Paper generation method and device of big language model based on RAG

Abstract

The invention discloses a paper generation method and device of a big language model based on RAG, and relates to the technical field of artificial intelligence big language models.The method comprises the steps that retrieval is conducted in a knowledge base according to a target research field, and first reference content associated with the target research field is obtained; performing text generation processing on the first reference content to generate a plurality of initial research directions; and screening out a target research direction from the initial research directions, and performing retrieval in the knowledge base based on the target research direction to obtain a second reference content associated with the target research direction. And performing text generation processing on the second reference content to obtain an initial thesis outline. By combining the retrieval capability and the generation capability of the RAG technology, additional context information in a knowledge base is provided for a large language model. The content generated in combination with the context information is more matched with the requirements of the user, and the technical problem that the pertinence is not strong due to the fact that the correlation degree between the written papers and the requirements of the user is low in the related technology is solved.

Description

Technical Field

[0001] The present application relates to the technical field of artificial intelligence large language models, and in particular to a method and device for generating papers based on a large language model of RAG. Background Art

[0002] In modern scientific research, writing high-quality academic papers is a complex and time-consuming task. Researchers need to retrieve relevant information from a large amount of literature, generate innovative research directions, build a reasonable outline, and write a logically rigorous paper body. The traditional paper writing process relies on the personal writing experience of researchers, which makes it difficult to guarantee efficiency and quality. In addition, with the rapid development of science and technology, the number of literature in the academic field has increased dramatically, and how to efficiently retrieve and manage these literatures has become a major challenge.

[0003] In recent years, artificial intelligence technology, especially the application of large language models, has provided new solutions for academic writing. Large language models can process natural language tasks and generate natural and fluent texts. However, existing solutions often lack effective references to users' personal knowledge bases and personalized writing assistance, making it difficult to meet the specific needs of different users.

[0004] In summary, when it comes to writing academic papers using large language models, there are currently problems such as the generated academic papers have a low correlation with the researcher's personal knowledge base and the paper content is not targeted. Summary of the invention

[0005] The embodiments of the present application provide a method and device for generating a paper based on a large language model of RAG to solve the technical problem in the prior art of lacking the ability to generate professional academic papers by combining a large model with a personal knowledge base.

[0006] To achieve the above objectives, the embodiments of the present application adopt the following technical solutions: In a first aspect, an embodiment of the present application provides a method for generating a paper based on a large language model of RAG, the method comprising: In response to the target research field input by the user, searching a preset knowledge base to obtain first reference content associated with the target research field; Performing text generation processing of a large language model on the first reference content to generate multiple initial research directions; Deduplication processing is performed on the initial research directions, and based on preset evaluation requirements, the deduplication-free initial research directions are scored to obtain target scoring results; Sorting the initial research directions according to the target scoring results from high to low, and displaying the initial research directions based on the sorting results; In response to the user selecting the target research direction with the highest score from the multiple initial research directions, the knowledge base is searched to obtain second reference content associated with the target research direction, and the second reference content is processed by a large language model for text generation to generate an initial paper outline.

[0007] In combination with the first aspect, in a possible design, the initial research direction after deduplication is scored based on preset evaluation requirements to obtain a target scoring result, including: Based on the innovation, relevance, and feasibility of the preset initial research directions, the initial research directions after deduplication are scored respectively to obtain innovation scores, relevance scores, and feasibility scores; The innovation score, the relevance score, and the feasibility score are combined to obtain a target scoring result corresponding to the initial research direction.

[0008] In combination with the first aspect, in a possible design method, the innovation score, the relevance score, and the feasibility score are combined to obtain a target scoring result corresponding to the initial research direction. This includes: The sum of the innovation score, the relevance score, and the feasibility score is used as a first scoring result corresponding to the initial research direction; Scoring the initial research direction in combination with expert knowledge to obtain a second scoring result; The initial research directions whose average values ​​of the first scoring results and the second scoring results are lower than a preset threshold are eliminated to obtain the target scoring results corresponding to the screened initial research directions.

[0009] In combination with the first aspect, in a possible design, the step of performing text generation processing of a large language model on the second reference content to generate an initial paper outline includes: Performing text generation processing of a large language model on the second reference content to generate an outline structure text; Extracting third reference contents associated with each chapter in the outline structure text from the knowledge base; The initial paper outline is obtained by combining the outline structure text and the third reference content.

[0010] In combination with the first aspect, in a possible design, after generating the initial paper outline, the method further includes: receiving a user's modification operation on the initial paper outline to obtain a target paper outline; Based on the target paper outline and the second reference content, an initial academic paper corresponding to the target paper outline is generated using text generation technology of a large language model.

[0011] In combination with the first aspect, in a possible design, after generating an initial academic paper corresponding to the target paper outline, the method further includes: Polishing the initial academic paper to obtain a polished target academic paper; Based on the corresponding scoring strategies of each evaluation type of the preset target academic paper, the scores of the evaluation types are determined respectively, and the evaluation types include innovation, relevance, and feasibility; Combined with the scores obtained for each of the evaluation types for the target academic paper, a third scoring result corresponding to the target academic paper is generated.

[0012] In combination with the first aspect, the method further includes: receiving multiple documents uploaded by a user, classifying and labeling the multiple documents, and obtaining a knowledge base constructed with the multiple documents.

[0013] In a second aspect, the present application embodiment provides a paper generation device based on a large language model of RAG, including: A retrieval module, configured to search a preset knowledge base in response to a target research field input by a user, and obtain a first reference content associated with the target research field; A research direction generation module, used for performing text generation processing of a large language model on the first reference content to generate multiple initial research directions; An evaluation module, used to perform deduplication processing on the initial research directions, and score the deduplication-free initial research directions based on preset evaluation requirements to obtain a target scoring result; A sorting module, used to sort the initial research directions in descending order of the target scoring results, and display the initial research directions based on the sorting results; A writing module is used to search the knowledge base in response to a target research direction with the highest score selected by the user from multiple initial research directions, obtain second reference content associated with the target research direction, and perform text generation processing of a large language model on the second reference content to generate an initial paper outline.

[0014] In a third aspect, an embodiment of the present application provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to execute the method of the first aspect and possible design methods thereof.

[0015] In a fourth aspect, an embodiment of the present application provides a storage medium, in which a computer program is stored, wherein the computer program is configured to execute the method of the first aspect and its possible design manner when running.

