Advanced technology field report text generation method and system based on deep research

Abstract

The invention provides a advanced technology field report text generation method and system based on deep research. The method comprises the following steps: step 1, constructing a multi-level chapter outline according to a research subject; step 2, on the basis of the generated chapter outline, performing subtitle-oriented content writing; and step 3, carrying out credibility evaluation on the content generated in the step 2, and establishing semantic association between the reference and the generated content through an entity extraction and semantic matching technology so as to identify and correct possible illusion phenomena. The method has the advantages that the defects that traditional literature retrieval is ubiquitous but not deep, and key information is missed are overcome; the accuracy of authenticity verification is greatly improved, and the core problems of structure disorder, information missing, content illusion and the like in traditional report generation are systematically solved.

Description

Technical Field

[0001] This application belongs to the field of information retrieval and natural language processing technology, specifically relating to a method and system for generating report text in a cutting-edge technology field based on in-depth research. Background Technology

[0002] Reports in cutting-edge technology fields are core documents supporting technology R&D planning, achievement transformation evaluation, and industry trend analysis. Their generation efficiency and quality directly affect industry innovation and the cultivation of core competitiveness. Faced with massive amounts of multi-source scientific and technological data that need to be processed in a timely manner, traditional methods often rely on fixed template filling or direct generation. While these methods can complete basic writing tasks, they lack flexibility, are difficult to adapt to the needs of dynamically expanding in-depth research, and are prone to logical inconsistencies, failing to meet the requirements of in-depth research for content rigor and adaptability.

[0003] In recent years, numerous innovative methods have emerged in the fields of text generation and deep research. AgentWrite, by simulating human deep writing thinking, employs a segmented generation strategy to overcome the limitations of the context window, providing support for the creation of long-form deep texts. STORM proposes a "planning-execution" two-stage generation framework, effectively enhancing the logical coherence and structural integrity of texts, adapting to the structured needs of deep research. Other research embeds planning mechanisms into the model architecture to achieve efficient single-round long-form text generation, improving the efficiency of deep research. CO-STORM continuously revises and expands information through multi-round dialogues to optimize text quality, facilitating content iteration in deep research. Meanwhile, for complex deep research tasks, DeepSeek improves inference performance through multi-stage training and a small amount of cold-start data, adapting to the complex inference needs of deep research. DeepResearch employs a reinforcement learning fine-tuning strategy, supporting autonomous planning, retrieval, and verification in dynamic environments, generating structured and credible research reports, enhancing the credibility of deep research. DeepResearcher, for the first time, undergoes end-to-end reinforcement learning training in a real network environment, enabling it to autonomously cope with the complexity of open networks, suitable for deep research tasks requiring multi-source fusion and cross-validation. However, existing methods are still weak in dealing with fuzzy tasks in in-depth research, unable to accurately retrieve and deeply analyze "relevant" content, and the models are prone to "illusions," introducing false information, which seriously affects the accuracy and credibility of in-depth research reports and makes it difficult to support high-quality in-depth research in cutting-edge technology fields. Summary of the Invention

[0004] The purpose of this application is to overcome the shortcomings of existing technologies, which cannot accurately retrieve and deeply analyze "relevance" content, and whose models are prone to "illusions," introducing false information, seriously affecting the accuracy and credibility of in-depth research reports, and making it difficult to support high-quality in-depth research in cutting-edge technology fields.

[0005] To achieve the above objectives, this application proposes a method for generating report texts in cutting-edge technology fields based on in-depth research, including: Step 1: Construct a multi-level chapter outline based on the research topic; Step 2: Based on the generated chapter outline, write content oriented towards subheadings; Step 3: Conduct a credibility assessment on the content generated in Step 2. Through entity extraction and semantic matching techniques, establish semantic relationships between the references and the generated content, thereby identifying and correcting any possible hallucinations.

[0006] As an improvement to the above method, step 1 includes: Step 1-1: For a given research topic, first, based on the general chapter system of the established frontier technology field report, and combined with the research topic-related information obtained by coarse-grained network retrieval and in-depth analysis, generate a multi-level outline with hierarchical logic. Steps 1-2: Using a general large language model, the multi-level outline is broken down into independent in-depth research units by chapter, and a writing plan is generated for each chapter.

[0007] As an improvement to the above method, step 1-1 includes: Web crawling technology is used to collect raw information related to the research topic, resulting in coarse-grained web retrieval of items and the output of raw retrieval information; The semantic similarity of the search results is calculated and ranked using the BGE model.

