A scientific research question and answer reasoning all-in-one machine based on multi-agent cooperation and a control method

Abstract

The application provides a scientific research question and answer reasoning all-in-one machine based on multi-agent cooperation and a control method, relates to the technical field of artificial intelligence and multi-agent reasoning, and comprises a first agent, which is used for performing semantic understanding on a scientific research question input by a user to obtain a semantic understanding result; performing semantic expansion on the semantic understanding result according to configured expansion rules to obtain a first query request corresponding to the scientific research question; a second agent, which is used for screening answer information associated with the query request sent by the first agent from a pre-constructed scientific research knowledge base and sending the corresponding answer information to the first agent; and a third agent, which is used for evaluating whether target answer information and corresponding semantic understanding result, first query request and second query request satisfy configured evaluation rules to obtain an evaluation result. The application realizes accurate, comprehensive and reliable answering of scientific research questions of users.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence and multi-agent reasoning technology, specifically, a scientific research question and answer reasoning all-in-one machine based on multi-agent cooperation and a control method. BACKGROUND

[0002] The existing question and answer platform or large model mainly has the following problems:

[0003] First, data sparsity and lack of professionalism: existing technologies usually rely on large-scale pre-training corpus, but the data coverage of professional fields (such as microbial natural products, drug research and development, biochemistry, etc.) is still insufficient, which limits the understanding of professional terms, complex mechanisms or edge knowledge of the model, and limits the professional reasoning ability.

[0004] Second, scientific research question and answer requires in-depth understanding of professional terms and specific concepts, and the current model often deviates or even misunderstands the semantics when facing relatively rare or highly professional expressions; for complex problems in the field of biochemistry, single round retrieval often cannot cover the required global knowledge. Professional problem solving usually requires multi-level, multi-dimensional knowledge association and reasoning.

[0005] Third, the type of information is single, and the cross-modal fusion is insufficient: most existing question and answer systems mainly take text as input and output, and it is difficult to efficiently process structured data, experimental atlas, spectral graph or gene sequence and other data forms; for scientific research problems, especially involving multi-level and multi-dimensional information, a single mode of information is often insufficient to provide comprehensive and accurate answers, and the system needs to be broken through in integrating multi-modal data.

[0006] Fourth, the task flow is not fully split, and the logical reasoning and causal relationship recognition are insufficient: facing problems in the scientific research field with multi-step or hierarchical structure, existing technologies often have difficulty in effectively splitting complex tasks into manageable subtasks, resulting in an incomplete overall solution process; scientific research problems usually require fine reasoning and causal analysis in details, and the lack of logical reasoning ability in existing systems may lead to errors or omissions in the solution process.

[0007] Fifth, there is a high risk of hallucination and false information, and real-time feedback and dynamic correction are insufficient: due to the lack of effective self-evaluation mechanism, existing technologies are prone to produce "hallucination" output that does not conform to the actual biological or chemical facts in the absence of auxiliary professional tools, and lack sufficient output monitoring mechanism, which cannot timely alarm when errors are found, and continuously update and correct related information through feedback mechanism, thereby reducing the risk of misuse of false information. SUMMARY

[0008] The purpose of this application is to provide a scientific research question-answering reasoning integrated machine and control method based on multi-agent collaboration, so as to solve the above-mentioned problems existing in the prior art. It can understand natural language tasks and complete complex multimodal scientific research question reasoning and answering through dialogue.

[0009] In a first aspect, the present invention provides a scientific research question-answering and reasoning integrated machine based on multi-agent collaboration, the integrated machine comprising:

[0010] A first intelligent agent performs semantic understanding on a research question input by a user to obtain a semantic understanding result; performs semantic expansion on the semantic understanding result according to configured expansion rules to obtain a first query request corresponding to the research question, and sends the first query request to a second intelligent agent; and receives a first response information sent by the second intelligent agent in response to the first query request; if the richness of the first response information does not meet a configured richness threshold, generates a second query request based on the first response information and the semantic understanding result, and sends the second query request to the second intelligent agent; and receives a second response information sent by the second intelligent agent in response to the second query request; if the lexical richness of the second response information meets a configured lexical richness threshold, the second response information is used as the target response information for the research question; generates an evaluation request and sends the evaluation request to a third intelligent agent; and receives an evaluation result sent by the third intelligent agent in response to the evaluation request; wherein the evaluation request includes: the semantic understanding result, the first query request, the second query request, and the target response information; if the evaluation result meets the configured evaluation rules, the target response information is output.

[0011] The second intelligent agent is used to filter response information associated with the query request sent by the first intelligent agent from a pre-built scientific research knowledge base, and send the corresponding response information to the first intelligent agent.

[0012] A third intelligent agent is used to evaluate whether the target response information and the corresponding semantic understanding results, the first query request and the second query request meet the configured evaluation rules, and obtain the evaluation results.

[0013] In an optional implementation, the all-in-one machine further includes: a knowledge base construction module and an interactive display device;

[0014] The interactive display device is used for the user to input the research question and to display the target response information output by the first intelligent agent;

[0015] The knowledge base construction module is specifically used for:

[0016] Obtaining multi-source scientific research data; wherein, the scientific research data comprises: each compound, comprehensive characteristics of different compounds, and scientific research literature containing any compound;

[0017] Based on the comprehensive characteristics of each compound and different compounds, a first database is constructed;

[0018] Based on the abstract of the scientific research literature containing each compound, an abstract vector of each scientific research literature is generated; based on the abstract vector of each scientific research literature and each scientific research literature, a second database is constructed;

[0019] From each scientific research literature, an image associated with any compound and an image description of the corresponding image are extracted, and a third database is constructed;

[0020] Taking each compound as a unique identifier, the comprehensive characteristics of each compound, the abstract vector of each scientific research literature and each image description are associated, and a three-dimensional knowledge graph is constructed;

[0021] Based on the first database, the second database, the third database, the three-dimensional knowledge graph and the pre-configured tool library, a scientific research knowledge base is generated.

[0022] In an optional implementation, the first agent is further configured to:

[0023] Obtaining the form of the scientific research question input by the user; wherein, the form of the scientific research question comprises: text form, image form and mixed form of text and image;

[0024] Matching the target pre-processing means corresponding to the form from the configured different forms and different pre-processing means configuration table;

[0025] Using the target pre-processing means to pre-process the scientific research question to obtain the pre-processed scientific research question.

[0026] In an optional implementation, the first agent is specifically configured to:

[0027] Performing semantic understanding on the pre-processed scientific research question to obtain a semantic understanding result; wherein, the semantic understanding result comprises: a to-be-queried entity contained in the scientific research question, a keyword, a first query intention corresponding to the scientific research question, and a semantic relationship between the to-be-queried entity and the query intention;

[0028] According to the semantic relationship, the first query intention is split to obtain a plurality of second query intentions;

[0029] According to the plurality of second query intentions and the to-be-queried entity, a plurality of first query questions are generated;

[0030] Filtering, from the semantic understanding result, the first query question, the second query question and the third query question to obtain at least one target query question;

[0031] Generating, according to the association relationship list and the to-be-queried entity, an expansion list;

[0032] Expanding, according to the expansion list, a plurality of first query questions to obtain a plurality of second query questions;

[0033] According to the configured chain transformation rule, the first query question and the second query question are transformed and adjusted to obtain a plurality of third query questions;

[0034] According to the semantic understanding result, the first query question, the second query question and the third query question are filtered to obtain at least one target query question;

[0035] According to the obtained target query question, a first query request corresponding to the scientific research question is generated.