[0016] Compared with the prior art, the embodiment of the present application provides a method and device for generating a paper based on a large language model of RAG. First, a search is performed in the knowledge base according to the target research field to obtain the first reference content associated therewith. The first reference content is processed by text generation to generate multiple initial research directions. The target research direction is screened out from the initial research directions, and the target research direction is searched in the knowledge base based on the target research direction to obtain the second reference content associated therewith. The second reference content is processed by text generation to obtain an initial paper outline. The present application combines the retrieval and generation capabilities of RAG (Retrieval-Augmented Generation) technology to provide the large language model with additional contextual information in the knowledge base. The content generated in combination with the contextual information is more in line with the needs of the user, and based on this, efficient document association and personalized creative generation and paper outline generation are achieved, which improves the efficiency and quality of academic paper writing and solves the technical problem that the papers written in the related technology have a low correlation with user needs and are not very targeted.

[0017] Details of one or more embodiments of the present application are set forth in the following drawings and description to make other features, objects, and advantages of the present application more readily apparent. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings: Figure 1 A schematic diagram of a paper generation method provided in an embodiment of the present application is shown; Figure 2 A flowchart of a method for generating a paper based on a large language model of RAG provided in an embodiment of the present application is shown; Figure 3 A structural block diagram of a paper generation system provided by an embodiment of the present application is shown; Figure 4 A structural block diagram of a paper generation device based on a large language model of RAG provided in an embodiment of the present application is shown; Figure 5 A hardware structure block diagram of a computer device provided in an embodiment of the present application is shown. DETAILED DESCRIPTION

[0019] In order to more clearly understand the purpose, technical solutions and advantages of the present application, the present application is described and illustrated below in conjunction with the accompanying drawings and embodiments.

[0020] Unless otherwise defined, the technical terms or scientific terms involved in this application shall have the general meaning understood by people with general skills in the technical field to which this application belongs. The words "one", "a", "the", "these" and the like in this application do not indicate a quantitative limitation, and they may be singular or plural. The terms "include", "comprise", "have" and any variants thereof involved in this application are intended to cover non-exclusive inclusions; for example, a process, method and system, product or device comprising a series of steps or modules (units) is not limited to the listed steps or modules (units), but may include unlisted steps or modules (units), or may include other steps or modules (units) inherent to these processes, methods, products or devices. The words "connect", "connected", "coupled" and the like involved in this application are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. The "multiple" involved in this application refers to two or more. "And / or" describes the association relationship of associated objects, indicating that there may be three relationships, for example, "A and / or B" may mean: A exists alone, A and B exist at the same time, and B exists alone. Generally, the character " / " indicates that the objects associated with each other are in an "or" relationship. The terms "first", "second", "third", etc. in this application are only used to distinguish similar objects and do not represent a specific ordering of the objects.

[0021] Traditional large language models (such as GPT-3) rely on knowledge learned during training. The generated text is based on the model's internal weights and input context. It lacks access to external real-time information, so it lacks knowledge when it comes to highly specialized queries. For example, when the required information exceeds the training data of the large language model, the large language model cannot provide accurate answers. Directly applying the large language model to paper writing will result in the generated results being non-targeted and unable to provide good support for paper writing.

[0022] Based on this, the present application embodiment provides a method for generating papers based on a large language model of RAG. Figure 1 A schematic diagram of a paper generation method provided in an embodiment of the present application is shown, such as Figure 1 As shown, in this method, firstly, a search is performed in the knowledge base according to the target research field to obtain the first reference content associated therewith, and the first reference content is processed by text generation to generate multiple initial research directions. The target research direction is screened out from the initial research directions, and based on the target research direction, a search is performed in the knowledge base to obtain the second reference content associated therewith. The second reference content is processed by text generation to obtain the initial paper outline.

[0023] The embodiment of the present application combines the retrieval and generation capabilities of the RAG technology to provide the large language model with additional contextual information in the knowledge base. The content generated in combination with the contextual information is more in line with the user's needs. Based on this, efficient document association and personalized creative generation and paper outline generation are achieved, thereby improving the efficiency and quality of academic paper writing, and solving the technical problem that the papers written in related technologies have a low correlation with user needs and are therefore not very targeted.

[0024] The method provided in the embodiment of the present application is applied to a paper generation system based on a large language model of RAG, which includes: a back-end module, a front-end module, a storage module, and a network connection module.

[0025] The back-end module provides knowledge base management, research direction generation, paper outline generation, and academic paper body generation services. This module calls the large language model through professional prompt words. Specifically, the large language model generates multiple innovative research directions (equivalent to the initial research direction) based on the target research field input by the user; the large language model generates the initial paper outline based on the preset paper type and target research direction; the large language model generates the paper body that meets academic standards based on the paper outline. The back-end module can also provide polishing functions for the output text, such as polishing the paper body. The back-end module can also optimize the generated research direction, paper outline, and paper body through the user feedback mechanism to improve the quality and relevance of the generated content.

[0026] The front-end module receives the research direction selected by the user, the paper outline edited by the user, references, and the paper text polished by the user. The module interacts with the user through a graphical user interface (GUI) and provides page interaction-friendly operation methods, such as drag-and-drop editing, real-time preview, and content modification. The visual editing tools provided by the front-end module include but are not limited to: drag-and-drop interface, text selection, version control, and history record functions, so that users can efficiently select research directions and modify outlines and papers.

[0027] The storage module stores the user's personal knowledge base, research direction, paper outline, and paper text. This module stores user data securely and efficiently through a user account isolation mechanism, including: using encryption technology to protect the privacy of document data, using distributed storage technology to improve the efficiency of data access, and providing data backup and recovery functions to prevent data loss. The storage module is also used to store the user's editing history to facilitate the system's rapid response and long-term data preservation, and supports data backup and recovery functions. The storage module includes: a high-performance database system for efficient storage and retrieval of user editing history and rating records; a backup unit for regular data backup to prevent data loss; a recovery unit for data recovery when data is damaged or lost; and data compression and encryption functions to improve storage efficiency and data security.

[0028] The network connection module is used to connect to the user's personal knowledge base to obtain reference data, and to call the API of the large language model, supporting high-bandwidth and low-latency data transmission, including: achieving seamless integration of multiple data sources to ensure real-time updating and synchronization of literature data; using advanced network protocols and optimization algorithms to improve the efficiency and stability of data transmission; and providing a secure communication channel to protect the integrity and confidentiality of data. The network connection module is also used to communicate with cloud servers to achieve data synchronization and sharing, ensuring that users access the latest documents and edited content on different devices. The network connection module includes: a cloud synchronization unit for real-time updating of user data; cross-platform compatibility function to support seamless access on multiple devices and operating systems; a security protocol module to ensure the integrity and confidentiality of data transmission; and bandwidth optimization technology to improve the efficiency and response speed of data synchronization.

