User personalized scientific research planning scheme generation method and system based on multi-agent interaction

Abstract

The invention provides a user personalized scientific research planning scheme generation method and system based on multi-agent interaction. The method comprises the following steps: firstly, receiving a question of a user, outputting a keyword through a keyword disassembling agent, further generating a related expert agent, integrating literature information, forming a theme through interactive discussion between the user and the multiple agents, and storing the theme into a global memory library; after a user selects a theme, entering an interactive inspiration and approval stage, generating an initial scheme by a scientific research scheme planning agent, and constructing a feasible literature reference database through second correlation screening; the keyword generation agent is combined with the database to generate literature keywords, a final combination is obtained through exchange and mutation operation, scheme feasibility is enhanced through the detailed agent, evaluation is conducted through the review agent, if optimization is needed, evaluation is conducted again after updating is conducted through the optimization agent, and if improvement is not needed, a final scheme is output. The framework links are compact, and abundant, efficient and accurate scientific research intelligent reply and planning schemes can be provided for users.

Description

Technical Field

[0001] The present invention relates to the field of scientific research planning scheme generation, and more specifically to a method and system for generating user personalized scientific research planning schemes based on multi-agent interaction. Background Art

[0002] Advancing scientific progress requires more innovative tools to help solve difficult problems. Since the early 1970s, automating the general scientific discovery process has been a long-term goal in this field. Traditionally, human researchers need to identify the target research area, collect background knowledge, draft a set of reasonable hypotheses and experimental plans, and conduct subsequent tests according to the plans. With the vigorous development of the field of artificial intelligence, the basic large language model (LLM) has made great progress, effectively promoting all stages of scientific discovery and achieving a certain degree of automation. With the help of the retrieval-augmented generation (RAG) algorithm, LLM can quickly grasp the key information of the target database and enhance the reliability of its output.

[0003] However, existing approaches to scientific research planning typically use LLMs, which reference single-domain databases, to assist in planning and management. This approach is limited in that it only considers literature information from a single domain, failing to account for the multidimensional relationships between documents and handling interdisciplinary or multi-domain queries. Summary of the Invention

[0004] In response to the defects existing in the above technical background, the present invention provides a method and system for generating user personalized scientific research planning schemes based on multi-agent interaction, which automatically generates personalized and innovative scientific research schemes.

[0005] The specific contents are as follows:

[0006] A method for generating a user-personalized scientific research plan based on multi-agent interaction includes the following steps:

[0007] S1, receiving user input questions, breaking down the intelligent agent into keywords and outputting a series of keywords;

[0008] S2: Based on user input and a series of keywords, relevant expert agents are generated based on the author database, relevant literature information is integrated based on the literature database, and users and multiple agents interact in discussions to obtain several discussion topics. All discussion content information will be stored in the global memory bank;

[0009] S3: The user selects a discussion topic and enters the interactive inspiration Spark stage and the interactive criticism Criticism stage. During this process, the global memory library is updated in real time.

[0010] S4, the research plan planning agent generates an initial research plan based on the global memory library and the discussion topic selected by the user;

[0011] S5. Based on the initial scientific research plan, a second relevance screening is performed in the literature database to obtain a number of documents to form a feasibility literature reference database; the keyword generation agent combines each document in the feasibility literature reference database to generate i document keywords respectively; a document exchange operation and a mutation operation are performed on the document keywords to obtain a final document keyword combination, and then the feasibility of the initial scientific research plan is enhanced by the refinement agent to generate a highly feasible scientific research plan;

[0012] S6. The review agent evaluates the highly feasible scientific research plan and provides improvement suggestions and innovation measurement. If the highly feasible scientific research plan needs to be optimized and improved, the highly feasible scientific research plan and the improvement suggestions will be sent back to the optimization agent for scientific research plan optimization and update, and then the review agent will evaluate it again until no improvement is needed. If no improvement is needed, the plan will be output as the final scientific research planning plan.

[0013] Furthermore, in S1, specifically: the series of keywords includes the combined core keywords and alternative keywords.