[0036] In an optional implementation manner, the first intelligent agent is further configured to:

[0037] Statistically count the number of words contained in the received response information;

[0038] Based on the number of words counted, the richness of the response information is determined.

[0039] In an optional implementation manner, the second intelligent agent is specifically configured to:

[0040] From the three-dimensional knowledge graph, filter the abstract vector, compound, comprehensive property and image description associated with the query request sent by the first intelligent agent;

[0041] Obtain, by using the tool library, the association data associated with the query request;

[0042] Determine the scientific research literature corresponding to the abstract vector as a target scientific research literature;

[0043] Search, from the target scientific research literature, a text block associated with the query request;

[0044] Calculate the cosine similarity between each text block;

[0045] According to the reference relationship of the target scientific research literature corresponding to each text block, determine the reference weight corresponding to each text block;

[0046] According to the cosine similarity between the text blocks and the corresponding reference weight, the text blocks are sorted;

[0047] According to the first screening rule configured, target text blocks are screened from the sorted text blocks;

[0048] According to the second screening rule configured, the compound, the comprehensive characteristics, the associated data and the image are screened respectively to obtain target compound, target comprehensive characteristics, target associated data and target image;

[0049] Based on the target text blocks, the target compound, target comprehensive characteristics, target associated data and target image, response information associated with the query request is generated.

[0050] In an optional implementation, the third agent is further configured to:

[0051] If the evaluation result does not satisfy the configured evaluation rule, a configured response deviation prompt is generated, and the response deviation prompt is sent to the first agent;

[0052] The first agent is further configured to:

[0053] In response to the received response deviation prompt, based on the response deviation prompt and the corresponding query request and response information, a new query request is generated.

[0054] In a second aspect, the present application provides a scientific research question and answer reasoning control method based on multi-agent cooperation, which comprises:

[0055] Obtaining a scientific research question input by a user;

[0056] Controlling the first agent to perform semantic understanding on the scientific research question to obtain a semantic understanding result; according to a configured expansion rule, the semantic understanding result is expanded to obtain a first query request corresponding to the scientific research question, and the first query request is sent to the second agent;

[0057] Controlling the second agent to screen first response information associated with the first query request from a pre-constructed scientific research knowledge base, and sending the first response information to the first agent;

[0058] Controlling the first agent to receive the first response information; if the richness of the first response information does not satisfy a configured richness threshold, a second query request is generated according to the first response information and the semantic understanding result, and the second query request is sent to the second agent;

[0059] Controlling the second agent to screen second response information associated with the second query request from the pre-constructed scientific research knowledge base, and sending the second response information to the first agent;

[0060] The first intelligent agent receives the second response information, and if a lexical richness of the second response information meets a configured lexical richness threshold, the second response information is taken as target response information of the scientific research question; an evaluation request is generated and sent to a third intelligent agent; the evaluation request includes the semantic understanding result, the first query request, the second query request, and the target response information.

[0061] The third intelligent agent evaluates whether the target response information and the corresponding semantic understanding result, the first query request, and the second query request meet a configured evaluation rule, and obtains an evaluation result.

[0062] The first intelligent agent receives the evaluation result, and if the evaluation result meets the configured evaluation rule, the target response information is output.

[0063] In a third aspect, the present application provides an electronic device, which comprises a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory complete mutual communication through the communication bus.

[0064] The memory is used to store a computer program.

[0065] The processor is used to execute the program stored on the memory, and realize the method of any one of the preceding embodiments.

[0066] In a fourth aspect, the present application provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to realize the method of any one of the preceding embodiments.

[0067] The present application establishes a professional literature vector library of a specific field, uses retrieval augmented generation (RAG) technology, and when answering a scientific research question, a model obtains accurate and traceable information support by retrieving vector representations of related field literatures, so as to improve the professionalism and credibility of the answer.

[0068] The present application constructs a multi-agent framework to ensure semantic consistency of question input and response; the first intelligent agent is responsible for semantic analysis and professional expression of the input question of the user, and the second intelligent agent is responsible for generating an answer based on retrieved domain knowledge. Through this multi-agent collaborative mechanism, semantic consistency from the input question to the generated response is ensured, so as to effectively avoid information bias and improve the professional quality of the answer.

[0069] The application designs an evaluation benchmark dataset for biochemistry, covering five key research scenarios in the field of natural products. To minimize human intervention, the application utilizes expert and large language models to evaluate the results for reasonableness. The experimental results show that the application significantly outperforms other test models in terms of performance and exhibits superior answer diversity in certain cases, highlighting its potential as an intelligent assistant for natural product research.

[0070] The application creates an external database construction paradigm, expanding the application capabilities of large language models in the fields of biology and chemistry, thereby improving data processing efficiency and scientific accuracy. It proposes a multi-agent framework to promote interdisciplinary integration between biology and chemistry, supporting collaborative efforts in these two scientific fields, particularly for applications such as natural product research. A set of sound evaluation criteria is established to evaluate the performance of the integrated machine, thereby verifying the effectiveness of the entire framework. The evaluation criteria of the application can be transferred to any related agent evaluation task. Through the multi-agent collaboration framework, the application endows the model with self-evaluation capabilities, enabling it to determine whether the current knowledge reserve is sufficient to solve the problem and trigger a more in-depth analysis process when necessary.

[0071] The application can adaptively allocate computing resources according to the complexity of the problem, perform multi-round progressive retrieval on key problems, and ensure the comprehensiveness and accuracy of knowledge coverage. The application establishes a strict information traceability process to ensure that all output content can be traced back to authoritative sources, effectively reducing the illusion risk and improving the credibility of the scientific research assistance system.

[0072] The application constructs various forms of vector databases based on the data characteristics of the biochemistry field to ensure the accuracy of retrieval and provide multi-modal output. Each LLM is processed differently to become an agent that performs its own functions. The application uses the thinking chain technology to chain and expand the problem, better outputting the expected response information of the user, and the response information is more comprehensive.

[0073] The first agent of the application uses the thinking chain technology and reinforcement learning to understand and reason the scientific research problems input by the user. The second agent retrieves or calls the corresponding scientific research tools based on the understanding and reasoning results of the first agent, returns the retrieval results to the first agent to generate a new analysis report, and returns it to the third agent. The third agent is used to trace and evaluate the retrieval results, thereby ensuring accurate and comprehensive answers to the user's scientific research questions and improving the factual reliability and practical application value of the answer content. BRIEF DESCRIPTION OF DRAWINGS

[0074] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments of the present application. It should be understood that the following drawings only show some embodiments of the present application, and therefore should not be regarded as a limitation on the scope, and for those skilled in the art, other related drawings can also be obtained without creative labor on the basis of these drawings.