[0008] As an improvement to the above method, step 2 includes: Step 2-1: Based on the chapter titles of the multi-level outline, use a large speech model to generate search keywords that match the semantics of the titles, and clarify the scope of literature search topics suitable for in-depth research. Step 2-2: Obtain the original reference set corresponding to the chapter title through multi-source deep network retrieval, filter high-confidence literature based on the pre-set confidence threshold, and store it in the cache library; at the same time, for the in-depth research knowledge gaps that still exist after the retrieval, the large language model supplements fine-grained keywords based on the prompting engineering to trigger a second deep retrieval. Steps 2-3 involve performing multiple rounds of deep semantic analysis and iterative reasoning on high-confidence documents in the cache library to uncover core information in the documents that meets the needs of in-depth research and to determine whether the documents adequately support the in-depth answer requirements of the chapter content. If the documents do not adequately support the content, a new round of retrieval is triggered. If the documents adequately support the content, a draft report is generated based on the refined core information of the documents according to the chapter writing plan.

[0009] As an improvement to the above method, steps 2-3 include: A hybrid retrieval mechanism is used to extract relevant literature evidence from the cache, the BM25 algorithm is used for keyword matching, and then Sentence-BERT is used to calculate semantic similarity. The large language model is then used to determine whether it is sufficient to support the answer to the question. If the literature support is sufficient, paragraph content is generated directly based on the large language model, combined with task-specific instruction templates and few-sample prompts. If the literature support is insufficient, a new round of literature retrieval cycle is triggered.

[0010] As an improvement to the above method, step 3 includes: Step 3-1: Using entity extraction technology, extract the reference entity set from high-confidence documents and the generated content entity set from the initial report draft to clarify the core objects for in-depth fact verification. Step 3-2: Use the BGE model to map each entity in the two entity sets to a vector representation. The difference in entity surface representation is eliminated through vector space transformation to ensure the semantic matching accuracy for in-depth research. Step 3-3: Based on vector representation, the cosine similarity algorithm is used to calculate the similarity between the generated content entities and the reference entities, and a preset threshold for adapting to in-depth research is set: if there is a reference entity whose similarity is greater than the threshold, the generated content entity is determined to have literature support; if all entities in the generated content entity set are successfully matched, the initial draft is determined to be free of illusions; if there are unmatched questionable entities, each sentence in the paragraph containing the questionable entity is subjected to secondary verification in combination with the background of in-depth research. If it is confirmed to be an illusion, a corrected statement based on literature is generated and the original erroneous content is replaced; if it is verified to be correct, the original content is retained, and finally, a true and credible report text that meets the requirements of in-depth research on cutting-edge technologies is output.

[0011] As an improvement to the above method, in step 3-3, if there are any unmatched suspicious entities, then a secondary verification is performed on each sentence in the paragraph containing the suspicious entity, taking into account the in-depth research background. If it is confirmed to be a hallucination, then a corrected statement based on literature evidence is generated and the original erroneous content is replaced. Specifically: For questionable entities that fail to match, each sentence in the paragraph is split into sentences and analyzed in depth by a large language model. The large language model determines whether the generated content can be inferred from the references, which serves as the basis for whether it constitutes hallucination content. If a sentence is confirmed to be a hallucination, the large language model is called to rewrite the sentence and replace the original content.

[0012] This application also provides a report text generation system for cutting-edge technology fields based on in-depth research, implemented using the above method, the system comprising: The framework planning module is used to construct multi-level chapter outlines based on the research topic. The multi-source deep mining module is used to write content based on subheadings, building upon the generated chapter outline. The fact-accurate verification module is used to evaluate the credibility of the content generated by the multi-source deep mining module. Through entity extraction and semantic matching technology, it establishes semantic associations between references and generated content, thereby identifying and correcting possible hallucination phenomena.