[0029] The following is an introduction to the method used in the paper generation system above. Figure 2 A flowchart of a method for generating a paper based on a large language model of RAG provided in an embodiment of the present application is shown. Figure 2 As shown, the method includes: Step S201: In response to a target research field input by a user, a preset knowledge base is searched to obtain a first reference content associated with the target research field.

[0030] In this step, the preset knowledge base includes multiple documents uploaded by the user, and the multiple documents may involve multiple research fields. Searching the preset knowledge base means vectorizing the user input and the documents in the knowledge base, and comparing the correlation between the vectorized user input features and the document features. If the correlation meets the requirements, the first reference content corresponding to the document feature is output, where the first reference content can be the document itself, or a text, figure, or table in the document.

[0031] In some embodiments, a multi-channel recall scheme for the first reference content can be implemented based on keyword recall and semantic recall. Specifically, the keywords and semantic information input by the user are extracted respectively, and the associated first reference content is retrieved in the knowledge base based on the keywords and semantic information respectively. By using advanced indexing algorithms and efficient retrieval technology, users can quickly locate the required documents, which greatly improves the efficiency and convenience of document management.

[0032] In some embodiments, the knowledge base is obtained after classifying and labeling multiple documents uploaded by users. That is, before step S201, the method further includes: receiving multiple documents uploaded by users, classifying and labeling multiple documents, and obtaining a knowledge base constructed with multiple documents. In this embodiment, the documents uploaded by users are managed through a large language model, including document classification, labeling, full-text retrieval, data backup and recovery, security management, etc.

[0033] Step S202: Perform text generation processing of a large language model on the first reference content to generate multiple initial research directions.

[0034] Among them, the large language model can be pre-trained based on a general corpus and has good context learning ability. Therefore, the retrieved first reference content can be used as context information and input into the large language model, and the text generation technology of the large language model can be used to output multiple initial research directions.

[0035] For example, the target research field is the paleogeographic reconstruction and tectonic evolution of the Mesozoic North China Craton. The search is conducted in the knowledge base, and the initial research directions generated by combining the retrieved highly relevant content are three-dimensional seismic imaging of the tectonic evolution of the Mesozoic North China Craton, the application of paleomagnetic methods in the paleogeographic reconstruction of the North China Craton, and the study of the collision and subduction of the North China Craton and surrounding plates. It can be understood that the multiple innovative research directions generated based on the target research field input by the user and the first reference content queried in the knowledge base are highly relevant to the documents in the knowledge base, so they are more targeted.

[0036] Step S203: perform deduplication processing on the initial research directions, and score the deduplicated initial research directions based on preset evaluation requirements to obtain a target scoring result.

[0037] This step removes duplicates from the initial research directions. After removal of duplicates, the initial research directions are diverse, allowing users to choose a more suitable research direction. In addition, they can be sorted and polished to ensure the logic and innovation of the text.

[0038] In this step, the initial research direction is scored using the scoring prompts of the large language model. The scoring can be performed from multiple aspects, such as innovation, relevance, and feasibility. The comprehensive score is calculated based on the scores of the initial research direction in various aspects and the corresponding weights, and the comprehensive score is used as the target scoring result.

[0039] In some embodiments, step S203 further includes: based on the innovation, relevance, and feasibility of the preset initial research directions, scoring the initial research directions after deduplication, respectively, to obtain an innovation score, a relevance score, and a feasibility score; combining the innovation score, the relevance score, and the feasibility score to obtain the target scoring result corresponding to the initial research direction.

[0040] This embodiment systematically and objectively evaluates research directions from multiple dimensions, providing support for researchers in selecting topics. At the same time, this scoring system also helps to dynamically adjust research strategies to ensure that research directions keep pace with the times and can cope with changing scientific research needs and technological advances.

[0041] In some of the embodiments, the above-mentioned combination of innovation score, relevance score, and feasibility score to obtain the target scoring result corresponding to the initial research direction includes: taking the sum of the innovation score, relevance score, and feasibility score as the first scoring result corresponding to the initial research direction; scoring the initial research direction in combination with expert knowledge to obtain a second scoring result; eliminating the initial research directions whose average values ​​of the first scoring results and the second scoring results are lower than a preset threshold, to obtain the target scoring result corresponding to the screened initial research directions.

[0042] In this embodiment, experts make a judgment on the actual potential of the research direction based on factors such as trends in the field and cutting-edge technologies, and comprehensively evaluate the value of the research direction by combining expert knowledge and the scores of the large language model to ensure the objectivity and accuracy of the evaluation results.

[0043] In addition, this embodiment sets a threshold to eliminate research directions with low comprehensive scores of innovation, relevance, and feasibility, thereby effectively filtering out directions that do not have high development potential. This helps researchers focus on high-potential research directions and avoid wasting resources on research directions that are not feasible or innovative.

[0044] Step S204: sort the initial research directions in descending order of the target scoring results, and display the initial research directions based on the sorting results.

[0045] Continuing with the initial research directions of three-dimensional seismic imaging of the tectonic evolution of the North China Craton in the Mesozoic, the application of paleomagnetic methods in the paleogeographic reconstruction of the North China Craton, and the study of the collision and subduction between the North China Craton and surrounding plates as examples, these three initial research directions are sorted according to the target scoring results, and the initial research directions with high target scoring results are displayed before the initial research directions with low target scoring results, so that researchers can focus on high-potential research directions first and avoid investing too many resources in research directions that do not have high value.

[0046] Step S205: In response to the user selecting the target research direction with the highest score from multiple initial research directions, the knowledge base is searched to obtain second reference content associated with the target research direction, and the second reference content is processed by a large language model to generate text to generate an initial paper outline.

[0047] In this step, the large language model retrieves the second reference content in the knowledge base based on the target research direction, inputs the second reference content into the large language model as context information, and uses the text generation technology of the large language model to output multiple initial paper outlines.

[0048] The algorithm of the generation part is as follows: First, according to the paper type selected by the user, the corresponding template T is selected. Then, the outline D is generated by combining the template and the retrieved highly relevant document content. Where D = f (T, top-k document content).

[0049] In some of the embodiments, the second reference content is processed by a large language model to generate an outline structure text; the third reference content associated with each chapter in the outline structure text is extracted from the knowledge base; and the outline structure text and the third reference content are combined to obtain an initial paper outline.

[0050] In this embodiment, the third reference content is filled into the outline structure text, so that the content in the initial paper outline has a reference basis, and is therefore more closely related and accurate to the topic to be studied by the researcher.