[0075] Figure 1 The structural diagram of the scientific research question and answer reasoning all-in-one machine based on multi-agent cooperation provided by the embodiments of the present application;

[0076] Figure 2 The structural diagram of the scientific research question and answer reasoning all-in-one machine based on multi-agent cooperation provided by the embodiments of the present application;

[0077] Figure 3 The flowchart of the scientific research question and answer reasoning control method based on multi-agent cooperation provided by the embodiments of the present application. DETAILED DESCRIPTION

[0078] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, and not all the embodiments. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0079] As shown in the figure, Figure 1 The scientific research question and answer reasoning all-in-one machine based on multi-agent cooperation provided by the embodiments of the present application is applied to scientific research question and answer in professional fields (such as microorganism natural products, drug research and development, biology, chemistry, etc.); as shown in the figure, Figure 2 The scientific research question and answer reasoning all-in-one machine based on multi-agent cooperation can include: an interactive display device, a first agent for question and answer, reading comprehension and reasoning, a second agent for retrieval and calling, a third agent for tracing and evaluation, a scientific research knowledge base and a knowledge base construction module; the first agent is in communication connection with the interactive display device, the second agent and the third agent respectively; the second agent is also in communication connection with the third agent and the scientific research knowledge base; the scientific research knowledge base is also in communication connection with the knowledge base construction module;

[0080] Specifically,

[0081] The interactive display device is used for user to input scientific research questions, and display target answer information output by the first agent;

[0082] The knowledge base construction module is used for constructing the scientific research knowledge base containing a first database, a second database, a third database, a three-dimensional knowledge graph and a pre-configured tool library.

[0083] A scientific knowledge base configured to store corresponding scientific data, a three-dimensional knowledge graph, or various tool calling interfaces through a first database, a second database, a third database, the three-dimensional knowledge graph, and a tool library;

[0084] A first intelligent agent configured to perform semantic understanding on a scientific question input by a user to obtain a semantic understanding result, perform semantic expansion on the semantic understanding result according to a configured expansion rule to obtain a first query request corresponding to the scientific question, and send the first query request to a second intelligent agent, receive first response information sent by the second intelligent agent in response to the first query request, generate a second query request according to the first response information and the semantic understanding result if the richness of the first response information does not satisfy a configured richness threshold, and send the second query request to the second intelligent agent, receive second response information sent by the second intelligent agent in response to the second query request, and if the lexical richness of the second response information satisfies a configured lexical richness threshold, take the second response information as target response information of the scientific question, generate an evaluation request and send the evaluation request to a third intelligent agent, receive an evaluation result sent by the third intelligent agent in response to the evaluation request, and if the evaluation result satisfies a configured evaluation rule, send the target response information to a display module.

[0085] A second intelligent agent configured to filter response information associated with the query request sent by the first intelligent agent from a pre-constructed scientific knowledge base, and send the corresponding response information to the first intelligent agent.

[0086] A third intelligent agent configured to evaluate whether the target response information and the corresponding semantic understanding result, the first query request, and the second query request satisfy a configured evaluation rule to obtain an evaluation result.

[0087] The first agent, the second agent, and the third agent can be carried in a processor or a server, or can be carried in a terminal with strong computing power. The processor can be two Nvidia A6000 48G; the server can be a physical server, or a server cluster or a distributed system formed by multiple physical servers, or a cloud server providing basic cloud computing services such as cloud service, cloud database, cloud computing, cloud function, cloud storage, network service, cloud communication, middleware service, domain name service, security service, content delivery network (CDN), and big data and artificial intelligence platform. The terminal can be a user equipment (UE) such as a mobile phone, a smart phone, a notebook computer, a digital broadcast receiver, a personal digital assistant (PDA), a tablet computer (PAD), a handheld device, a vehicle-mounted device, a wearable device, a computing device, or other processing devices connected to a wireless modem, a mobile station (MS), a mobile terminal (Mobile Terminal), etc. The terminal and the server can be directly or indirectly connected through wired or wireless communication, which is not limited in the present application.

[0088] In the embodiment of the present application, the knowledge base construction module is specifically configured to:

[0089] obtain multi-source scientific research data; construct a first database based on the comprehensive characteristics of each compound and different compounds; generate an abstract vector of each scientific research literature based on the abstract of the scientific research literature containing each compound; construct a second database based on the abstract vector of each scientific research literature and each scientific research literature; extract images associated with any compound and image descriptions of the corresponding images from each scientific research literature to construct a third database; associate the comprehensive characteristics of each compound, the abstract vector of each scientific research literature, and each image description with each compound as a unique identifier to construct a three-dimensional knowledge graph; and generate a scientific research knowledge base based on the first database, the second database, the third database, the three-dimensional knowledge graph, and a pre-configured tool library.

[0090] In the embodiment of the present application, the scientific research data includes each compound, the comprehensive characteristics of different compounds, and scientific research literature containing any compound. Specifically, the compound includes microbial secondary metabolites and chemical compounds; the chemical compound includes type II polyketide compounds; the comprehensive characteristics of microbial secondary metabolites include gene cluster data, chemical structure (SMILES), biological activity, gene sequence (FASTA), and functional domain prediction results; the gene cluster data includes gene cluster ID, species source, gene composition, and functional annotation.

[0091] In the embodiments of the present application, the comprehensive characteristics of each compound and different compounds are derived from the structured data in MIBiG (Microbial Secondary Metabolites Database), PubChem (Chemical Compound Database) and other professional databases, as well as the text data of literature mining (Web of Science) and expert annotation information.

[0092] In the embodiments of the present application, based on the comprehensive characteristics of each compound and different compounds, a first database is constructed, including:

[0093] The multi-source scientific research data is normalized, and an article table articles is generated according to each processed scientific literature; a product table products is generated according to the compound name, chemical structure and biological activity of each processed type II polyketide compound; a gene cluster table BGCs is generated according to the gene cluster ID, species source and functional annotation of each processed compound; a gene table genes is generated according to the gene sequence and functional domain prediction result of each processed compound; a literature-product association table article_product is generated by processing the synthesis / characterization association between the processed literature and each compound through a many-to-many relationship mapping; a gene cluster-product association table BGC_product is generated by establishing the biosynthesis path mapping of the gene cluster and its synthetic product; the chemical research (articles), gene function (genes) and synthetic biology data (BGCs) in the scientific literature are associated through the product table as a hub; the article table, the product table, the gene cluster table and the gene table are used as entity tables; and the first database is generated according to the association data of the entity tables and the gene cluster-product association table, the literature-product association table and the product table.