[0014] Furthermore, in S2, the generation of relevant expert agents based on the author database includes:

[0015] First, we searched the author database based on a series of keywords to obtain an author list. Each author in the list includes the following indicators: author personal information, the author's published papers, paper citations, impact factors, and research directions;

[0016] Each author in the author list is scored using a comprehensive author scoring method, and the top N authors are selected as experts. The titles and abstracts of all their published articles are extracted as an expert knowledge base. N expert agents are constructed based on each expert's name, title, focus area, number of citations, and collaborator names, and the expert agents are introduced into a multi-agent interaction framework.

[0017] Furthermore, in S2, the integration of relevant literature information based on the literature database is specifically as follows:

[0018] After searching the literature database based on the series of keywords, a number of roughly related documents are obtained. The user agent performs a first relevance screening on the documents to obtain a number of finely related documents to form a related information database. The related information database is embedded into the expert knowledge base of each expert agent for reference;

[0019] The user interacts with multiple agents to discuss and obtain several discussion topics including:

[0020] Pass the user input to each expert agent, and then each expert agent responds to the user input;

[0021] After the reply is completed, the discussion topic agent generates several discussion topics based on the user input and the reply content of each expert agent.

[0022] Furthermore, the comprehensive scoring method is: combining the relevance score, author influence and activity to obtain a comprehensive score;

[0023] Among them, the relevance score is obtained by the similarity between all papers published by the author and the series of keywords; the author's influence is obtained by the number of paper citations, impact factor, and the number of conference or journal papers published in recent years; the activity is obtained by counting the total number of citations of the author's papers in recent years and the total number of papers in recent years.

[0024] The first relevance screening specifically includes: using the user agent to measure the relevance of the document title and document abstract content with the user input, and screening out highly relevant documents;

[0025] The expert agents respond to user input specifically as follows:

[0026] The memory bank of each expert agent stores the communication and reply content with the user in real time. When the communication reply exceeds N episode After N times, the contents of the memory bank are compressed and summarized, and the last N remain The communication reply is stored in the memory bank; the user input is then transmitted to the expert intelligent agent, which compares the user input with its own knowledge base and obtains relevant content with the help of the retrieval enhancement generation method. The relevant content is combined with the user input and the compressed summary content in the memory bank to give a final reply.

[0027] Furthermore, in S3, the interactive inspiration Spark stage is specifically as follows:

[0028] The questioning agent asks heuristic questions based on the global memory bank to the user or any expert agent based on the selected discussion topic. The questioned person responds, and the expert agent's response refers to its historical discussion content. Other expert agents then comment on and follow up on the questioned person's response, updating the global memory bank.

[0029] The interactive criticism stage is specifically as follows:

[0030] The critical agent criticizes the output content of each expert agent in the global memory library. The criticized expert agent will respond with reference to its historical discussion content and update the global memory library.

[0031] Furthermore, in S5, the second correlation screening is specifically as follows: the scientific research plan and all the documents in the literature database are respectively subjected to vector embedding processing using the all-MiniLM-L6-v2 model, and the correlation between the scientific research plan and each document is measured using cosine similarity; several documents with high correlation are selected as a feasibility literature reference database.

[0032] Furthermore, in S5, the keyword generation agent combines each document in the feasibility document reference database to generate document keywords. Specifically: the keyword generation agent is used to combine each document title and abstract in the feasibility document reference database to generate corresponding i document keywords for each document; then the all-MiniLM-L6-v2 model is used to perform vector embedding processing on all document keywords. In order to identify semantically similar document keywords, the classic K-means clustering algorithm is used to classify all document keywords to obtain K document keyword classes.

[0033] Furthermore, in S5, the document exchange operation and mutation operation are specifically as follows:

[0034] In the exchange operation, n documents are randomly selected from a number of documents in the feasible document reference database to obtain a set of i*n document keywords for the n documents, as well as the corresponding n document titles and abstracts; then, i document keywords are randomly extracted from these i*n document keywords, and the random extraction is repeated several times to generate several different sets of initial document keyword combinations;

[0035] In the mutation operation, for any n document keywords in each group of document keyword combinations, other document keywords are randomly selected from the document keyword class to which they belong, aligned and replaced to achieve a diverse combination of document keywords and obtain the final document keyword combination.