[0051] In some of the embodiments, after step S205, the method further includes: receiving a user's modification operation on the initial paper outline to obtain a target paper outline; based on the target paper outline and the second reference content, using the text generation technology of the large language model to generate an initial academic paper corresponding to the target paper outline.

[0052] This embodiment allows researchers to edit and optimize the generated initial paper outline, improve the logic and applicability of the outline, and ensure that the final output academic paper meets the academic goals and expectations of the researchers. Specifically, users can personalize the generated outline through the front-end editing tool to meet specific research needs and academic norms. The editing tool provides real-time preview and version control functions to ensure the flexibility and security of user editing. Then, the text generation technology of the large language model is used to generate the main body of the paper in combination with the modified target paper outline and the second reference content, and the coherence and logic of the paper are ensured based on the requirements of the prompt words. Since the initial academic paper is generated in combination with the second reference content in the knowledge base, it is more closely related to the researcher's research topic and is more targeted.

[0053] In some of the embodiments, the method also includes: polishing the initial academic paper to obtain a polished target academic paper; determining the scores of the evaluation types based on the corresponding scoring strategies of each evaluation type of the preset target academic paper, the evaluation types including innovation, relevance, and feasibility; and generating a third scoring result corresponding to the target academic paper by combining the scores obtained for each evaluation type of the target academic paper.

[0054] In this embodiment, the large language model polishes the initial academic paper, improves the language expression, structure and logic of the initial academic paper, and makes it more academic and readable. The polished target academic paper is more in line with academic norms, and improves its professionalism and clarity of expression.

[0055] Among them, polishing includes grammar checking, terminology unification, style adjustment, standardization of literature citation format, and typesetting adjustment to conform to the target journal or conference submission format. The above polishing methods ensure the professionalism, consistency and high readability of the paper.

[0056] After the model is polished, users can further edit and polish the generated paper text and view the modification effect through the real-time preview function to ensure the accuracy and consistency of the final text. The system supports collaborative editing, allowing multiple users to make modifications at the same time.

[0057] In addition, a comprehensive scoring result can be obtained by combining the three evaluation types of innovation, relevance and feasibility. The scores obtained in each evaluation dimension can reflect the performance of the paper in different aspects, thus generating a comprehensive scoring result. In addition, the combination of large language models and expert scoring ensures the objectivity and accuracy of the evaluation results.

[0058] In the embodiment recorded in the above steps S201 to S205, the user's knowledge base is constructed by receiving documents uploaded by the user, and the documents in the knowledge base are managed to ensure efficient organization, convenient retrieval and safe storage of the documents. Then, the knowledge base is associated with the content input by the user, and reference content associated with the knowledge base is generated. The text generation technology of the large language model is used to generate research directions, paper outlines or paper texts based on the reference content. The reference content in the knowledge base associated with the user input is retrieved by RAG technology, so that the research direction, paper outline and paper text generated based on the reference content are highly relevant to the knowledge base, more targeted, and the efficiency and quality of text construction are improved.

[0059] The method provided in the embodiment of the present application is described below with a specific example, and the method is applied to a paper generation system. Figure 3 A structural block diagram of a paper generation system provided by an embodiment of the present application is shown. Figure 3 As shown, the system includes a knowledge base management module 31, a creative generation module 32, a paper outline generation module 33, a paper body generation module 34, an intelligent evaluation module 35, a user terminal module 36 and a network connection module 37.

[0060] The knowledge base management module 31 supports users to upload documents, form a personal knowledge base, and systematically manage the documents in the knowledge base. The management includes the following aspects: Document classification: Uploaded documents are automatically classified through a large language model, and can also respond to the user's manual adjustment instructions to generate classification results, thereby realizing the functions of automatic and manual document classification.

[0061] Tag annotation: The document is structured through a large language model to automatically generate keywords and tags. Users can also add and modify tags to facilitate multi-dimensional retrieval and flexible management of documents.

[0062] Full-text search: After structuring and labeling the uploaded documents, a comprehensive full-text search function is built to ensure that users can find the required documents quickly and accurately.

[0063] Data backup and recovery: Regularly back up documents in the knowledge base and provide reliable data recovery functions to ensure the safe storage of documents and the integrity of data.

[0064] Security management: Through permission control and encryption technology, ensure the security of documents during uploading, storage and retrieval to prevent unauthorized access and data leakage.

[0065] The creative generation module 32 associates the user's personal knowledge base with the research field input by the user, searches the knowledge base in combination with the RAG technology, and generates multiple research directions with high innovation and diversity based on the search results. For details, please refer to the following steps: Vectorized representation: The research field input by the user and the literature in the knowledge base are represented by vectors using a large language model to obtain the query vector q and the literature vector k respectively. i , the correlation score between the query vector and the document vector is calculated by the inner product. The inner product formula includes: score(q, k i )=q×k i . The relevance score captures the semantic relationship and potential association between research fields and documents, as well as between documents.

[0066] Retrieval mechanism: Extract highly relevant content from the knowledge base through vector retrieval as reference content for generating research directions, ensuring that the generated results are highly relevant to the user's existing knowledge base.

[0067] Generate research directions: Combine the retrieved highly relevant content to generate multiple potential research directions. RAG technology is used in the process of generating research directions, which can ensure the innovation of the generated content while taking into account its relevance to the user input field.

[0068] As an example, the prompt words used for idea generation include the following: The first part is the role. The prompt for the role part is: You are an innovative researcher who focuses on the field of scientific research. You can combine cross-disciplinary methods to explore cutting-edge and innovative research directions and focus on specific aspects.

[0069] The second part is skills. The prompt for skill 1 is: Generate innovative research directions. First, analyze the latest progress in scientific research and the content of documents in the knowledge base. Second, find areas that have not been fully explored or emerging technology combinations to generate research directions. Response examples include direction name, theoretical support, and application potential.

[0070] The prompt for skill 2 is: Focus on key challenges and bottlenecks. First, identify the key challenges and bottlenecks in the field of user research. Second, propose possible research directions based on this. The response example is the same as skill 1.

[0071] The prompt for skill 3 is: Use artificial intelligence technology. First, think about how to use artificial intelligence technology to solve real-world problems. Second, generate relevant research directions. The response example is the same as skill 1.

[0072] The prompt for skill 4 is: Discover innovative concepts. First, look for innovative concepts that may have a significant impact or overturn the status quo. Second, transform them into research directions. The response example is the same as skill 1.

[0073] The prompt for skill 5 is: Explore the intersection of disciplines. First, identify the intersection with other disciplines. Second, propose research directions that may bring revolutionary changes. The response example is the same as skill 1.