[0013] Compared with existing technologies, the advantages of this application are: 1. This invention breaks through the limitations of traditional generation models and constructs a deep research mechanism of "in-depth research-oriented framework planning - multi-source in-depth mining - accurate fact verification": In the framework building stage, it innovatively integrates the semantic ranking of the BGE model with the in-depth analysis of cutting-edge technology thematic units to generate multi-level outlines, accurately matching the research dimensions of the topic and solving the problem of insufficient adaptability between traditional outlines and the needs of in-depth research; In the literature processing stage, it establishes a literature processing mechanism based on "multi-dimensional confidence screening + multi-round iterative retrieval", which not only enhances the in-depth research value of the included literature, but also ensures the integrity of information coverage, avoiding the defects of "broad but not deep" and "missing key information" in traditional literature retrieval; In the authenticity verification stage, it proposes an illusion correction method of "vector mapping + secondary in-depth verification" that combines the characteristics of cutting-edge technology entities. It eliminates expression differences through accurate vector encoding of entities such as technical terms, parameters, and conclusions, and then implements secondary verification by tracing the original text and cross-validating related data in combination with the background of in-depth research, which greatly improves the accuracy of authenticity verification and systematically solves the core problems of structural chaos, missing information, and content illusion in traditional report generation.

[0014] 2. This invention designs a key technical path that integrates "semantic analysis + iterative optimization + precise verification" to ensure a balance between generation efficiency and quality: In the planning stage, the semantic ranking and topic unit analysis of the BGE model are used to ensure that the outline conforms to the domain specifications and is adapted to the specific topic, without relying on a fixed template; in the literature processing stage, confidence screening and multi-round iterative retrieval are used to ensure high-quality information and no coverage gaps; in the illusion correction stage, entity vector mapping and secondary verification are used to accurately remove fictitious information and avoid misjudgment due to differences in expression; the overall technical path reduces the cost of manual intervention and promotes the upgrade of report generation from "one-time output" to "dynamic optimization", which improves the quality of generation and accelerates the efficiency of business processes. Attached Figure Description

[0015] Figure 1 The diagram shows a flowchart of a method for generating report text in cutting-edge technology fields based on in-depth research. Figure 2 The diagram shown is a schematic of the framework planning module; Figure 3 The diagram shown is a schematic of the multi-source deep mining module; Figure 4 The diagram shown is a schematic of the fact-accurate verification module. Detailed Implementation

[0016] The technical solution of this application will be described in detail below with reference to the accompanying drawings.

[0017] The purpose of this invention is to address the key bottlenecks in report generation tasks using large language models in cutting-edge technology fields, especially in in-depth research scenarios, and to propose a systematic and deeply optimized intelligent generation method. Existing technologies generally suffer from insufficient understanding of fuzzy or open-ended in-depth research tasks, difficulty in accurately identifying and deeply analyzing core intelligence information highly "relevant" to the task, and a tendency to generate factual "illusions" and introduce false or fabricated content during the generation process. This results in reports with loose structures, redundant or missing key information, and insufficient credibility, failing to meet the high requirements for rigor, accuracy, and credibility in in-depth research scenarios such as scientific research assessment, strategic planning, or technical intelligence.

[0018] Example 1 This invention proposes a method for generating report text in cutting-edge technology fields based on in-depth research. By constructing a comprehensive optimization mechanism encompassing "in-depth research-oriented framework planning—multi-source in-depth mining—accurate fact verification," it achieves a closed-loop improvement from understanding in-depth research tasks, acquiring high-value knowledge, generating in-depth content, to verifying authenticity. This method not only deepens the semantic analysis and target focusing capabilities for fuzzy research tasks, accurately retrieves, filters, and integrates highly relevant intelligence clues suitable for in-depth research needs, but also deepens content depth through multi-round iterative reasoning and strictly controls information authenticity through entity-level fact verification mechanisms, significantly suppressing model illusions and ensuring the logical rigor and factual reliability of report content. Ultimately, this invention aims to comprehensively improve the efficiency of clue screening, the quality of classification and compilation, and the credibility of report generation in in-depth research scenarios, providing efficient, reliable, and traceable intelligent text generation support for decision support and knowledge innovation in cutting-edge technology fields. This invention systematically solves the technical pain points of traditional report generation, such as structural chaos, missing information, and fabricated (illusionary) content, and is particularly suitable for the high requirements of report rigor, information depth, and credibility in in-depth research scenarios. The specific technical solution is as follows: Step 1) Taking the needs of in-depth research topics in cutting-edge technology fields as the core input, the report generation is provided with a unified and rigorous structural constraint that adapts to the logic of in-depth research through a two-step method of "multi-level outline generation - chapter decomposition and writing plan formulation". Specifically, it includes: Step 1-1) For a given in-depth research topic on cutting-edge technology, first, based on the general chapter system of reports in the field of cutting-edge technology, and combined with the relevant information on the topic (including technology context, research hotspots, core data, etc.) obtained by coarse-grained network retrieval and in-depth analysis, generate a multi-level outline with hierarchical logic, and clarify the overall logical architecture of the report to meet the needs of in-depth research. Steps 1-2) Break down the above multi-level outline into independent in-depth research units by chapter, and generate a detailed writing plan for each chapter that includes "core research points, information needs, and logical connection requirements" to provide precise guidance for the subsequent generation of in-depth content.