[0036] A user personalized scientific research planning scheme generation system based on multi-agent interaction is obtained by using any of the user personalized scientific research planning scheme generation methods based on multi-agent interaction.

[0037] The beneficial effects are:

[0038] It effectively analyzes user intent, intelligently deconstructs keywords, automatically builds multi-domain agents, and searches and extracts relevant information for interactive discussions and critiques that are both professional and cross-disciplinary. This provides users with a detailed discussion process and richer responses. Furthermore, the entire discussion process is effectively stored, forming a global memory visible to each agent. Each agent's reply can selectively incorporate historical discussion content, which helps continuously improve response quality and enables the system to dynamically adapt to complex and ever-changing scientific research interaction scenarios.

[0039] Ultimately, the research plan planning agent will refine specialized research plans from the global memory bank, iteratively improve the quality of the review and optimization agents, and refine the agents to enhance their feasibility, ultimately generating highly innovative and feasible research plans. This multi-agent interactive framework, with its tightly linked components, provides users with richer, more efficient, and more accurate research intelligent responses and plan generation, significantly enhancing its generalization capabilities and cross-disciplinary perspective, promoting cutting-edge exploration across interdisciplinary fields, and meeting the diverse needs of users. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 A flowchart of related document information search and expert agent establishment according to an embodiment of the present invention;

[0041] Figure 2 A schematic diagram of a multi-agent framework flow for generating innovative scientific research solutions according to an embodiment of the present invention. DETAILED DESCRIPTION

[0042] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. The embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0043] Example 1:

[0044] More specifically, based on Figure 1 and 2 , for detailed explanation.

[0045] First, the user input is obtained, and the user input is decomposed into keywords and the agent outputs a series of keywords;

[0046] The keyword disassembly agent prompt template is as follows:

[0047] You are a professional scientist. Based on the user's question,

[0048] Return a valid JSON in strict accordance with the format:

[0049] {"core": [...], "option": [...]}

[0050] rule:

[0051] 1.core: up to 3 key terms with high relevance

[0052] 2. Optional: Up to 5 additional terms (mechanism, method, context, etc.)

[0053] 3. Don’t include any explanations or additional keywords

[0054] User question:{input}

[0055] The keyword decomposition agent will output a json format text containing the core keyword key i and alternative keywords key ri ,After combining the core keywords and alternative keywords, searches were conducted in the author database and the literature database respectively.

[0056] Among them, after searching in the author database, a list of authors is obtained Each author a i All include: author's personal information (name, unit, etc.); author's published papers Indicators such as paper citations, H-index, and research direction.

[0057] The author's overall score can be given by the following formula:

[0058] Score(a i )=w rel ·Rel(a i )+w inf ·Inf(a i )+w rec ·Rec(a i )

[0059] where w rel ,w inf ,w rec These are the weights of the relevance, author influence, and activity indicators, which can be adjusted according to actual conditions. The meaning of each indicator is described below.

[0060] Relevance score Rel(ai ):To the author a i All papers published with the series keyword K * The similarity calculation is obtained

[0061]

[0062] Among them, Embed(x) means that the text x is vector-embedded using the all-MiniLM-L6-v2 model to form a vector v(x); Sim(v1,v2) means using cosine similarity to measure the correlation between two vectors v1 and v2. Specifically expressed as:

[0063]

[0064] Here v1·v2 is the dot product of vectors v1 and v2. ‖v1‖, ‖v2‖ are the Euclidean norms of vectors v1 and v2.

[0065] Author influenceInf(a i ):Citations(a i ), impact factor Hindex, number of papers published in journals or conferences in the past five years j , where j represents a journal. Here we only count the number of papers published in Nature, Science, Cell and their sub-journals.