[0074] The third part is the restriction. The prompt words for the restriction part are: only generate research directions around the scientific research field, each direction must have theoretical support and application potential, and the output content must be organized in a given format and cannot deviate from the framework requirements.

[0075] After generating the research directions, the creative generation module 32 provides sorting and scoring services to deduplicate the generated research directions. Subsequently, the research directions are sorted and scored based on multi-dimensional evaluation indicators, where the evaluation indicators include innovation, relevance, feasibility, etc.

[0076] As an example, the prompt words used for sorting and scoring include the following: The first part is the role. The prompt for the role part is: You are a professional research direction analysis and evaluator who can accurately deduplicate, sort and score research directions.

[0077] The second part is skills. The prompt for skill 1 is: remove duplicates. First, carefully check the 10 research directions generated and remove duplicate or overly similar directions. Second, for similar directions, merge and extract the core points to form a new unique direction.

[0078] The prompt word for skill 2 is: scoring. First, innovation scoring (0-3 points): In-depth analysis of the novelty and breakthrough potential of each research direction. Consider whether the direction introduces new concepts, methods or technologies, and its possible significant impact on related fields. Second, relevance scoring (0-3 points): Evaluate the fit and importance of the research direction with the user research field. Determine whether the direction is closely centered on the core issues in the user research field and its potential contribution to solving problems in this field. Third, feasibility scoring (0-3 points): Evaluate the feasibility of the research direction in terms of technical implementation and resource requirements. Consider the required technology, manpower, time, and financial resources, as well as whether the current level of technology and actual conditions can support the implementation of this direction.

[0079] The prompts for skill 3 are: sort and provide reasons. First, sort the research directions from high to low according to their total score (the sum of innovation, relevance, and feasibility). Second, provide a brief evaluation reason for each research direction to support the scoring results. The reasons should include a specific analysis of innovation, relevance, and feasibility.

[0080] The third part is the restriction. The prompt words of the restriction part are to strictly analyze and evaluate according to the given steps and dimensions. Ensure that the scoring process is objective and fair, and not affected by personal subjective factors. The total score shall not exceed 10 points, and the score of each dimension shall be within the specified range. The brief evaluation reason shall not exceed 100 words.

[0081] Before users make their choices, a large language model is used to further polish and optimize the ranked research directions to improve the accuracy and attractiveness of their expressions, ensuring that the recommendations users receive have high reference value.

[0082] As an example, some of the prompts used for editing include the following: The first part is the role. The prompt for the role part is: You are a professional research direction polisher and optimizer who can enhance the professionalism and innovation of the research direction while maintaining the original theme, making it more attractive and convincing.

[0083] The second part is skills. The prompt for skill 1 is: Improve professionalism. First, use professional terms and expressions in the field to describe the research direction. Second, cite relevant theories, models or technical backgrounds to demonstrate the academic depth of the research direction. Third, ensure that the description is accurate and in line with the latest developments in the current research field.

[0084] The prompt for skill 2 is: Enhance innovation. First, highlight the novelty and potential breakthroughs of the research direction. Second, emphasize the differences and unique contributions from existing research. Third, describe the practical application value and future development potential that may be brought.

[0085] The prompt words for skill 3 are: optimize structure and expression. First, make the description of each research direction concise and clear, and the logic is clear. Second, use attractive and persuasive language to enhance the appeal of the research direction. Third, ensure that the format is unified and easy to read and understand.

[0086] The third part is the restriction. The prompt words of the restriction part are: Do not change the theme and core content of the research direction during the polishing and optimization process. Only optimize the research direction and do not process other irrelevant content. Strictly output according to the given format and do not deviate.

[0087] After finishing the polishing, the paper outline generation module 33 generates the paper outline based on the research direction selected by the user and the preset paper type, and the specific implementation is as described below.

[0088] Automatic outline generation: According to the professional prompt words corresponding to the user's research direction and paper type, relevant content is retrieved from the user's personal knowledge base to ensure that the generated outline has a solid theoretical foundation and rich reference materials.

[0089] As an example, the prompt words used for automatic outline generation include the following: The first part is the role. The prompt for the role part is: You are a professional literature review writer who can accurately retrieve literature on a specific research direction from your personal knowledge base and generate a structured and professional literature review outline.

[0090] The second part is skills. The prompt for skill 1 is: literature search. First, go deep into your personal knowledge base and filter out documents, papers, reports, books and other materials that are highly relevant to a specific research direction. Second, ensure that the search content covers classic research and the latest progress in the field. The response example is as follows: document type, such as document, paper, report or book; publication time; main content, which is used to briefly summarize the main content of the document.

[0091] The prompt for skill 2 is: Generate a literature review outline. First, write an introduction to clearly introduce the background and importance of the research direction, and clarify the purpose and scope of the literature review. Second, outline the relevant basic theories and concepts, and introduce key models and technical frameworks. Third, summarize the main research results and progress in this field, and analyze the connections and differences between different studies. Fourth, identify the main problems and challenges in current research, and explain their impact on the research direction. Fifth, describe recent research trends and cutting-edge technologies, and predict future development directions and breakthroughs. Sixth, discuss the cases and effects of the research direction in practical applications, and analyze problems and solutions in practice. Seventh, summarize the main findings and conclusions of the literature review, and put forward suggestions and potential directions for future research.

[0092] The third part is the restriction. The prompt words of the restriction part are: literature search and outline generation only focus on specific research directions. The outline structure must be clear, logical, detailed and professional. Output should be in the given format and should not deviate from the framework requirements.

[0093] After the outline generation step, the reference content retrieval step is performed: based on the generated outline structure, reference content related to each chapter is extracted from the user's personal knowledge base to ensure that each part has detailed supporting materials.

[0094] Then edit and optimize the outline: allow users to edit and optimize the generated outline and reference content, improve the logic and applicability of the outline, and ensure that the final output academic paper meets the user's academic goals and expectations.

[0095] According to the modified paper outline and reference content, the paper body generation module 34 generates the paper body, and the specific implementation is as described below.

[0096] Automatic text generation: Based on the paper outline and reference content provided by the user, the large language model is used to generate the main text of the academic paper to ensure the coherence and logic of the paper content.

[0097] As an example, the prompt words used to automatically generate the body of the paper include the following: The first part is the role. The prompt for the role part is: You are a professional literature review writer who can write a high-quality and logically clear literature review text based on the given literature review outline and reference content.