[0094] In the embodiments of the present application, the article table contains the metadata and abstract of each scientific literature; each entity table in the first database stores each compound entity; and each entity table in the first database is provided with a foreign key constraint (such as product_id) to ensure the consistency of cross-references of chemical entities and biological entities; the present application establishes the biosynthesis relationship of “gene cluster-compound” through the BGC_product table to associate microbiology and natural product chemistry data; and the synthesis method (chemistry) in the literature and the product activity data (pharmacology) are cross-linked through the article_product table to realize the fusion of multi-disciplinary scientific research data.

[0095] In the embodiments of the present application, the normalization processing includes: normalizing the name of each compound; structuring the gene cluster data; extracting the metadata of the scientific literature through a regular expression to separate and store the metadata and experimental description content in the scientific literature; and performing block processing on each scientific literature to obtain multiple text blocks of each scientific literature; the metadata includes: author, title, journal, year and digital object unique identifier DOI.

[0096] In the embodiment of the present application, since the compound names of PubChem (a chemical compound database), InChIKey and the compound names stored in other professional databases may be different, the names of each compound need to be normalized mapping to eliminate the naming differences.

[0097] In the embodiment of the present application, the gene cluster data is structured and processed, including extracting the genomic composition, functional annotation and other fields of BGC (biosynthetic gene cluster) from MIBiG, and establishing the association with the product.

[0098] In the embodiment of the present application, based on the abstracts of the scientific research literatures containing each compound, the abstract vectors of each scientific research literature are generated, including:

[0099] The abstracts of each scientific research literature are vectorized using the field-adapted embedding model embedding-3-small to obtain the abstract vectors of each scientific research literature.

[0100] In the embodiment of the present application, each scientific research literature is processed by blocking, including: using pdfminer to extract the pure text of each scientific research literature, and eliminating the header and / or footer and / or watermark noise; using a hybrid blocking strategy (fixed length + overlapping window) to process each scientific research literature by blocking to balance the context coherence and retrieval granularity; using the field-adapted embedding model embedding-3-small to vectorize the blocked text blocks to retain semantic features. The present application balances the text semantic integrity (fixed blocking) and context overlap requirements (sliding window) to improve the fragment retrieval accuracy of long literatures (such as reviews).

[0101] In the embodiment of the present application, the image descriptions of each image are obtained by manually labeling the experimental flowchart and chemical structure chart in the scientific research literature, storing the picture URL and standardized description.

[0102] In the embodiment of the present application, the third database further includes: the structural formula images of each compound generated in real time by tool calling (such as chemical drawing software) to supplement the missing data.

[0103] In the embodiments of the present application, the tool library includes: a functional annotation tool, a PubChem interface tool, and a multiple sequence alignment tool, a bioinformatics tool, a scientific mapping statistical tool, and a computational chemistry tool; wherein the functional annotation tool is used to align the gene sequence with the type II polyketide gene database, and filter the conserved domains (such as KR and ACP) through the Bit-score threshold; the PubChem interface tool is used to retrieve physicochemical property data according to the product name, and supplement the chemical context; the multiple sequence alignment tool is used to identify the evolutionary relationship of the gene cluster through the MAFFT algorithm, and support the generation of synthetic biology modification suggestions; the bioinformatics tool can include BLAST, DIAMOND, SPADES, and HISAT2, etc.; the scientific mapping statistical tool can include R and python, etc.; and the computational chemistry tool can include Gromacs and Gaussian.

[0104] In the embodiments of the present application, the three-dimensional knowledge graph realizes cross-referencing of relational data (structured attributes, i.e., compounds and comprehensive characteristics in the first database), vector data (semantics, i.e., abstract vectors), and image data (vision) through unique identifiers (such as DOI and compound ID).

[0105] In the embodiments of the present application, the first intelligent agent is a mentor Mentor, the second intelligent agent is an executor Executor, and the third intelligent agent is an evaluator Evaluator; the mentor Mentor and the executor Executor perform t rounds of interaction. In each round, the mentor Mentor submits a query request to the executor Executor, the executor Executor processes the query and retrieves relevant context information from various knowledge bases; the information can include text, images, references, and other supporting materials to help address the current query. Based on the user's original intention and the results returned by the executor Executor in the previous rounds, the mentor Mentor formulates the next query. After completing t rounds of interaction, the mentor Mentor summarizes the questions and answers generated in the entire interaction process to form a final response (i.e., response information) and returns it to the user. Evaluation process: the evaluator Evaluator interacts with the executor Executor during the entire process of retrieving and generating answers (before, during, and after) to improve the quality of the answers. Error reporting: if the evaluator Evaluator finds abnormal conditions during the retrieval process of the executor Executor, the evaluator Evaluator will feed back these problems to the mentor Mentor.

[0106] In the embodiments of the present application, the first intelligent agent is also used for:

[0107] Obtaining a form of a scientific research question input by a user; matching a target preprocessing means corresponding to the form from a configuration table of different forms and different preprocessing means; preprocessing the scientific research question by using the target preprocessing means to obtain a preprocessed scientific research question.

[0108] In the embodiment of the present application, the form of the scientific research question includes: a text form, an image form and a mixed form of text and image.

[0109] In the embodiment of the present application, the target preprocessing means corresponding to the text form includes: using internal natural language processing (NLP) technology to perform word segmentation, syntax analysis and semantic extraction on the scientific research question to obtain keywords, entities, context relationships and question intentions corresponding to the scientific research question; and identifying the fields and problem backgrounds possibly involved in the scientific research question based on the keywords and semantic units.

[0110] In the embodiment of the present application, the target preprocessing means corresponding to the image form includes: performing visual feature extraction and local pattern recognition on the image to obtain scene information, object labels, color information and other descriptive features from the image; and generating image labels and descriptions which are integrated into subsequent analysis in the form of text.

[0111] In the embodiment of the present application, the first intelligent agent is further configured to:

[0112] performing semantic understanding on the preprocessed scientific research question to obtain a semantic understanding result; splitting the first question intention according to the semantic relationship to obtain a plurality of second question intentions; generating a plurality of first query questions according to the plurality of second question intentions and the to-be-queried entity; screening an association relationship list associated with the to-be-queried entity and the keyword from the three-dimensional knowledge graph; generating an expansion list according to the association relationship list and the to-be-queried entity; expanding the plurality of first query questions according to the expansion list to obtain a plurality of second query questions; performing deformation adjustment on the plurality of first query questions and the plurality of second query questions according to a configured chain deformation rule to obtain a plurality of third query questions; filtering the plurality of first query questions, the plurality of second query questions and the plurality of third query questions according to the semantic understanding result to obtain at least one target query question; and generating a first query request corresponding to the scientific research question according to the obtained target query question.

[0113] In the embodiment of the present application, the semantic understanding result includes: a to-be-queried entity, a keyword, a first question intention corresponding to the scientific research question and a semantic relationship between the to-be-queried entity and the question intention contained in the scientific research question.

[0114] In the embodiment of the present application, the first question intention corresponding to the scientific research question includes a research direction (such as biosynthesis, chemical modification and structure optimization, function, application and classification) involved in the scientific research question.