[0019] Step 2) Guided by the multi-level outline chapter titles output in Step 1), adopt the deep research strategy of "multi-source deep retrieval - semantic discrimination - iterative optimization" to achieve comprehensive acquisition of information required for in-depth research and in-depth content refinement; Specifically, it includes: Step 2-1) Based on the chapter titles of the multi-level outline, generate search keywords that are highly matched with the semantics of the titles in combination with the in-depth research objectives, and clarify the scope of literature search topics suitable for in-depth research; Step 2-2) Obtain the original reference set corresponding to the chapter title through multi-source deep network retrieval, filter high-confidence literature based on the pre-set confidence threshold, and store it in a dedicated cache library; at the same time, for the knowledge gaps in in-depth research that still exist after the retrieval (such as the lack of comparison of certain technical parameters or cross-domain application cases), the large language model supplements fine-grained keywords (such as "XX technology 2023-2025 measured data") based on the prompting engineering and triggers a second deep retrieval to ensure that the information coverage meets the integrity requirements of in-depth research; Steps 2-3) Perform multiple rounds of deep semantic analysis and iterative reasoning on high-confidence documents in the cache library to accurately mine the core information in the documents that is suitable for in-depth research needs and determine whether the documents fully support the in-depth answering needs of the chapter content: if there are insufficient documents, a new round of retrieval cycle is triggered; if there are sufficient documents, based on the refined core information of the documents, a logically coherent, data-rich, and in-depth analysis draft report is generated according to the chapter writing plan.

[0020] Step 3) Based on the initial draft of the report output in Step 2) and the high-confidence documents in the cache library, the fictitious information is accurately identified and eliminated through the deep fact verification path of "entity extraction - vector mapping - similarity matching - secondary verification" to ensure the authenticity of the in-depth research content.

[0021] Specifically, it includes: Step 3-1) Using entity extraction technology, extract the reference entity set from high-confidence documents and the generated content entity set from the initial report draft to clarify the core objects for in-depth fact verification; Step 3-2) Use an encoding function to map each entity in the two entity sets into a vector representation. Eliminate the differences in entity surface representation through vector space transformation to ensure the semantic matching accuracy for in-depth research. Step 3-3) Calculate the similarity between the generated content entities and the reference entities using the cosine similarity algorithm based on vector representation, and set a preset threshold suitable for in-depth research: if there is a reference entity whose similarity is greater than the threshold, the generated content entity is determined to have literature support; if all entities in the generated content entity set are successfully matched, the initial draft is determined to be free of illusions; if there are unmatched questionable entities (such as technical conclusions or data values ​​for which no supporting evidence was found), then each sentence in the paragraph containing the questionable entity is subjected to secondary verification in combination with the background of in-depth research. If it is confirmed to be an illusion, a corrected statement based on literature evidence is generated and the original erroneous content is replaced. If it is verified to be correct, the original content is retained, and finally, a true and credible report text that meets the requirements of in-depth research on cutting-edge technologies is output.

[0022] like Figure 1As shown, the overall implementation process of this invention revolves around a closed-loop system built around "in-depth research-oriented framework planning—multi-source in-depth mining—accurate fact verification." Through the connection and optimization of these three core modules, it systematically solves the problems of structural chaos, missing key information, and content illusion in traditional report generation, meeting the high requirements of in-depth research in cutting-edge technology fields for report rigor, information depth, and credibility. Specifically, the framework planning module is responsible for building a report structure and writing guidelines adapted to the logic of in-depth research, providing unified constraints for subsequent content generation; the multi-source in-depth mining module focuses on the acquisition and in-depth analysis of high-value information, ensuring sufficient content support through multiple rounds of retrieval and reasoning; and the accurate fact verification module conducts entity-level verification and illusion correction on the generated content to ensure the report's authenticity. The overall process specifically includes the following steps: Step 1) Construct a multi-level outline for the research report based on the research topic, specifically including the following steps: When there is a research question... During input, a two-step method of "multi-level outline generation - chapter decomposition and writing plan formulation" is used to provide a unified and rigorous structural constraint for report generation.