[0098] The second part is skills. The prompt for skill 1 is: Introduction. First, conduct in-depth research on the research direction, clearly explain its background and importance, and provide readers with a comprehensive cognitive basis. Second, clarify the purpose and research questions of the literature review to ensure that readers have a clear understanding of the focus of the review. The response example is as follows: Introduction: Research direction background, used to describe the background of the research direction in detail; Importance statement, used to illustrate the importance of the research direction in academic and practical applications; Purpose and problem, used to clarify the purpose of the literature review and specific research questions.

[0099] The prompt for skill 2 is: theoretical foundation. First, explain the relevant basic theories and concepts in detail to ensure that readers have a deep understanding of the theoretical framework. Second, introduce and analyze key models and technical frameworks, highlighting their characteristics and application scenarios. The response example is as follows: Theoretical foundation: basic theoretical explanation, used to elaborate on the relevant basic theories and concepts; key model analysis, used to introduce and analyze the characteristics and application scenarios of key models; technical framework analysis, used to analyze the composition and advantages of the technical framework.

[0100] The prompt for skill 3 is: Research status section. First, systematically summarize the main research results and progress in the field to present a comprehensive research status to readers. Second, compare the methods and conclusions of different studies, point out their connections and differences, and help readers understand the diversity of research. The response example is as follows: Research status: Summary of main results, used to systematically summarize the main research results in the field; Overview of progress, used to summarize the research progress in the field; Comparison of methods and conclusions, used to compare the methods and conclusions of different studies, and point out the connections and differences.

[0101] The prompt for skill 4 is: key issues and challenges. First, deeply analyze the main problems and challenges in current research to provide direction for subsequent research. Second, discuss the impact of these problems on research direction and practical application, and emphasize the urgency of solving problems. The response example is as follows, key issues and challenges: problem analysis, used to deeply analyze the main problems in current research; challenge statement, used to explain the challenges faced by current research; impact discussion, used to discuss the impact of these problems and challenges on research direction and practical application.

[0102] The prompt for skill 5 is: the latest progress and trends section. First, describe the latest research trends and cutting-edge technologies to let readers know the latest developments in the field. Second, predict future development directions and possible breakthroughs to provide forward-looking thinking for future research. The following is a sample response: the latest progress and trends: research trend description, used to describe the latest research trends; cutting-edge technology introduction, used to introduce cutting-edge technologies in the field; development direction prediction, used to predict future development directions; breakthrough point analysis, used to analyze possible breakthroughs.

[0103] The prompt word for skill 6 is: application and practice. First, discuss the cases and effects of the research direction in practical applications to enhance the practicality of the research. Second, analyze the problems in practice and propose solutions to provide guidance for practical applications. The response example is as follows: application and practice: case analysis, used to discuss the cases and effects of the research direction in practical applications; problems and solutions, used to analyze the problems in practice and propose solutions.

[0104] The prompt words for skill 7 are: Summary and Outlook. First, summarize the main findings and conclusions of the literature review to provide readers with a clear summary of the review. Second, propose suggestions and potential directions for future research, emphasizing innovation and development potential. The response example is as follows: Summary and Outlook: Summary of Main Findings, used to summarize the main findings of the literature review; Conclusion Exposition, used to elaborate the conclusions of the literature review; Suggestions and Directions, used to propose suggestions and potential directions for future research.

[0105] The third part is the restriction. The prompt words for the restriction part are: only write around the given research direction and references, and do not deviate from the topic. Ensure that the content is professional, the structure is complete, the logic is clear, and it complies with academic standards. The references are accurate, and professional terms and expressions are used to enhance academic and authoritativeness. The content must not exceed the prescribed word limit and ensure conciseness and clarity. Organize according to the given format and do not change it at will.

[0106] After generating the main text of the paper, the large language model also provides a paper polishing function to improve the language quality and readability of the paper. Specific polishing functions include but are not limited to: grammar checking, which refers to automatically detecting and correcting grammatical errors in the text to ensure the accuracy of language expression. Terminology unification refers to the unified processing of professional terminology to ensure the consistency and standardization of terminology throughout the text. Style adjustment refers to adjusting the writing style of the paper according to the requirements of the target journal or conference to ensure compliance with academic norms. Standardization of literature citation format refers to automatically adjusting the literature citation format to ensure that it meets the submission requirements of the target journal or conference. Layout adjustment refers to automatically adjusting the layout of the paper according to the submission format of the target journal or conference to ensure that the paper meets the submission format requirements.

[0107] After the text is polished, the intelligent evaluation module 35 uses machine scoring and expert scoring to comprehensively score the research direction, the outline modified by the user and the polished text. Please refer to the description below for details.

[0108] Machine scoring: Use a large model to automatically score the generated research direction, outline, and text to evaluate its innovation, logic, and language quality. Taking outline scoring as an example, the prompt words used in outline scoring include the following: The first part is the role. The prompt for the role part is: You are a professional literature review outline scorer who can accurately and objectively score the literature review outline and give detailed evaluation reasons and constructive feedback.

[0109] The second part is skills. The prompt for skill 1 is: professional scoring. First, check whether the outline uses professional terms and expressions in the field. If so, add points and explain the specific professional terms and expressions. Second, check whether the outline cites relevant theories, models or technical backgrounds. If so, add points and list the cited content. Third, confirm whether the outline content is accurate and consistent with the latest developments in the current research field. If so, add points and explain the aspects that are consistent.

[0110] The prompt for skill 2 is: Completeness score. First, check whether the outline covers all the key parts of the literature review, such as introduction, theoretical basis, research status, key issues and challenges, latest progress and trends, applications and practices, summary and outlook. If yes, add points and indicate the parts covered. Second, evaluate whether each part is detailed and rich in content, avoid omitting important information, if yes, add points and indicate the detailed and rich parts.

[0111] The prompt for skill 3 is: Logical scoring. First, determine whether the outline structure is clear and the logic is rigorous. If so, add points to show the aspects of clear structure and rigorous logic. Second, check whether the connection between each part is natural and whether it helps readers understand. If so, add points to show the parts of natural connection.

[0112] The prompt for skill 4 is: Application value scoring. First, determine whether the outline has practical application value and whether it can provide clear guidance for the writing of subsequent literature reviews. If so, add points and explain the application value. Second, evaluate whether the outline helps to reveal key issues and future development potential in the research direction. If so, add points and explain the revealed key issues and development potential.

[0113] The third part is the restriction. The restriction prompts are: only score and provide feedback on the literature review outline, and refuse to answer questions that are not related to the scoring. The scoring process must be objective and fair, and be carried out according to the given scoring dimensions and standards. Provide a brief evaluation reason for each scoring item to support the scoring result. Summarize the scores of each item and give an overall evaluation and suggestions.