[0115] According to the semantic relationship, the first question intention is split to obtain a plurality of second question intentions, including:

[0116] For a scientific research question in a text form, the first question intention is split into a plurality of fine-grained questions (i.e., second question intentions) according to logical relationships or semantic relationships, such as "definition of key words", "principle exploration", "data relationship", and the like.

[0117] For a scientific research question in an image form, semantic extraction is performed on the image to obtain the first question intention; based on the image semantic extraction result, a follow-up question (i.e., a second question intention) for the image content is generated; for example, "what is the object relationship in the image", "how is the scene related to the question", and the like.

[0118] For a scientific research question in a mixed form of text and image, the description generated from the image and the text information are integrated, visual cues and language details are taken into account when the question is comprehended as a whole, and more comprehensive information decomposition is ensured.

[0119] In the embodiments of the present application, when the question intention of the user is obtained, a query request is not directly generated according to the question intention of the user, but the question intention is further expanded and deformed, so as to obtain a plurality of questions, so that the response information is sufficient and perfect.

[0120] For example, assuming that the input scientific research question is "how to improve the disease resistance of rice through gene editing technology"; first, the question is segmented and analyzed, and key entities (such as "gene editing technology", "rice", and "disease resistance") and semantic relationships (such as "improve" and "through") are extracted; according to the semantic relationship, the semantic understanding result is obtained: the core intention of the user is to understand how the gene editing technology acts on the rice to enhance its disease resistance.

[0121] According to the semantic understanding result, the first question intention is split into a plurality of second question intentions: intention Figure 1 : what is the mechanism of action of gene editing technology; intention Figure 2 : which genes of rice are related to disease resistance; intention Figure 3 : how to apply gene editing technology to rice genetic modification;

[0122] Combined with the query entities ("gene editing technology", "rice", and "disease resistance"), a plurality of first query questions are generated: question 1: what is the working principle of gene editing technology; question 2: which genes of rice are related to disease resistance; question 3: how to modify the genes of rice by using gene editing technology.

[0123] From the three-dimensional knowledge graph, a list of associated relationships related to "gene editing technology", "rice", and "disease resistance ability" is screened: relationship 1: gene editing technology → CRISPR-Cas9 system; relationship 2: rice gene → OsNPR1 (related to disease resistance ability); relationship 3: disease resistance ability → immune signal pathway;

[0124] According to the list of associated relationships and the entity to be queried, an expansion list is generated: expansion 1: specific operation process of CRISPR-Cas9 system; expansion 2: function of OsNPR1 gene and its regulation mechanism; expansion 3: role of immune signal pathway in plant disease resistance;

[0125] According to the expansion list, the first query question is expanded to obtain multiple second query questions: question 4: how does CRISPR-Cas9 system realize gene editing; question 5: function of OsNPR1 gene and its role in rice disease resistance; question 6: how does immune signal pathway affect plant disease resistance ability.

[0126] According to the configured chain transformation rule, the first query question and the second query question are transformed and adjusted to obtain multiple third query questions: transformation 1: how to edit OsNPR1 gene in rice using CRISPR-Cas9 system; transformation 2: whether mutation of OsNPR1 gene will affect disease resistance ability of rice; transformation 3: whether activation of immune signal pathway can enhance disease resistance ability of rice.

[0127] According to the semantic understanding result, the first query question, the second query question and the third query question are filtered to finally obtain the target query question:

[0128] Target question 1: how to edit OsNPR1 gene in rice using CRISPR-Cas9 system to enhance disease resistance ability; target question 2: function of OsNPR1 gene and its specific role in rice disease resistance.

[0129] According to the target query question, a first query request corresponding to the scientific research question is generated: query request: please provide relevant literature and experimental data on editing OsNPR1 gene in rice using CRISPR-Cas9 system to enhance disease resistance ability.

[0130] In the embodiment of the application, the first intelligent agent is also used to: count the number of words contained in the received response information; determine the richness of the response information based on the number of words counted; when the richness of the response information meets a preset richness threshold, the first intelligent agent determines that the response information can be output; if not, the first intelligent agent needs to generate a new query request again according to the response information and the scientific research question to obtain more rich second response.

[0131] In another embodiment of the present application, the method for determining the richness of the answer information comprises:

[0132] extracting all key information points from the answer; the key information points include entities, attributes and relationships; counting the number of each key information point to obtain a first score result;

[0133] classifying the answer information by type (such as definition, explanation, example, application, etc.); counting the number of information of each type; scoring according to the diversity and quantity of information types to obtain a second score result;

[0134] counting the vocabulary quantity of the vocabulary in the answer information, and determining the weight of each vocabulary according to the degree of association of each vocabulary with the entity to be queried, the keyword and the questioning intention;

[0135] calculating a third evaluation result according to the frequency, quantity, distribution in the sentence and weight of each vocabulary;

[0136] weighting and summing the first score result, the second score result and the third score result to obtain the richness of the answer information.

[0137] Specifically, the weights of the first score result, the second score result and the third score result can be obtained by manual configuration.

[0138] In an embodiment of the present application, the first agent is pre-trained using reinforcement learning, so that the first agent can construct multi-step logic from single-step reasoning; positive feedback is given when the first agent successfully decomposes a biochemistry problem and guides the second agent to complete the task; when the first agent decomposes a natural product research problem, it not only directly gives guidance, but also records the complete thinking process; the first agent automatically finds background knowledge related to natural products and organizes information fragments in logical order.

[0139] In an embodiment of the present application, the generation of the second query request comprises:

[0140] Obtain the richness of the first response information; according to the first response information, extract the question intention without response information from the question intention corresponding to the first query request and the scientific research question, that is, the unanswered question intention, to obtain the third question intention; extract a plurality of entities from the first query request and the first response information; extract a new association relationship list associated with the plurality of entities; according to the first response information, split the third question intention into a plurality of fourth question intentions; based on the configured expansion rule, expand the new association relationship list and the fourth question intention to obtain a fifth question intention; generate a plurality of fourth query questions according to the third question intention and the extracted entities; expand the fourth query question according to the expansion list to generate a fifth query question; according to the configured chain transformation rule, the fourth query question and the fifth query question are transformed and adjusted to generate a sixth query question; according to the semantic understanding result, filter the generated fourth query question, fifth query question and sixth query question to remove redundant, repeated or irrelevant questions to obtain a new target query question; and generate a second query request based on the new target query question.

[0141] In the embodiment of the present application, the generation of the query question adopts a hierarchical progressive reasoning method of the instructor, that is, starting from the overall problem, gradually refining to the specific sub-problems, and generating corresponding reasoning steps at each level: preliminary summary: based on the preprocessing result, the core intention and context of the user question are summarized. Intermediate reasoning link: for each subdivision, generate intermediate reasoning conclusions and extract key information to form a plurality of intermediate reasoning units. Final synthesis: integrate all intermediate results to form a logically coherent answer, and verify and correct it according to the historical dialogue record.