[0066]

[0067] α1 is the influence weight of the number of paper citations, α2 is the Hindex weight of the author, and α3 is the influence weight of the number of publications in journals or conferences in the past five years;

[0068] Activity Rec(a i ): A collection of papers by the statistical author in the past five years The number of citations to the paper p in the collection (Cites(p), as well as the number of papers published by the author in the past five years Organize and calculate the total number of citations and the total number of papers published by the author in the past five years.

[0069]

[0070] Among them, β1 is the weight of the normalized term of the total number of citations of papers in the past five years, and β2 is the weight of the normalized term of the number of papers in the past five years.

[0071] Based on the comprehensive scores, we selected the top three experts and extracted information such as the titles and abstracts of all their published articles as their expert knowledge base. Based on the expert name, title, focus area, citations, and collaborators, we constructed three expert agents and established their prompt words as follows:

[0072] You are {name}, known for {title}. You specialize in {focus_area}. Your work has been cited in {citations}. Your collaborators include {collaborators}. Please provide users with expert analysis and recommendations from your perspective in this field.

[0073] The three set expert agents are introduced into the multi-agent interaction framework.

[0074] After searching the literature database, we obtain the literature roughly related to the keyword. The roughly related literature is given to the user agent for relevance screening. We measure the relevance of the literature title ref_title and the literature abstract abs_text to the user input, and the prompt words are established as follows:

[0075] User input: {input}

[0076] Reference title: {ref_title}

[0077] Abstract of reference: {abs_text}

[0078] Give me your thoughts and reasons, and then decide whether this document is relevant to the user input

[0079] The exact JSON format returned is: {"Decision":"<RELATED or UNRELATED> "}

[0080] Then give me your thoughts.

[0081] The selected relevant literature is integrated and embedded into the knowledge base of each expert agent for reference.

[0082] When entering the discussion stage between humans and multi-agents, all content information will be stored in the global memory.

[0083] Based on the user input, each expert agent answers in a random order.

[0084] The memory bank of each expert agent stores the communication and reply content with the user in real time. When the communication reply exceeds N episodeAfter N times, all the contents in the memory bank are compressed and summarized, and the last N remain Exchange reply.

[0085] The prompt words of the compressed summary are as follows:

[0086] If the memory compression summary has not been performed before, the prompt word is,

[0087] f"Please summarize the core content of this memory and extract the most critical part: {messages}"

[0088] If memory compression summary has been performed, the prompt word is,

[0089] f"This is the summary of the conversation so far:\n{summary_old}\n\nPlease update and extend the original condensed summary based on the following new direct conversations between you and the user:{messages}"

[0090] Here, messages is the communication reply content between the user and the agent, summary_old is the original memory compressed summary content. The new memory compressed summary content will be recorded as summary_new.

[0091] After compressing and summarizing the content in the memory bank, the user input is transmitted to the expert agent. The expert agent compares the user input with its own memory bank, uses retrieval-augmented generation (RAG) to find relevant content, and then combines the user input with the compressed summary in the memory bank to provide a final response.

[0092] After each expert agent responds in turn, the discussion topic agent will generate multiple discussion topics and establish the following prompt words for the discussion topic agent:

[0093] You are a professional meeting note-taker. Please collect scientific research points based on the following conversation as possible meeting discussion topics:

[0094] <convo>

[0095] {conversation}

[0096] < / convo>

[0097] And output in JSON format.

[0098] Here, conversation represents the set of user input and all expert agent responses. The output is all discussion topics, and the user can select one of them, which is denoted as theme.

[0099] Next, we enter the multi-agent interaction inspiration Spark stage. First, we randomly select an agent or user as the target, denoted as target_name. The questioning agent creates a prompt word based on the discussion topic information theme and the conversation_content of the target in the global memory:

[0100] You are a questioner, a professional host, and a scientific researcher with extremely innovative thinking.

[0101] We are currently discussing: {theme}

[0102] You are now facing {target_name} and asking questions based on his / her recent response: {conversation_content}

[0103] You can consider some valuable, innovative, and inspiring questions, or you can consider some special and specific small questions.