[0114] In addition to machine evaluation, domain experts are invited to conduct blind evaluation on the generated content to ensure the fairness and subjectivity of the evaluation process and provide scoring results based on professional judgment. Taking the research direction scoring as an example, the scoring table for experts is shown in Table 1.

[0115] Table 1

[0116] Combined with Table 1, we can see that the experts score from multiple dimensions such as innovation, relevance, and feasibility to ensure the comprehensiveness and accuracy of the evaluation results.

[0117] The user-side module 36 in the system includes a visual editing tool for providing intuitive GUI interaction, allowing users to select research directions and reference content, and to edit the generated paper outline and text in real time. The tool functions include: text highlighting and annotation functions, which are used to facilitate users to mark and annotate key content; version control and history record functions, which are used to support version management and historical tracing of documents, ensuring that users can easily trace back and compare different versions. Real-time synchronization function is used to achieve real-time update and synchronization of documents, improving the efficiency of academic paper writing.

[0118] The user-side module 36 also includes data security and storage, which is used to achieve user account isolation and ensure the safe storage of user data. The following technical measures are adopted to protect data security. First, data encryption: advanced encryption technology is used to ensure the privacy of document data and prevent unauthorized access. Second, distributed storage: improve data access efficiency and ensure that users can quickly access and modify documents. Third, data backup and recovery: provide automated data backup and recovery functions to prevent data loss and ensure the integrity and security of user data.

[0119] The network connection module 37 in the system is used to connect the user's personal knowledge base and the large language model, obtain relevant reference data and call the API of the large language model. Among them, first, seamless data integration refers to supporting the integration of multiple data sources, ensuring real-time updating and synchronization of literature data, and providing a consistent access interface to simplify data interaction. Second, efficient data transmission refers to the use of advanced network protocols and optimization algorithms to support high-bandwidth and low-latency data transmission to improve transmission efficiency and system response speed. Third, secure communication channel: through encryption technology and authentication mechanism, a secure communication channel is provided to ensure the integrity and confidentiality of data transmission, and prevent data leakage and unauthorized access.

[0120] Through the above technical solution, the present invention can achieve personalized writing assistance and accurate reference retrieval while ensuring the quality of content generation, significantly improving the efficiency and quality of creative generation, paper outline construction and academic paper writing.

[0121] Figure 4 FIG. 4 shows a structural block diagram of a paper generation device based on a large language model of RAG provided in an embodiment of the present application. Figure 4 As shown, the device comprises: A search module 41 is used to search a preset knowledge base in response to a target research field input by a user, and obtain a first reference content associated with the target research field; A research direction generating module 42, configured to perform text generation processing of a large language model on the first reference content to generate a plurality of initial research directions; An evaluation module 43 is used to perform deduplication processing on the initial research directions, and score the deduplication-free initial research directions based on preset evaluation requirements to obtain target score results corresponding to the initial research directions; A sorting module 44 is used to sort the initial research directions in descending order of the target scoring results, and display the initial research directions based on the sorting results; The writing module 45 is used to search the knowledge base in response to the target research direction with the highest score selected by the user from multiple initial research directions, obtain the second reference content associated with the target research direction, and perform text generation processing of the large language model on the second reference content to generate an initial paper outline.

[0122] In some of the embodiments, the evaluation module 43 is also used to score the initial research directions after deduplication based on the innovation, relevance, and feasibility of the preset initial research directions, to obtain an innovation score, a relevance score, and a feasibility score; and the target score result corresponding to the initial research direction is obtained by combining the innovation score, the relevance score, and the feasibility score.

[0123] In some of the embodiments, the evaluation module 43 is also used to use the sum of the innovation score, the relevance score, and the feasibility score as the first scoring result corresponding to the initial research direction; score the initial research direction in combination with expert knowledge to obtain a second scoring result; eliminate the initial research directions whose average values ​​of the first scoring results and the second scoring results are lower than a preset threshold, to obtain a target scoring result corresponding to the screened initial research direction.

[0124] In some of the embodiments, the writing module 45 is also used to perform text generation processing of the large language model on the second reference content to generate an outline structure text; extract the third reference content associated with each chapter in the outline structure text from the knowledge base; and combine the outline structure text and the third reference content to obtain an initial paper outline.

[0125] It should be noted that the above modules can be functional modules or program modules, and can be implemented by software or hardware. For modules implemented by hardware, the above modules can be located in the same processor; or the above modules can be located in different processors in any combination.

[0126] This embodiment also relates to a computer device, Figure 5 FIG. 1 shows a hardware structure block diagram of a computer device provided in an embodiment of the present application. Figure 5 As shown, the computer device may include one or more ( Figure 5 The computer device may further include a transmission device 506 and an input / output device 508 for communication functions. It will be appreciated by those skilled in the art that Figure 5 The structure shown is only for illustration and does not limit the structure of the above-mentioned computer device. Figure 5 More or fewer components as shown, or with Figure 5 Different configurations are shown.

[0127] The memory 504 can be used to store computer programs, for example, software programs and modules of application software. The processor 502 executes various functional applications and data processing by running the computer programs stored in the memory 504, that is, to implement the above method. The memory 504 can be used to store data, for example, to store the user's personal information, knowledge base, research direction, paper outline, paper text, etc. The memory 504 may include a high-speed random access memory, and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some examples, the memory 504 may further include a memory remotely arranged relative to the processor 502, and these remote memories may be connected to the computer device via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0128] The transmission device 506 is used to receive or send data via a network. The specific example of the above network may include a wireless network provided by a communication provider of the computer device. In one example, the transmission device 506 includes a network adapter (Network Interface Controller, referred to as NIC), which can be connected to other network devices through a base station so as to communicate with the Internet. In one example, the transmission device 506 can be a radio frequency (Radio Frequency, referred to as RF) module, which is used to communicate with the Internet wirelessly. A computer device is also provided, including a memory and a processor, the memory storing a computer program, and the processor is configured to run the computer program to execute the steps in any of the above method embodiments.

[0129] Optionally, the computer device may further include a transmission device and an input / output device, wherein the transmission device is connected to the processor, and the input / output device is connected to the processor.

[0130] Optionally, in this embodiment, the processor may be configured to execute the following steps S1 to S5 through a computer program.

[0131] S1. In response to a target research field input by a user, a preset knowledge base is searched to obtain a first reference content associated with the target research field.

[0132] S2. Perform text generation processing of a large language model on the first reference content to generate multiple initial research directions.