[0023] like Figure 2 As shown, based on the general chapter system of reports in cutting-edge technology fields, and combined with topic-related information obtained through coarse-grained web retrieval, a multi-level outline with a hierarchical structure is generated to clarify the overall logical architecture of the report. Here, mature web crawling technology is used to comprehensively collect raw information related to the topic, ultimately obtaining the output of coarse-grained web retrieval for each item and its raw retrieval information.

[0024] Then, the BGE model (BAAI General Embedding) is used to calculate and rank the semantic similarity of the search results. The BGE model performs excellently in semantic representation and similarity calculation; it can transform each retrieved text into a semantic vector and accurately measure the semantic relevance of the text by calculating the cosine similarity between vectors. The search results are then ranked according to the calculated similarity values ​​to obtain an ordered dataset. Subsequently, we analyzed the data concentration and... We will gather relevant information, organize the thematic units, and further develop an outline. ,Right now:

[0025] in, This refers to the output of coarse-grained web retrieval of entries and their raw retrieval information; This represents the ordered dataset obtained after sorting using the BGE model. The output; Indicates based on the question and dataset Generate outline The function is actually a coarse-grained outline generation module based on a large language model (LLM). Its core mechanism is to use the prompting capability of LLM to automatically generate a structured preliminary outline O based on the input question q and the context dataset D.

[0026] Subsequently, the framework planning module invokes a general-purpose language model to perform structured parsing of the generated hierarchical outline, breaking it down into several independent chapters. Then, it further refines the core content of each chapter, developing corresponding writing plans. Finally, the chapter divisions and corresponding writing plans are integrated to form a clearly structured, task-specific outline with a plan, providing a systematic collaborative guide for subsequent writing.

[0027] Step 2) The multi-source deep mining module, based on the chapter outline generated in Step 1), conducts subheading-oriented deep mining. The content writing task is as follows. This process centers on structured problem decomposition, combined with literature retrieval, caching optimization, and iterative reasoning mechanisms, to achieve a closed-loop process from knowledge acquisition to content generation, ensuring that the output content has sufficient factual support and logical rigor.

[0028] like Figure 3 As shown, firstly, regarding the input subheading... The large language model, based on its own knowledge and experience, deeply analyzes the potential n related issues within the problem itself. By using web search technology, we obtained the original reference set for these n questions. Next, duplicate documents obtained from multiple questions are deduplicated, and documents with high confidence levels are selected for inclusion in the cache repository. The cache repository construction process can be formalized as follows:

[0029] in, This represents the merged and deduplicated results of the original documents in the multi-subject problem. This is a preset confidence threshold for documents, used to measure their relevance, authority, and timeliness. Only when a document's overall confidence score reaches or exceeds this threshold will it be considered a confidence level. Only then is it included in the cache library. This mechanism effectively filters out low-quality or irrelevant literature, ensuring that cached content has high information density and academic credibility, thereby supporting the generation of subsequent high-quality content.

[0030] After the cache library is built, this module performs multiple rounds of in-depth thinking on the selected documents, using an iterative reasoning framework based on a large language model to simulate the cognitive process of human experts "reading-understanding-reflection-integration". Each round of reasoning includes the following three core steps: 1. Semantic parsing: Utilizing a large language model to generate summaries, identify key entities, and extract events from document content, thereby extracting structured information; 2. Topic Summarization: Identify co-occurring concepts and potential logical relationships through cluster analysis or hierarchical topic modeling; 3. Logical deduction: Combining chain-thinking prompts, conduct cross-document causal reasoning and contradiction detection.

[0031] Unlike traditional one-time text generation models, this framework employs an iterative reasoning framework. Through layer-by-layer analysis of document content, thematic summarization, and logical deduction, it gradually constructs an understanding of the problem. A comprehensive cognitive model. This process simulates the "reading-understanding-reflection-integration" thinking path of human experts, achieving a leap from shallow information extraction to deep knowledge reasoning.