[0104] You may wish to elaborate further, but please be direct and friendly in your question. Speak in {target_name}'s language.

[0105] The questioner's question content will be recorded as question_content. If the questionee is a user, they are required to manually reply. If the questionee is an expert agent, the question content question_content is directly handed over to the expert agent for reply, and the reply content is recorded as responses_content. Another expert agent random_hero is randomly selected to comment and answer based on question_content, responses_content, and its own memory and knowledge base content. The corresponding prompt words are as follows:

[0106] The questioner's question is: '{question_content}',

[0107] {target_name}'s response is: '{responses_content}'.

[0108] {random_hero}, please answer using {target_name}'s words. What are your thoughts on {target_name}'s question and answer? Do you have any further perspective?

[0109] In the interactive criticism phase, a criticized expert agent is first randomly selected and recorded as target_nameC. The criticizing agent will create the following prompt words based on the entire conversation content of target_nameC in the global memory bank:

[0110] You are a professional scientific researcher.

[0111] We are currently discussing the following:\n{theme_details}\n

[0112] You are now facing {target_nameC}. You need to condemn him. The core of the condemnation should focus on the content of his previous chat:

[0113] {conversation_content}

[0114] You should criticize directly and harshly, pointing out loopholes and shortcomings.

[0115] Because he is likely exaggerating. Use the language {target_nameC} just used to strongly criticize him.

[0116] The output criticism content of the critical agent is recorded as criticism_content. This part of the critical content will be directly transmitted to the criticized expert agent for reply and obtain its response_content.

[0117] The contents of the above stages will be updated and stored in the global memory in real time.

[0118] The research plan planning agent will intelligently summarize the global memory to obtain a preliminary research plan and establish the following prompt words:

[0119] The user has conducted detailed discussions with multiple experts. Please collect multiple innovative scientific research solutions based on the following conversation content:

[0120]

[0121]

[0122] In order to further enhance the feasibility of the preliminary scientific research plan and focus on key scientific research issues, we take a preliminary scientific research plan as an example and organize the literature database:

[0123] At the same time, the scientific research plan and all the documents in the literature database are respectively vector-embedded using the all-MiniLM-L6-v2 model to obtain the scientific research plan vector v sci and each document vector v in the literature database i (i=1,2,…,P), where P represents the total number of documents. The cosine similarity is used to measure the relevance of the research plan with each document Sim(v sci ,v i ). Literature with a correlation higher than 0.7 and ranked in the top 10 were selected as the feasibility literature reference database.

[0124] Then, we use the keyword generation agent to combine the title and abstract of each document in the feasibility literature reference database to generate 5 corresponding keywords for each document. The corresponding prompt words are as follows:

[0125] Based on the following research topic and abstract, please generate no more than 5 relevant keywords, separated by commas:

[0126] Topic: {topic}

[0127] Abstract:

[0128] Please only input the 5 most valuable keywords, directly in English, separated by commas.

[0129] So we obtained 5 corresponding keywords for each of our 10 documents.

[0130] Next, we use the all-MiniLM-L6-v2 model to embed all keywords. To identify semantically similar keywords, we use the classic K-means clustering algorithm.

[0131] K-means is an unsupervised learning method based on iterative optimization. Its goal is to partition an input sample set into K non-overlapping subsets (clusters), ensuring that samples within each subset are highly similar, while samples across clusters are highly diverse. In this solution, a sample is each vectorized keyword, and its vector represents the semantic information of the keyword.

[0132] The first step of the K-means algorithm is to select the initial K cluster centers (i.e., centroids), denoted as μ1, μ2, …, μK. These centroids are vectors randomly selected from the entire dataset, and the vectors of K keywords are randomly selected as the initial centroids.

[0133] The second step of the K-means algorithm is the allocation phase. Each keyword vector x iwill be assigned to the cluster center μ with the closest Euclidean distance k . Its mathematical expression is:

[0134] dist(x i ,μ k )=‖x i -μ k ‖ 2

[0135] Among them, x i : represents the i-th keyword vector; μ k : The centroid of the current k-th cluster; ‖·‖: represents the Euclidean norm (L2 norm), which measures the distance between two vectors. The goal of this stage is to classify each keyword into a nearest cluster C k , thus forming a new cluster distribution.