[0142] In the embodiment of the present application, the instructor, that is, the first intelligent agent, not only subdivides the current question, but also generates new subsequent questions (GENERATE NEXT QUESTION) according to the existing analysis, so as to further explore the details of the question or confirm the accuracy of the information, thereby ensuring the continuity of the reasoning chain, the continuous feedback and correction in the interaction with the user, and the gradual improvement of the knowledge logic.

[0143] In the embodiment of the present application, the generation of the first query request includes two steps: generating subsequent questions based on the historical dialogue and the reasoning state to promote in-depth discussion of the question; and constructing a detailed chain thinking prompt to disassemble the complex question into a plurality of sub-questions.

[0144] In the embodiment of the present application, the generation of the first response information includes four steps: generating a follow-up question based on the historical dialogue and the inference state to promote in-depth discussion of the question; constructing a detailed chain of thought prompt to break down complex questions into multiple sub-questions; feeding back the current progress to the user at the intermediate stage, while requesting additional information to further refine and accurately meet the needs; integrating the outputs of each sub-question to generate a final comprehensive scientific report or answer document.

[0145] In the embodiment of the present application, the first agent is further configured to: in response to the received response deviation prompt, generate a new query request based on the response deviation prompt and the corresponding query request and response information; the specific steps are similar to the generation process of the second query request, which will not be described here.

[0146] In the embodiment of the present application, the third agent will evaluate the query request generated by the first agent and the response information of the second agent, evaluate whether the query request and the response information are associated with the scientific research question input by the user, when there is no association, the third agent will generate a deviation prompt and send it to the first agent and the second agent, to prompt the first agent to generate a new query request, so that the second agent retrieves response information based on the new query request; at the same time, when the scientific research question input by the user is any question in the configured refused answer question library, the third agent is used to feedback the corresponding refused answer prompt of the question to the second agent and the first agent, to refuse to answer the corresponding question of the user; the refused answer question library is configured by the user, and contains multiple questions and corresponding refused answer prompts of the questions.

[0147] In the embodiment of the present application, the second agent is specifically configured to:

[0148] From the three-dimensional knowledge graph, filter the abstract vector, compound, comprehensive property and image description associated with the query request sent by the first agent; use the tool library to obtain the associated data associated with the query request; determine the abstract vector corresponding scientific research literature as the target scientific research literature; search the text blocks associated with the query request from the target scientific research literature; calculate the cosine similarity between each text block; determine the reference weight corresponding to each text block according to the reference relationship of the target scientific research literature corresponding to each text block; sort each text block according to the cosine similarity between each text block and the corresponding reference weight; according to the configured first screening rule, the target text block is obtained from the sorted text blocks; according to the configured second screening rule, the compound, the comprehensive property, the associated data and the image are screened respectively to obtain the target compound, the target comprehensive property, the target associated data and the target image; based on the target text block, the target compound, the target comprehensive property, the target associated data and the target image, the response information associated with the query request is generated.

[0149] In the embodiment of the present application, the second intelligent agent adopts hierarchical retrieval to quickly locate relevant articles through matching of the abstract vector of the scientific literature; searches for text blocks in the literature (i.e., the target scientific literature), and optimizes the result ranking by combining cosine similarity calculation and citation relationship weight; the DOI of the original text associated with each text block in the vector database (i.e., the second database) realizes bidirectional jump with the articles table in the relational database.

[0150] In the embodiment of the present application, the screening of the image description associated with the query request sent by the first intelligent agent includes: using the BM25 algorithm to match the image title keywords (such as "polyketide synthase"), avoiding the problem of insufficient sensitivity of traditional semantic embedding to professional terms; the search result is embedded into the LLM response through the URL to realize the mixed output of text and graphics.

[0151] For example, when the user queries "clorobiocin biosynthesis":

[0152] The vector database (i.e., the second database) returns relevant literature fragments→the relational database (i.e., the first database) extracts the corresponding product (clorobiocin)→associates the BGC table to obtain the synthetic gene cluster→the tool library calls ClustalW to align homologous gene sequences. The retrieval process automatically connects multi-dimensional data of chemistry, genetics, and bioinformatics.

[0153] In the embodiment of the present application, in the screening process, low-relevance text blocks (cosine similarity <0.7) are excluded at the second database level; the final ranking is generated by combining the BM25 score, the number of citations, and the expert annotation weight, and the multidisciplinary cross-evidence is preferentially returned; through the mixed block strategy, the text semantic integrity (fixed block) and the context overlap demand (sliding window) are balanced, and the fragment retrieval accuracy of long literature (such as review) is improved; hierarchical vector retrieval is adopted, and through two-level screening of "abstract→full text", the calculation overhead is reduced, while the fine-grained association capability is retained.

[0154] In the embodiment of the present application, the evaluation request includes: semantic understanding result, first query request, second query request and target response information;

[0155] In the embodiment of the present application, the third intelligent agent is further configured to: if the evaluation result does not satisfy the configured evaluation rule, generate a configured response deviation prompt; and send the response deviation prompt to the first intelligent agent.

[0156] In the embodiment of the present application, the evaluation rule includes: quality rule, association rule and reference rule; wherein the quality rule is used to ensure the quality of the response information; the association rule is used to ensure that the response information is associated with the semantic understanding result and the corresponding query request; and the reference rule is used to ensure that the response information must have a reference source, thereby eliminating illusory information.

[0157] In the embodiments of the present application, the evaluation of the third intelligent agent specifically includes:

[0158] The evaluation request is judged by using the configured quality rules; if the query question in the evaluation request and the questioning intention in the semantic understanding belong to one of the pre-configured rejection list or exceed the pre-configured answer range, a questioning out-of-range prompt is generated, and the corresponding question or question intention is pre-audited by an all-in-one machine or manually, and a decision is made on whether to allow answering, so as to avoid the system providing answers to questions that should not be answered or are risky;

[0159] Irrelevant information detection (Irrelevant Information Detection) is performed on the evaluation request by using the configured association rules; the answer information and the corresponding query question and questioning intention are subjected to semantic analysis and matching, and parts that do not need to be displayed are filtered out; if all the contents are determined to be irrelevant, a feedback mechanism is triggered, that is, the evaluation answer information and the corresponding question are irrelevant, and a deviated answer prompt is generated;

[0160] Tracking citation (Tracking Citation) is performed on the evaluation request by using the configured reference rules: whether the references contained in the evaluation answer information conform to the predetermined reference format and logic is checked, so as to ensure the accuracy of the reference information in the answer information and avoid improper use of references (for example, references are made to contents that are not actually used or contents that have not been referenced are used); it is checked whether each support information block correctly defines the reference category, and a guidance prompt is returned for the answer with problems, requiring the executor to regenerate the answer that meets the reference requirements according to the feedback.

[0161] In the embodiments of the present application, when the evaluation finds that the answer information does not meet the evaluation rules, the evaluator will specifically point out the problems, such as rejection judgment failure, excessive irrelevant information, or reference errors. The feedback content is conveyed to the instructor in the form of a guidance prompt, and it is explicitly stated which part of the content does not meet the expected standard; the instructor, that is, the first intelligent agent, may need to adjust the original question or the current chain-of-thought prompt (CoT) and make changes in generating the next search or question refinement; if the evaluator removes a large amount of irrelevant information, the instructor may regenerate a more accurate search keyword and query parameter, thereby improving the relevance of the subsequent search results.