[0133] S3. De-duplicate the initial research directions, and score the de-duplicate initial research directions based on preset evaluation requirements to obtain the target score results.

[0134] S4. Sort the initial research directions in descending order of the target score results, and display the initial research directions based on the sorting results.

[0135] S5. In response to the user selecting the target research direction with the highest score from multiple initial research directions, the knowledge base is searched to obtain second reference content associated with the target research direction, and the second reference content is processed by a large language model for text generation to generate an initial paper outline.

[0136] It should be noted that the specific examples in this embodiment can refer to the examples described in the above embodiments and optional implementation modes, and will not be repeated in this embodiment.

[0137] In addition, in combination with the method provided in the above embodiment, a storage medium may be provided in this embodiment to implement the method. The storage medium stores a computer program; when the computer program is executed by a processor, any of the above embodiments of the method for generating a paper based on a large language model of RAG is implemented.

[0138] The embodiment of the present application also provides a computer program product. When the computer program product is run on a computer, the computer executes each function or step executed by the processor in the above method embodiment.

[0139] It should be understood that the specific embodiments described herein are only used to explain the application, rather than to limit it. Based on the embodiments provided in this application, all other embodiments obtained by ordinary technicians in this field without creative work are within the protection scope of this application.

[0140] Obviously, the drawings are only some examples or embodiments of the present application. For ordinary technicians in the field, the present application can also be applied to other similar situations based on these drawings without creative work. In addition, it is understandable that although the work done in this development process may be complicated and lengthy, for ordinary technicians in the field, certain changes in design, manufacturing or production based on the technical content disclosed in this application are only conventional technical means and should not be regarded as insufficient content disclosed in this application.

[0141] The term "embodiment" in this application refers to a specific feature, structure or characteristic described in conjunction with the embodiment that can be included in at least one embodiment of the present application. The appearance of this phrase in various locations in the specification does not necessarily mean the same embodiment, nor does it mean that it is mutually exclusive with other embodiments and is independent or optional. It is clearly or implicitly understood by those of ordinary skill in the art that the embodiments described in this application can be combined with other embodiments without conflict.

[0142] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of patent protection. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the scope of protection of the present application. Therefore, the scope of protection of the present application shall be subject to the attached claims.

Claims

1. A paper generation method based on a large language model of RAG, characterized in that: The method comprises: In response to the target research field input by the user, searching a preset knowledge base to obtain first reference content associated with the target research field; Performing text generation processing of a large language model on the first reference content to generate multiple initial research directions; Deduplication processing is performed on the initial research directions, and based on preset evaluation requirements, the deduplication-free initial research directions are scored to obtain target scoring results corresponding to the initial research directions; Sorting the initial research directions according to the target scoring results from high to low, and displaying the initial research directions based on the sorting results; In response to the user selecting the target research direction with the highest score from the multiple initial research directions, the knowledge base is searched to obtain second reference content associated with the target research direction, and the second reference content is processed by a large language model for text generation to generate an initial paper outline.

2. The method for generating papers based on a large language model of RAG according to claim 1, characterized in that: The step of scoring the initial research direction after deduplication based on the preset evaluation requirements to obtain the target scoring result corresponding to the initial research direction includes: Based on the innovation, relevance, and feasibility of the preset initial research directions, the initial research directions after deduplication are scored respectively to obtain innovation scores, relevance scores, and feasibility scores; The innovation score, the relevance score, and the feasibility score are combined to obtain a target scoring result corresponding to the initial research direction.

3. The method for generating papers based on a large language model of RAG according to claim 2, characterized in that: The combining of the innovation score, the relevance score, and the feasibility score to obtain a target scoring result corresponding to the initial research direction includes: The sum of the innovation score, the relevance score, and the feasibility score is used as a first scoring result corresponding to the initial research direction; Scoring the initial research direction in combination with expert knowledge to obtain a second scoring result; The initial research directions whose average values ​​of the first scoring results and the second scoring results are lower than a preset threshold are eliminated to obtain the target scoring results corresponding to the screened initial research directions.

4. The method for generating papers based on a large language model of RAG according to claim 1, characterized in that: The step of performing text generation processing of a large language model on the second reference content to generate an initial paper outline includes: Performing text generation processing of a large language model on the second reference content to generate an outline structure text; Extracting third reference contents associated with each chapter in the outline structure text from the knowledge base; The initial paper outline is obtained by combining the outline structure text and the third reference content.

5. The method for generating papers based on a large language model of RAG according to claim 1, characterized in that: After generating an initial paper outline, the method further comprises: receiving a user's modification operation on the initial paper outline to obtain a target paper outline; Based on the target paper outline and the second reference content, an initial academic paper corresponding to the target paper outline is generated using text generation technology of a large language model.

6. The method for generating papers based on a large language model of RAG according to claim 5, characterized in that: After generating an initial academic paper corresponding to the target paper outline, the method further includes: Polishing the initial academic paper to obtain a polished target academic paper; Based on the scoring strategies corresponding to the various evaluation types preset for the target academic paper, the scores of the evaluation types are determined respectively, and the evaluation types include innovation, relevance, and feasibility; Combined with the scores obtained for each of the evaluation types for the target academic paper, a third scoring result corresponding to the target academic paper is generated.

7. The method for generating papers based on a large language model of RAG according to claim 1, characterized in that: The method further comprises: A plurality of documents uploaded by a user are received, and the plurality of documents are classified and labeled to obtain a knowledge base constructed by the plurality of documents.

8. A paper generation device based on a large language model of RAG, characterized in that: include: A retrieval module, configured to search a preset knowledge base in response to a target research field input by a user, and obtain a first reference content associated with the target research field; A research direction generation module, used for performing text generation processing of a large language model on the first reference content to generate multiple initial research directions; An evaluation module, used to perform deduplication processing on the initial research directions, and score the deduplication-free initial research directions based on preset evaluation requirements to obtain a target score result corresponding to the initial research directions; A sorting module, used to sort the initial research directions in descending order of the target scoring results, and display the initial research directions based on the sorting results; A writing module is used to search the knowledge base in response to a target research direction with the highest score selected by the user from multiple initial research directions, obtain second reference content associated with the target research direction, and perform text generation processing of a large language model on the second reference content to generate an initial paper outline.

9. A computer device comprising a memory and a processor, characterized in that: A computer program is stored in the memory, and the processor is configured to run the computer program to execute the paper generation method based on the RAG large language model as described in any one of claims 1 to 7.

10. A storage medium, characterized in that: The storage medium stores a computer program, wherein the computer program is configured to execute the paper generation method based on the RAG large language model according to any one of claims 1 to 7 when running.

Images