[0032] Based on this, a hybrid retrieval mechanism is used to retrieve data from the cache. Relevant literature evidence was extracted. The specific method was as follows: first, the BM25 algorithm was used for keyword matching; then, Sentence-BERT was combined to calculate semantic similarity, improving recall accuracy and relevance; finally, a large language model was used to determine whether it was sufficient to support the answer to the question. The solution involves several steps. If sufficient literature support is available, the system directly generates well-structured and logically clear paragraphs based on the large language model, combined with task-specific instruction templates and few-sample prompts. If insufficient literature support is available, a new round of literature retrieval is triggered. In this new round of retrieval, the system uses the semantic analysis results from the previous round as contextual clues to assist in optimizing the retrieval strategy, achieving a dynamic optimization process of "retrieval—analysis—feedback—re-retrieval".

[0033] Finally, the multi-source deep mining module generates a preliminary report based on sufficient literature support and multiple rounds of reasoning results. Its formal expression is:

[0034] in This indicates that the multi-source deep mining module is based on the problem. and cache library cache library Generate the first draft.

[0035] Step 3) as Figure 4 As shown, the core mechanism of the credibility assessment of the content generated in step 2 in this stage lies in establishing a semantic relationship between the references and the generated content through entity extraction and semantic matching techniques, thereby identifying and correcting potential hallucinations. Specifically, the fact-accurate verification module first uses entity extraction technology to extract information from the references... With generated content Extract key entities and construct the corresponding reference entity set. With generated content entity set Subsequently, the BGE model was used for vectorization encoding. Each entity is mapped to a vector representation, that is:

[0036] This eliminates discrepancies in surface representation and ensures the accuracy of semantic matching. Based on the vectorized results, the agent calculates entities in the generated content. With reference entities The similarity between them, and the matching degree is calculated by the following formula:

[0037] If it exists Make (in If a preset threshold is set, the generated content entity is considered to have supporting references. If all entities in the generated content entity set have a match, the generated content is determined not to be a hallucination; otherwise, a preliminary determination is made that a hallucination exists. For questionable entities that fail to match, this module processes the paragraphs... Each sentence in Sentence-by-sentence breakdown and deep analysis using a large language model are performed. The fact-accurate verification module uses the large language model to determine whether the generated content can be inferred from the references, serving as the basis for whether it constitutes hallucination content. If a sentence is confirmed to be hallucination, the large language model is invoked to rewrite the sentence, generating a semantically consistent and fact-accurate corrected statement, replacing the original sentence content. This process can be formally represented as follows: If the large language model determines that the reasoning result is correct, the original generated content is retained and reassembled into coherent paragraphs, ultimately resulting in a verified final draft of the chapter. Its formal expression is:

[0038] This application proposes a method for generating report texts in cutting-edge technology fields based on in-depth research. It systematically improves the quality of generated reports in these fields by constructing a structured framework, conducting multi-dimensional in-depth data mining, and implementing credible verification and tracing. This method integrates three core in-depth research stages: First, it builds a clearly structured and logically rigorous report framework based on the standards and task requirements of the cutting-edge technology field, providing structured guidance for in-depth research. Second, it conducts multi-dimensional information mining and content refinement around the framework nodes to fill information gaps and deepen the content, enhancing the richness of the in-depth research. Third, it combines credible data sources to conduct full-process verification and tracing of the generated content, accurately identifying and eliminating fictitious information to ensure the accuracy and credibility of the in-depth research report.

[0039] This application proposes a method for generating report texts in cutting-edge technology fields based on in-depth research. It replaces the traditional generation mode by constructing an in-depth research mechanism of "in-depth research-oriented framework planning - multi-source in-depth mining - accurate fact verification". It combines a multi-level outline generation method that integrates BGE model semantic ranking and topic unit analysis to enhance topic adaptability, a literature processing mechanism based on confidence screening and multi-round iterative retrieval to improve literature quality and coverage completeness, and an illusion correction method based on entity vector mapping and secondary verification to improve the accuracy of authenticity verification. It comprehensively solves the problems of structural chaos, missing information, and content illusion in traditional report generation.

[0040] Furthermore, this application restructures the entire generation process into a traceable, interventionist, and iteratively optimized dynamic optimization system. Each operation retains structured logs and intermediate products (such as problem decomposition results, search literature lists, entity matching records, etc.), making the generation process highly interpretable. This design not only improves the controllability and credibility of the generated results but also provides a technical foundation for compliance review, version backtracking, and knowledge tracing of complex technical reports. Overall, this method systematically solves key problems commonly found in traditional technical report generation, such as loose structure, missing information, and factual illusions. While improving the report's structure, content completeness, and credibility, it significantly reduces the cost of manual intervention and effectively accelerates the efficiency of knowledge output and decision-making in scientific research and business scenarios.