[0136] The third step of the K-means algorithm is the update phase. After completing the assignment of keyword vectors, the centroid of each cluster needs to be updated. The new centroid μ k is the mean of all keyword vectors in the cluster, and is calculated as follows:

[0137]

[0138] Among them, C k : represents the current k-th cluster; |C k |: the number of data points (keyword vectors) in the cluster; x i ∈C k : The keyword vector belonging to the kth cluster. Through this step, the center of each cluster will move toward the average position of its internal samples.

[0139] K-means is an iterative algorithm that alternates between the allocation phase and the update phase until any of the following termination conditions is met:

[0140] 1. The position of the cluster center does not change between two consecutive iterations, or the change is less than the set threshold;

[0141] 2. Reach the preset maximum number of iterations.

[0142] At this point, the clustering process is considered to have converged, and the final cluster division results are output to obtain K keyword classes.

[0143] Next, perform document exchange and mutation operations.

[0144] In the swap operation, we randomly selected three documents from the 10 selected documents, obtained their 15 keyword sets, and the corresponding three document titles and abstracts. We then randomly extracted five keywords from these 15 keywords, performing five random extractions, to generate five different initial keyword combinations.

[0145] In the mutation operation, for any three keywords in each keyword combination, other keywords are randomly selected from the keyword class to which they belong, aligned and replaced to achieve a diverse combination of keywords and obtain the final keyword combination.

[0146] After the literature exchange and mutation operations, the five keywords in each final keyword combination are integrated with the corresponding three literature abstracts and the research plan content, and then handed over to the refinement agent for processing and optimization. The agent prompt words are as follows:

[0147] You are a research proposal optimization agent. Your task is to enhance the feasibility and innovation of the research proposal based on the research proposal text, keywords, and the titles and abstracts of three references I provide. During the optimization process, you must always focus on the keywords and comprehensively reference the core content and ideas of the three references to enhance the value and advancement of the research proposal in terms of method design, technical path, and theoretical foundation.

[0148] The specific requirements are as follows:

[0149] 1. Input content:

[0150] -Research proposal title: {idea_description}

[0151] -Research proposal details: {idea_details}

[0152] -Keywords:{keywords}

[0153] -References: {references}

[0154] 2. Optimization objectives:

[0155] - The “Research Background” section should include the latest developments in the current field and highlight the key challenges and research gaps.

[0156] -Use keywords as core clues, analyze and explain how they should be fully utilized in scientific research plans to demonstrate the feasibility and innovation of the plans.

[0157] - Refer to the ideas or methods of the three papers and organically integrate the valuable ideas into the "Problem Statement", "Research Motivation" and "Proposed Methods" modules of the scientific research plan, highlighting the areas for improvement.

[0158] - The advantages of the proposed solution compared with existing methods, and the innovative value of the proposed technology or model at the theoretical and practical levels.

[0159] 3. Output format:

[0160] Please strictly follow the following JSON array format to output the optimized research plan content. Do not output any additional text or comments to ensure that the JSON syntax is legal.

[0161]

[0162]

[0163] The resulting innovative scientific research solutions will be handed over to the review agent for review and evaluation of optimization directions. The prompts for the review agent are as follows:

[0164] We are planning to publish in a top journal. Please provide a detailed review of the following research proposal from the perspective of a professional reviewer:

[0165] Research proposal title: {idea_description}

[0166] Research proposal details:{idea_details}

[0167] Please provide feedback on the following:

[0168] 1. How to make the practical problems and innovations faced by the idea clearer and more specific?

[0169] 2. Further clarify the experimental plan to make it more specific and feasible?

[0170] 3. Are there any assumptions that need further verification or loopholes in the experimental design? How can they be adjusted?

[0171] Please give your thoughts first, then your responses. Follow the format below:

[0172] THOUGHT:

[0173] <thought>

[0174] RESPONSE:

[0175] ```json

[0176] <json>

[0177] ```

[0178] exist <thought>In the example above, first briefly reason about the idea and identify any doubts that might help you make a decision.

