapping rules, and logical connection markers such as "firstly," "furthermore," "at the same time," and "in conclusion" are used to optimize the transition between adjacent paragraphs, thereby forming a clear, logically coherent, and complete text structure. The final result is the generation of the target technical report, output in a standard document format such as Word, PDF, or a structured data file.

[0069] By standardizing and assembling the cited information in a unified manner, the generated content achieves consistency and completeness in terms of academic norms, chapter structure, and logical expression, thereby improving the standardization, readability, and traceability of evidence in automatically generated scientific and technological reports.

[0070] In summary, addressing the technical problems of existing automatic scientific report generation technologies, such as structural instability, topic divergence, difficulty in matching paragraphs with evidence, and repetition within chapters due to reliance on templates or end-to-end generation, this invention provides an automatic scientific report generation method driven by a structured outline. This method first determines the scope of the primary research topic based on search intent, then generates a secondary structure based on research questions, and finally refines the secondary structure to a tertiary level by enhancing it with opinion sentences. Since each level of structure is generated driven by interpretable semantic features from the literature collection, the report's chapter structure can adaptively form according to the topic distribution of the literature collection, thereby improving the stability of the report structure and the rationality of topic coverage, and avoiding structural drift caused by relying solely on templates or end-to-end generation.

[0071] By setting up a search intent recognition and document redistribution mechanism, the relevance score between the research topic and the description of the primary research topic category is used as the priority basis. When necessary, the original primary classification of the documents is covered, and the document collection is uniformly classified into the target primary structure node. The top-level semantic boundary of the report is defined in the structured outline construction stage, so that the subsequent secondary and tertiary divisions are carried out under the same research direction constraint, thereby improving the consistency between the report and the user's research objectives and reducing the risk of topic divergence.

[0072] By breaking down the text generation task into several section units at the structured outline level, integrating the generated text, and standardizing the citations after generation, the system can maintain high throughput and stable output even with a large document size, while ensuring consistent citation formats for easier subsequent formatting and review.

[0073] Example 2, based on the same inventive concept as the structured outline-driven automatic generation method for scientific and technological reports in the foregoing examples, such as... Figure 2 As shown, this application provides a structured outline-driven automatic technology report generation system, wherein the structured outline-driven automatic technology report generation system includes: The document input module 11 receives multiple scientific and technological documents on the same research topic, preprocesses them to obtain multiple standard scientific and technological documents, and stores them on a data storage server. The structure node generation module 12 performs semantic analysis and matching on the user-input research topic to generate first-level structure nodes. It then extracts and divides the research questions from the multiple standard scientific and technological documents according to the first-level structure nodes to construct second-level structure nodes. The viewpoint extraction module 13 extracts and refines viewpoint sentences based on the second-level structure nodes to generate third-level structure nodes. It then aggregates the viewpoint sentences from the third-level structure nodes to generate core viewpoints and bind them with evidence, establishing explicit binding relationships. The structured outline generation module 14 represents the first-level, second-level, and third-level structure nodes and their corresponding explicit binding relationships in a hierarchical data structure to generate a structured outline. The report integration and output module 15 performs task parsing and modeling on the structured outline, constructs a set of section generation tasks, generates and integrates the main text of the set of section generation tasks according to the hierarchical order of the structured outline, generates the target scientific and technological report, and outputs it.

[0074] Furthermore, the structure node generation module 12 is also used to: predefine multiple primary research topic category description texts; perform semantic analysis and matching quantification on the user-input research topic based on the multiple primary research topic category description texts, and calculate semantic relevance scores; retain primary research topics in the multiple primary research topic category description texts whose semantic relevance scores are higher than a preset threshold, and obtain research intent analysis results; and use the research intent analysis results as target research directions to generate primary structure nodes.

[0075] Furthermore, the structure node generation module 12 is also used to: uniformly classify and redistribute the multiple standard scientific and technological documents according to the first-level structure node to obtain multiple classified scientific and technological documents; extract research questions from the multiple classified scientific and technological documents belonging to the same first-level structure node to obtain research question text; combine the research question text and document metadata to construct a semantic feature vector, wherein the document metadata includes document title, abstract and keywords; and perform semantic clustering on the semantic feature vector to construct a second-level structure node.

[0076] Furthermore, the opinion extraction module 13 is also used to: extract author opinion sentences from the collection of scientific and technological documents corresponding to the secondary structure node; perform semantic clustering by using the author opinion sentences and the research question text as semantic feature inputs, and at the same time introduce the semantic information of the same level secondary structure node as mutual exclusion constraints to generate a tertiary structure node.