[0041] Example 2 This application also provides a report text generation system for cutting-edge technology fields based on in-depth research, implemented using the above method, the system comprising: The framework planning module is used to construct multi-level chapter outlines based on the research topic. The multi-source deep mining module is used to write content based on subheadings, building upon the generated chapter outline. The fact-accurate verification module is used to evaluate the credibility of the content generated by the multi-source deep mining module. Through entity extraction and semantic matching technology, it establishes semantic associations between references and generated content, thereby identifying and correcting possible hallucination phenomena.

[0042] This application may also provide a computer device, including: at least one processor, memory, at least one network interface, and a user interface. The various components in this device are coupled together via a bus system. It is understood that the bus system is used to implement communication between these components. In addition to a data bus, the bus system also includes a power bus, a control bus, and a status signal bus.

[0043] The user interface can include a display, keyboard, or clicking device. Examples include a mouse, trackball, touchpad, or touchscreen.

[0044] It is understood that the memory in the embodiments disclosed in this application may be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. The non-volatile memory may be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory may be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as Static Random Access Memory (SRAM), Dynamic Random Access Memory (DRAM), Synchronous DRAM (SDRAM), Double Data Rate SDRAM (DDRSDRAM), Enhanced Synchronous DRAM (ESDRAM), Synchlink DRAM (SLDRAM), and Direct Rambus RAM (DRRAM). The memories described herein are intended to include, but are not limited to, these and any other suitable types of memory.

[0045] In some implementations, the memory stores elements such as executable modules or data structures, or subsets thereof, or extended sets thereof: operating systems and applications.

[0046] The operating system includes various system programs, such as the framework layer, core library layer, and driver layer, used to implement various basic business functions and handle hardware-based tasks. The application programs include various applications, such as media players and browsers, used to implement various application functions. Programs implementing the methods of the embodiments of this disclosure can be included in the application programs.

[0047] In the above embodiments, the processor can also invoke programs or instructions stored in memory, specifically programs or instructions stored in an application program, for the following purposes: Follow the steps described above.

[0048] The above methods can be applied to or implemented by a processor. The processor may be an integrated circuit chip with signal processing capabilities. During implementation, each step of the above methods can be completed by integrated logic circuits in the processor's hardware or by software instructions. The processor can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic diagrams disclosed above. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the disclosed methods can be directly implemented by a hardware decoding processor, or by a combination of hardware and software modules in the decoding processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory; the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above methods.

[0049] It is understood that the embodiments described in this application can be implemented using hardware, software, firmware, middleware, microcode, or a combination thereof. For hardware implementation, the processing unit can be implemented in one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), general-purpose processors, controllers, microcontrollers, microprocessors, other electronic units for performing the functions described in this application, or combinations thereof.

[0050] For software implementation, the technology of this application can be implemented by executing the functional modules (e.g., procedures, functions, etc.) of this application. The software code can be stored in memory and executed by a processor. The memory can be implemented in the processor or outside the processor.

[0051] This application may also provide a non-volatile storage medium for storing a computer program. When the computer program is executed by a processor, it can implement the steps in the above method embodiments.

[0052] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application and are not intended to limit it. Although this application has been described in detail with reference to the embodiments, those skilled in the art should understand that modifications or equivalent substitutions to the technical solutions of this application do not depart from the spirit and scope of the technical solutions of this application, and should all be covered within the scope of the claims of this application.

Claims

1. A method for generating report text in cutting-edge technology fields based on in-depth research, comprising: Step 1: Construct a multi-level chapter outline based on the research topic; Step 2: Based on the generated chapter outline, write content oriented towards subheadings; Step 3: Conduct a credibility assessment on the content generated in Step 2. Through entity extraction and semantic matching techniques, establish semantic relationships between the references and the generated content, thereby identifying and correcting any possible hallucinations.

2. The method for generating report text in cutting-edge technology fields based on in-depth research according to claim 1, characterized in that, Step 1 includes: Step 1-1: For a given research topic, first, based on the general chapter system of the established frontier technology field report, and combined with the research topic-related information obtained by coarse-grained network retrieval and in-depth analysis, generate a multi-level outline with hierarchical logic. Steps 1-2: Using a general large language model, the multi-level outline is broken down into independent in-depth research units by chapter, and a writing plan is generated for each chapter.