[0179] If you've made a decision, add "Decision made: Solution complete." or "Decision made:

[0180] The plan is incomplete." Then give the reasons why.

[0181] If the review agent deems the solution incomplete, it will pass the review feedback content evaluation_response and the corresponding scientific research solution science_idea to the optimization agent for optimization and update, using the following prompt words:

[0182] The original research plan is: {science_idea}

[0183] After review by the evaluation experts, the feedback is as follows: {evaluation_response}

[0184] Please give your thoughts first, and then give the updated research plan in JSON format.

[0185] The format is as follows:

[0186]

[0187]

[0188] When the review agent gives a decision that the plan is complete, the scientific research plan optimization stops, and the final scientific research plan is output, which is the user-personalized scientific research planning plan based on multi-agent interaction.

[0189] The above embodiments are intended only to illustrate the design concepts and features of the present invention. Their purpose is to enable those skilled in the art to understand the contents of the present invention and implement them accordingly. The scope of protection of the present invention is not limited to the above embodiments. Therefore, any equivalent changes or modifications made based on the principles and design concepts disclosed in the present invention are within the scope of protection of the present invention.

[0190] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art could make various modifications and improvements without departing from the spirit of the present application, all of which fall within the scope of protection of the present application. Therefore, the scope of protection of the present patent application shall be determined by the appended claims.< / thought> < / json> < / thought>

Claims

1. A method for generating user-personalized scientific research planning schemes based on multi-agent interaction, characterized in that: The following steps are involved: S1, receiving user input questions, breaking down the intelligent agent into keywords and outputting a series of keywords; S2: Based on user input and a series of keywords, relevant expert agents are generated based on the author database, relevant literature information is integrated based on the literature database, and users and multiple agents interact in discussions to obtain several discussion topics. All discussion content information will be stored in the global memory bank; S3: The user selects a discussion topic and enters the interactive inspiration Spark stage and the interactive criticism Criticism stage. During this process, the global memory library is updated in real time. S4, the research plan planning agent generates an initial research plan based on the global memory library and the discussion topic selected by the user; S5. Based on the initial scientific research plan, a second relevance screening is performed in the literature database to obtain a number of documents to form a feasibility literature reference database; the keyword generation agent combines each document in the feasibility literature reference database to generate i document keywords respectively; a document exchange operation and a mutation operation are performed on the document keywords to obtain a final document keyword combination, and then the feasibility of the initial scientific research plan is enhanced by the refinement agent to generate a highly feasible scientific research plan; S6. The review agent evaluates the highly feasible scientific research plan and provides improvement suggestions and innovation measurement. If the highly feasible scientific research plan needs to be optimized and improved, the highly feasible scientific research plan and the improvement suggestions will be sent back to the optimization agent for scientific research plan optimization and update, and then the review agent will evaluate it again until no improvement is needed. If no improvement is needed, the plan will be output as the final scientific research planning plan.

2. The method for generating a user-personalized scientific research plan based on multi-agent interaction according to claim 1, characterized in that: In S1, specifically: the series of keywords include the combined core keywords and alternative keywords.

3. The method for generating a user-personalized scientific research plan based on multi-agent interaction according to claim 1, characterized in that: In S2, the generation of relevant expert agents based on the author database includes: First, we searched the author database based on a series of keywords to obtain an author list. Each author in the list includes the following indicators: author personal information, the author's published papers, paper citations, impact factors, and research directions; Each author in the author list is scored using a comprehensive author scoring method, and the top N authors are selected as experts. The titles and abstracts of all their published articles are extracted as an expert knowledge base. N expert agents are constructed based on each expert's name, title, focus area, number of citations, and collaborator names, and the expert agents are introduced into a multi-agent interaction framework.