[0077] Furthermore, the opinion extraction module 13 is also used to: collect opinion sentences belonging to the same third-level structure node in the third-level structure node to obtain original opinion sentences; input the original opinion sentences into the large language model for induction and summarization to generate node core opinions, and establish an explicit binding relationship between the node core opinions, the original opinion sentences and the document identifier.

[0078] Furthermore, the report integration and output module 15 is also used to: parse and split the structured outline to obtain multiple section generation units; define data structure fields for each section in the multiple section generation units, the data structure fields including unique identifier, paragraph type, title path and core viewpoint information; and perform task modeling based on the data structure fields of the multiple section generation units to construct a set of section generation tasks.

[0079] Furthermore, the report integration and output module 15 is also used to: retrieve the original text of the viewpoint sentences and document metadata corresponding to the multiple section generation units from the data storage server according to the explicit binding relationship, forming a paragraph-level evidence set; construct a generation prompt word template according to the type of the multiple section generation units; use the paragraph-level evidence set as a generation constraint, call the large language model to generate the main text of the section generation task set based on the prompt word template and the generation constraint, and obtain an initial section generation text set; assemble and integrate the initial section generation text set according to the hierarchical order of the structured outline, generate the target scientific and technological report, and output it.

[0080] Furthermore, the report integration and output module 15 is also used to: standardize the citation format in the initial section generated text set to obtain a standard section generated text set; assemble and integrate the standard section generated text set according to the hierarchical order of the structured outline to generate and output the target scientific and technological report.

[0081] Furthermore, the report integration output module 15 is also used to: perform regular expression matching and structured parsing on the citation format in the initial section generated text set, uniformly convert the identified citation tags into a preset format, and perform rule-based conversion on multiple citations in parallel and compound citation strings to obtain a standard section generated text set.

[0082] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The method and specific examples of automatic generation of scientific and technological reports based on structured outlines in the foregoing embodiment one are also applicable to the automatic generation system of scientific and technological reports based on structured outlines in this embodiment. Through the foregoing detailed description of the method of automatic generation of scientific and technological reports based on structured outlines, those skilled in the art can clearly understand the automatic generation system of scientific and technological reports based on structured outlines in this embodiment. Therefore, for the sake of brevity, it will not be described in detail here.

[0083] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0084] Obviously, those skilled in the art can make several improvements and modifications to this application without departing from the principles of this application, and these improvements and modifications also fall within the protection scope of this application.

Claims

1. A method for automatically generating scientific and technological reports based on structured outlines, characterized in that, The method includes: The system receives multiple scientific and technological documents on the same research topic, preprocesses these documents to obtain multiple standard scientific and technological documents, and stores these standard scientific and technological documents in a data storage server. Semantic analysis and matching are performed on the research topic input by the user to generate a first-level structure node. The research questions of the multiple standard scientific and technological documents are extracted and divided according to the first-level structure node to construct a second-level structure node. Based on the second-level structure nodes, the viewpoint sentences are extracted and refined to generate third-level structure nodes. The viewpoint sentences of the third-level structure nodes are then collected to generate core viewpoints and bind evidence, establishing explicit binding relationships. The first-level structure nodes, the second-level structure nodes, the third-level structure nodes, and the corresponding explicit binding relationships are represented in a hierarchical data structure form to generate a structured outline; The structured outline is parsed and modeled to generate a set of tasks for each section. The task set for each section is generated, assembled, and integrated according to the hierarchical order of the structured outline to generate and output the target scientific and technological report.

2. The method for automatically generating scientific and technological reports based on structured outlines as described in claim 1, characterized in that, Generate first-level structure nodes, including: Predefined descriptive texts for multiple primary research topic categories; Based on the description text of the multiple primary research topic categories, semantic analysis and matching quantification are performed on the user-input research topic to calculate the semantic relevance score. The primary research topics whose semantic relevance scores in the description text of the multiple primary research topic categories are higher than a preset threshold are retained to obtain the research intent analysis results. The results of the research intent analysis are used as the target research direction to generate primary structure nodes.