[0162] In one embodiment of the present application, when the third intelligent agent judges that the answer information does not meet the evaluation rules, the executor may be directly required to initiate a new information search, or a prompt is issued to the user to supplement or clarify the question, so as to form a more accurate context.

[0163] In the embodiments of the present application, the third intelligent agent, i.e., the evaluator, pre-constructs a biochemistry professional evaluation environment to perform real-time evaluation on the output of the executor; the third intelligent agent evaluates the scientific accuracy of natural product information according to the configured scoring standard; adjusts the evaluation standard according to external feedback to improve the ability to judge scientific accuracy; optimizes the expression of the biochemistry content generated by the executor to ensure professionalism and understandability.

[0164] As shown in Figure 1 , the t-round interaction is performed between the mentor and the executor in the scientific research question and answer reasoning all-in-one machine based on multi-agent cooperation. In each round, the mentor submits a query request to the executor, the executor processes the query and retrieves relevant context information from various knowledge bases. This information may include text, images, references and other supporting materials to help with the current query. Based on the user's original intention and the results returned by the executor in the previous rounds, the mentor will formulate the next query. After completing the t-round interaction, the mentor will summarize the questions and answers generated in the entire interaction process to form the final response and return it to the user. Evaluation process: the evaluator interacts with the executor during the entire process of retrieving and generating answers (before, during and after) to improve the quality of the answers. Error reporting: if the evaluator finds abnormal situations during the retrieval process of the executor, the evaluator will feed back these problems to the mentor.

[0165] In the embodiments of the present application, the first intelligent agent, i.e., the mentor, in view of the fact that users may need to interact with large language models multiple times when solving complex problems, point out misunderstandings in their answers or point out deficiencies in temporary answers until the model provides the desired answer, the system introduces a mentor as an intermediary between the user and the system executor. After the user submits the input, the mentor uses the chain-of-thought (CoT) technology to analyze and process the user's query step by step with the executor to guide the executor to retrieve sufficient information, and then generate a complete response and present it to the user (see Figure 1 ). In addition, the mentor also plays an important role in providing feedback to the user, guiding him to improve the query or elaborating the problem in detail after identifying the abnormal situations found by the evaluator (see Figure 1 ). The function of the mentor is similar to the cognitive function of the brain, which can complete tasks such as thinking, finding and correcting errors, and its dynamic ability is not limited to simple query processing, but can also generate insights beyond the scope of the original problem.

[0166] In the embodiments of the present application, the second intelligent agent, i.e., the executor:

[0167] To address the problem of hallucinated information in domain-specific queries, the executor of ChatT2 employs a Retrieval-Augmented Generation (RAG) framework, retrieving relevant information from external databases and combining it with the contextual learning capabilities of large language models to generate traceable answers. As shown in Figure 1 The executor is responsible for understanding the input delivered by the mentor and making corresponding action decisions based on historical memory. In responding to the mentor's query, the executor retrieves additional information from the database or calls tools based on the mentor's input and historical conversation records, and then integrates the information from different sources into a structured text format (e.g., XML), which is used as the context input for the large language model to generate the final answer. The final supporting context includes four parts: <text>(text), <figure>(image) and <tool>(tool).

[0168] In the process of retrieving information from relational and vector databases, queries are converted into SQL statements and text embeddings. SQL statements are executed to retrieve table information related to the query, which is represented as a vector in the vector database.< / tool> < / figure> < / text> (Table), Figure 3 s v v s <text>Tag presentation. For image database retrieval, the query is matched with the image title by BM2.5 score, the image with high score is identified and its URL is saved as image information, to <figure>Tag presentation.

[0169] In the tool repository retrieval process, the executor parses the key query information into the input of pre-packaged execution scripts, and the execution results of these scripts are regarded as tool repository information and encapsulated in <tool>The four different databases work together to ensure that the executor can obtain the maximum amount of traceable supporting materials, so as to effectively deal with the queries raised by the mentor. The detailed database construction method is described in the method part.

[0170] In the embodiments of the present application, the third intelligent agent, i.e., the evaluator:

[0171] Although the executor takes various measures to ensure that ChatT2 can provide accurate answers to domain-specific problems, another key goal is to ensure that ChatT2 can refuse to answer content beyond its capabilities, thereby improving the credibility of the results, which is particularly important for knowledge-intensive tasks such as question-answering. To achieve this goal, the present application designs an additional intelligent agent, i.e., the evaluator. The evaluator interacts with the executor and performs three key tasks, i.e., refusal to answer, irrelevant information filtering, and reference tracking, in the following three stages: (1) the preprocessing stage when the user query is submitted; (2) the intermediate processing stage when the executor retrieves contextual information; and (3) the post-processing stage when the executor generates the final context. This process ensures the credibility of the final results and prevents the inclusion of false or fabricated information. For queries that ChatT2 cannot answer or the executor fails to retrieve relevant information, the evaluator provides feedback by refusing to answer; for answerable queries, the evaluator ensures that the executor includes the correct reference source in the final answer, thereby enhancing the factual reliability and practical application value of the answer, which is particularly beneficial for researchers when analyzing the results.

[0172] The preferred embodiments of the present application are described below in conjunction with the accompanying drawings of the specification, it should be understood that the preferred embodiments described herein are only used to illustrate and explain the present application, and are not used to limit the present application, and the embodiments in the present application and the features in the embodiments can be combined with each other without conflict.

[0173] Figure 3 A flowchart of a scientific research question-answering reasoning control method based on multi-agent cooperation provided in the embodiments of the present application. As shown in Figure 1 , the method can include:

[0174] Step S310, obtaining a scientific research question input by a user; controlling a first intelligent agent to perform semantic understanding on the scientific research question to obtain a semantic understanding result; performing semantic expansion on the semantic understanding result according to a configured expansion rule to obtain a first query request corresponding to the scientific research question, and sending the first query request to a second intelligent agent.

[0175] Step S320, controlling the second intelligent agent to filter first answer information associated with the first query request from a pre-constructed scientific research knowledge base, and sending the first answer information to the first intelligent agent; controlling the first intelligent agent to receive the first answer information.

[0176] Step S330, if the richness of the first response information does not satisfy the configured richness threshold, a second query request is generated according to the first response information and the semantic understanding result, and the second query request is sent to the second agent.

[0177] Step S340, the second agent is controlled to filter the second response information associated with the second query request from the pre-constructed scientific research knowledge base, and the second response information is sent to the first agent; the first agent is controlled to receive the second response information.

[0178] Step S350, if the lexical richness of the second response information satisfies the configured lexical richness threshold, the second response information is taken as the target response information of the scientific research question; an evaluation request is generated, and the evaluation request is sent to the third agent.