4. The method for generating a user-personalized scientific research plan based on multi-agent interaction according to claim 1, characterized in that: In S2, the integration of relevant literature information based on the literature database is specifically as follows: After searching the literature database based on the series of keywords, a number of roughly related documents are obtained. The user agent performs a first relevance screening on the documents to obtain a number of finely related documents to form a related information database. The related information database is embedded into the expert knowledge base of each expert agent for reference; The user interacts with multiple agents to discuss and obtain several discussion topics including: Pass the user input to each expert agent, and then each expert agent responds to the user input; After the reply is completed, the discussion topic agent generates several discussion topics based on the user input and the reply content of each expert agent.

5. The method for generating a user-personalized scientific research plan based on multi-agent interaction according to claim 3, characterized in that: The comprehensive scoring method is: combining the relevance score, author influence and activity to obtain a comprehensive score; The relevance score is obtained by the similarity between all papers published by the author and the series of keywords; the author's influence is obtained by the number of paper citations, impact factor, and the number of conference or journal papers published in recent years; the activity is obtained by counting the total number of citations of the author's papers in recent years and the total number of papers in recent years; The first relevance screening specifically includes: using the user agent to measure the relevance of the document title and document abstract content with the user input, and screening out highly relevant documents; The expert agents respond to user input specifically as follows: The memory bank of each expert agent stores the communication and reply content with the user in real time. When the communication reply exceeds N episode After N times, the contents of the memory bank are compressed and summarized, and the last N remain The communication reply is stored in the memory bank; the user input is then transmitted to the expert intelligent agent, which compares the user input with its own knowledge base and obtains relevant content with the help of the retrieval enhancement generation method. The relevant content is combined with the user input and the compressed summary content in the memory bank to give a final reply.

6. The method for generating a user-personalized scientific research plan based on multi-agent interaction according to claim 1, characterized in that: In S3, the interactive inspiration Spark stage is specifically as follows: The questioning agent asks heuristic questions based on the global memory bank to the user or any expert agent based on the selected discussion topic. The questioned person responds, and the expert agent's response refers to its historical discussion content. Other expert agents then comment on and follow up on the questioned person's response, updating the global memory bank. The interactive criticism stage is specifically as follows: The critical agent criticizes the output content of each expert agent in the global memory library. The criticized expert agent will respond with reference to its historical discussion content and update the global memory library.

7. The method for generating a user-personalized scientific research plan based on multi-agent interaction according to claim 1, characterized in that: In S5, the second correlation screening is specifically as follows: the scientific research plan and all the documents in the literature database are vector-embedded using the all-MiniLM-L6-v2 model, and the correlation between the scientific research plan and each document is measured using cosine similarity; several documents with high correlation are selected as a feasibility literature reference database.

8. The method for generating a user-personalized scientific research plan based on multi-agent interaction according to claim 1, characterized in that: In S5, the keyword generation agent generates document keywords in combination with each document in the feasibility document reference database. Specifically, the keyword generation agent is used to generate corresponding i document keywords for each document in combination with each document title and abstract in the feasibility document reference database; then the all-MiniLM-L6-v2 model is used to perform vector embedding processing on all document keywords. In order to identify semantically similar document keywords, the classic K-means clustering algorithm is used to classify all document keywords to obtain K document keyword classes.

9. The method for generating a user-personalized scientific research plan based on multi-agent interaction according to claim 8, characterized in that: In S5, the document exchange operation and mutation operation are specifically as follows: In the exchange operation, n documents are randomly selected from a number of documents in the feasible document reference database to obtain a set of i*n document keywords for the n documents, as well as the corresponding n document titles and abstracts; then, i document keywords are randomly extracted from these i*n document keywords, and the random extraction is repeated several times to generate several different sets of initial document keyword combinations; In the mutation operation, for any n document keywords in each group of document keyword combinations, other document keywords are randomly selected from the document keyword class to which they belong, aligned and replaced to achieve a diverse combination of document keywords and obtain the final document keyword combination.

10. A user-personalized scientific research planning scheme generation system based on multi-agent interaction, characterized in that: It is obtained by adopting the user personalized scientific research planning scheme generation method based on multi-agent interaction as described in any one of claims 1-9.

Images