3. The method for automatically generating scientific and technological reports based on structured outlines as described in claim 1, characterized in that, Constructing secondary structure nodes includes: According to the first-level structural nodes, the multiple standard scientific and technological documents are uniformly assigned and reassigned to obtain multiple classified scientific and technological documents. Research questions are extracted from multiple classified scientific and technological documents belonging to the same first-level structural node to obtain the research question text; By combining the research question text and document metadata, a semantic feature vector is constructed, wherein the document metadata includes document title, abstract and keywords; The semantic feature vectors are semantically clustered to construct a two-level structure node.

4. The method for automatically generating scientific and technological reports based on a structured outline as described in claim 3, characterized in that, Generate a three-level structure of nodes, including: Extract author viewpoint sentences from the collection of scientific and technological documents corresponding to the secondary structure nodes; Semantic clustering is performed by taking the author's viewpoint sentence and the research question text as semantic feature inputs, and semantic information of the second-level structural nodes at the same level is introduced as mutual exclusion constraints to generate third-level structural nodes.

5. The method for automatically generating scientific and technological reports based on structured outlines as described in claim 1, characterized in that, Establishing explicit binding relationships includes: The viewpoint sentences belonging to the same third-level structure node are collected to obtain the original viewpoint sentences; The original viewpoint sentence is input into a large language model for induction and summarization, generating the core viewpoint of the node, and establishing an explicit binding relationship between the core viewpoint of the node, the original viewpoint sentence, and the document identifier.

6. The method for automatically generating scientific and technological reports based on structured outlines as described in claim 1, characterized in that, Construct a section to generate a set of tasks, including: The structured outline is parsed and split to obtain multiple section generation units; Define a data structure field for each section in the plurality of section generation units. The data structure field includes a unique identifier, paragraph type, title path, and core viewpoint information. Task modeling is performed based on the data structure fields of the multiple section generation units to construct a set of section generation tasks.

7. The method for automatically generating scientific and technological reports based on a structured outline as described in claim 6, characterized in that, Generate and output the target technology report, including: Based on the explicit binding relationship, the original text of the viewpoint sentences and the document metadata corresponding to the multiple section generation units are retrieved from the data storage server to form a paragraph-level evidence set; Construct a prompt word template according to the types of the multiple section generation units; Using the paragraph-level evidence set as generation constraints, the large language model is invoked to generate the main text of the section generation task set based on the prompt word template and the generation constraints, thereby obtaining the initial section generation text set. The initial section text set is assembled and integrated according to the hierarchical order of the structured outline to generate the target scientific and technological report and output it.

8. The method for automatically generating scientific and technological reports based on structured outlines as described in claim 7, characterized in that, The initial section text set is assembled and integrated according to the hierarchical order of the structured outline to generate and output the target scientific and technological report, including: The reference format in the initial section generated text set is standardized to obtain a standard section generated text set; The standard section text sets are assembled and integrated according to the hierarchical order of the structured outline to generate the target scientific and technological report and output it.

9. The method for automatically generating scientific and technological reports based on structured outlines as described in claim 8, characterized in that, The standard section generates a set of text, including: Regular expression matching and structured parsing are performed on the citation format in the initial section generated text set. The identified citation tags are uniformly converted into a preset format. The cases of multiple citations in parallel and compound citation strings are converted according to rules to obtain the standard section generated text set.

10. A structured outline-driven automatic technical report generation system, characterized in that, The steps for implementing the automatic generation method of scientific and technological reports based on structured outlines as described in any one of claims 1 to 9 include: The document input module is used to receive multiple scientific and technological documents on the same research topic, preprocess the multiple scientific and technological documents to obtain multiple standard scientific and technological documents, and store the multiple standard scientific and technological documents to the data storage server. The structural node generation module is used to perform semantic analysis and matching on the research topic input by the user, generate first-level structural nodes, extract and divide the research questions of the multiple standard scientific and technological documents according to the first-level structural nodes, and construct second-level structural nodes. The viewpoint extraction module is used to extract and refine viewpoint sentences based on the secondary structure nodes, generate tertiary structure nodes, collect the viewpoint sentences of the tertiary structure nodes, generate core viewpoints and bind evidence, and establish explicit binding relationships. The structured outline generation module is used to represent the first-level structure nodes, the second-level structure nodes, the third-level structure nodes and the corresponding explicit binding relationships in a hierarchical data structure form to generate a structured outline. The report integration and output module is used to perform task parsing and modeling on the structured outline, construct a set of task generation sections, generate and assemble the main text of the task generation sections according to the hierarchical order of the structured outline, generate the target scientific and technological report, and output it.

Images