3. The method for generating report text in cutting-edge technology fields based on in-depth research according to claim 2, characterized in that, Step 1-1 includes: Web crawling technology is used to collect raw information related to the research topic, resulting in coarse-grained web retrieval of items and the output of raw retrieval information; The semantic similarity of the search results is calculated and ranked using the BGE model.

4. The method for generating report text in cutting-edge technology fields based on in-depth research according to claim 1, characterized in that, Step 2 includes: Step 2-1: Based on the chapter titles of the multi-level outline, use a large speech model to generate search keywords that match the semantics of the titles, and clarify the scope of literature search topics suitable for in-depth research. Step 2-2: Obtain the original reference set corresponding to the chapter title through multi-source deep network retrieval, filter high-confidence literature based on the pre-set confidence threshold, and store it in the cache library; at the same time, for the in-depth research knowledge gaps that still exist after the retrieval, the large language model supplements fine-grained keywords based on the prompting engineering to trigger a second deep retrieval. Steps 2-3 involve performing multiple rounds of deep semantic analysis and iterative reasoning on high-confidence documents in the cache library to uncover core information in the documents that meets the needs of in-depth research and to determine whether the documents adequately support the in-depth answer requirements of the chapter content. If the documents do not adequately support the content, a new round of retrieval is triggered. If the documents adequately support the content, a draft report is generated based on the refined core information of the documents according to the chapter writing plan.

5. The method for generating report text in cutting-edge technology fields based on in-depth research according to claim 4, characterized in that, Steps 2-3 include: A hybrid retrieval mechanism is used to extract relevant literature evidence from the cache, the BM25 algorithm is used for keyword matching, and then Sentence-BERT is used to calculate semantic similarity. The large language model is then used to determine whether it is sufficient to support the answer to the question. If the literature support is sufficient, paragraph content is generated directly based on the large language model, combined with task-specific instruction templates and few-sample prompts. If the literature support is insufficient, a new round of literature retrieval cycle is triggered.

6. The method for generating report text in cutting-edge technology fields based on in-depth research according to claim 1, characterized in that, Step 3 includes: Step 3-1: Using entity extraction technology, extract the reference entity set from high-confidence documents and the generated content entity set from the initial report draft to clarify the core objects for in-depth fact verification. Step 3-2: Use the BGE model to map each entity in the two entity sets to a vector representation. The difference in entity surface representation is eliminated through vector space transformation to ensure the semantic matching accuracy for in-depth research. Step 3-3: Based on vector representation, the cosine similarity algorithm is used to calculate the similarity between the generated content entities and the reference entities, and a preset threshold for adapting to in-depth research is set: if there is a reference entity whose similarity is greater than the threshold, the generated content entity is determined to have literature support; if all entities in the generated content entity set are successfully matched, the initial draft is determined to be free of illusions; if there are unmatched questionable entities, each sentence in the paragraph containing the questionable entity is subjected to secondary verification in combination with the background of in-depth research. If it is confirmed to be an illusion, a corrected statement based on literature is generated and the original erroneous content is replaced; if it is verified to be correct, the original content is retained, and finally, a true and credible report text that meets the requirements of in-depth research on cutting-edge technologies is output.

7. The method for generating report text in cutting-edge technology fields based on in-depth research according to claim 6, characterized in that, In step 3-3, if there are any unmatched suspicious entities, a secondary verification is performed on each sentence in the paragraph containing the suspicious entity, taking into account the in-depth research background. If it is confirmed to be a hallucination, a corrected statement based on literature evidence is generated and the original erroneous content is replaced. Specifically: For questionable entities that fail to match, each sentence in the paragraph is split into sentences and analyzed in depth by a large language model. The large language model determines whether the generated content can be inferred from the references, which serves as the basis for whether it constitutes hallucination content. If a sentence is confirmed to be a hallucination, the large language model is called to rewrite the sentence and replace the original content.

8. A report text generation system for cutting-edge technology fields based on in-depth research, implemented based on the method described in any one of claims 1-7, characterized in that, The system includes: The framework planning module is used to construct multi-level chapter outlines based on the research topic. The multi-source deep mining module is used to write content based on subheadings, building upon the generated chapter outline; and The fact-accurate verification module is used to evaluate the credibility of the content generated by the multi-source deep mining module. Through entity extraction and semantic matching technology, it establishes semantic associations between references and generated content, thereby identifying and correcting possible hallucination phenomena.

Images