[0179] Step S360, the third agent is controlled to evaluate whether the target response information and the corresponding semantic understanding result, the first query request and the second query request satisfy the configured evaluation rules, and an evaluation result is obtained.

[0180] Step S370, the first agent is controlled to receive the evaluation result; if the evaluation result satisfies the configured evaluation rules, the target response information is output.

[0181] In the embodiment of the present application, the evaluation request includes: the semantic understanding result, the first query request, the second query request and the target response information.

[0182] For example, the user: "Please predict the possible antibacterial activity of compound Y";

[0183] 1. Tutor thought chain:

[0184] 1.1 Analyze the pharmacophore of compound Y;

[0185] 1.2 Determine the similar structure database to be searched;

[0186] 1.3 Design an activity prediction methodology framework;

[0187] 2. Performer performs:

[0188] 2.1 Query similar compounds according to the tutor's guidance;

[0189] 2.2 Apply QSAR model for prediction;

[0190] 2.3 Generate a preliminary activity prediction report;

[0191] 3. Evaluator evaluates:

[0192] 3.1 Check the rationality of the prediction (whether it conforms to the structure-activity relationship rule);

[0193] 3.2 Compare with known similar compound data;

[0194] 3.3 Propose revision and score;

[0195] 4. Feedback loop:

[0196] 4.1 Adjust executor's prediction strategy according to score;

[0197] 4.2 Tutor adjusts the direction of the thinking chain according to the feedback;

[0198] 4.3 System integrates the final prediction result.

[0199] In a further embodiment provided by the present application, a computer readable storage medium is also provided, which stores instructions, when running on a computer, causes the computer to perform the multi-agent collaboration based scientific research question and answer reasoning control method in any of the above embodiments.

[0200] In a further embodiment provided by the present application, a computer program product containing instructions is also provided, when running on a computer, causes the computer to perform the multi-agent collaboration based scientific research question and answer reasoning control method in any of the above embodiments.

[0201] Those skilled in the art should understand that the embodiments in the present application can be provided as methods, systems or computer program products. Therefore, the embodiments in the present application can be in the form of entirely hardware embodiments, entirely software embodiments or embodiments combining software and hardware aspects. Moreover, the embodiments in the present application can be in the form of computer program products implemented on one or more computer usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer usable program code.

[0202] The embodiments in the present application are described with reference to the flowcharts and / or block diagrams according to the methods, devices (systems) and computer program products of the embodiments in the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of the flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general purpose computer, a special purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device produce a device that implements the functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in one or more flows and / or blocks Figure 1 The means for performing the functions specified in one or more flows and / or blocks.

[0203] These computer program instructions can also be stored in a computer- readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instructions which implement the Figure 1 function specified in the flow or flows and / or blocks Figure 1 of the block or blocks.

[0204] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions that are executed on the computer or other programmable apparatus provide steps for implementing the Figure 1 function specified in the flow or flows and / or blocks Label presentation; meanwhile, the relevant literature sources are collected to form a literature set denoted as I. The text embedding part is used for hierarchical semantic similarity matching based on cosine similarity. First, a set of semantically similar literature is determined denoted as I, and then the semantic similarity of the document fragments obtained from I∪I is matched using the embedding vector of the query, so as to extract the main text of the paper related to the query. This information is presented in the form of a list of the top 10 documents, and the semantic similarity of each document is calculated and displayed. of the block or blocks.

[0205] While preferred embodiments in the application have been described, those skilled in the art will be able to devise various modifications and alternatives to embodiments of the application without departing from the spirit and scope of the embodiments of the application. Accordingly, it is intended that the appended claims cover all such modifications and alternatives as fall within the scope of the embodiments of the application.

[0206] It is apparent that a person skilled in the art can make various changes and modifications to the embodiments of the application without departing from the spirit and scope of the embodiments of the application. Thus, it is intended to include these changes and modifications in the embodiments of the application. The embodiments of the application are intended to cover the application as falling within the scope of the claims and their equivalents.< / tool> < / figure> < / text> ​

Claims

1. A scientific research question and answer reasoning all-in-one machine based on multi-agent cooperation, characterized in that, The all-in-one machine is applied to scientific research question and answer in the field of natural products of microorganisms, drug research and development, biology and chemistry, and comprises: A first intelligent agent is configured to perform semantic understanding on a scientific research question input by a user to obtain a semantic understanding result, perform semantic expansion on the semantic understanding result according to a configured expansion rule to obtain a first query request corresponding to the scientific research question, and send the first query request to a second intelligent agent; and receive first response information sent by the second intelligent agent in response to the first query request. If the richness of the first response information does not satisfy a configured richness threshold, a second query request is generated according to the first response information and the semantic understanding result, and the second query request is sent to the second intelligent agent. Second response information sent by the second intelligent agent in response to the second query request is received. If the lexical richness of the second response information satisfies a configured lexical richness threshold, the second response information is taken as target response information of the scientific research question. An evaluation request is generated and sent to a third intelligent agent. An evaluation result sent by the third intelligent agent in response to the evaluation request is received. The evaluation request comprises the semantic understanding result, the first query request, the second query request and the target response information. If the evaluation result satisfies a configured evaluation rule, the target response information is output. A second intelligent agent is configured to filter response information associated with the query request sent by the first intelligent agent from a pre-constructed scientific research knowledge base, and send the corresponding response information to the first intelligent agent. A third intelligent agent is configured to evaluate whether the target response information and the corresponding semantic understanding result, first query request and second query request satisfy a configured evaluation rule to obtain an evaluation result. The first intelligent agent is further configured to: extract all key information points from the response information; the key information points comprise entities, attributes and relationships; and count the number of each key information point to obtain a first score result; classify the response information according to types; count the number of information of each type; and score according to the diversity and quantity of information types to obtain a second score result; count the lexical quantity of the words in the response information, and determine the weight of each word according to the association degree of each word with the queried entity, keyword and questioning intention; calculate a third evaluation result according to the frequency, quantity, distribution in a sentence and weight of each word; weight and sum the first score result, second score result and third score result to obtain the richness of the response information.

2. The all-in-one machine of claim 1, wherein, The all-in-one machine further comprises a knowledge base construction module and an interactive display device. The interactive display device is configured to input the scientific research question by the user and display the target response information output by the first intelligent agent. The knowledge base construction module is specifically configured to: obtain multi-source scientific research data; wherein the scientific research data comprises compounds, comprehensive characteristics of different compounds and scientific research literature containing any compound; construct a first database based on the comprehensive characteristics of the compounds and different compounds; generating a summary vector of each scientific research literature based on an abstract of a scientific research literature containing each compound; and constructing a second database based on the summary vector of each scientific research literature and each scientific research literature; extracting images associated with any compound and image descriptions of the corresponding images from each scientific research literature, and constructing a third database; associating the comprehensive characteristics of each compound, the summary vector of each scientific research literature, and the image descriptions with each other as unique identifiers, and constructing a three-dimensional knowledge graph; generating a scientific research knowledge base based on the first database, the second database, the third database, the three-dimensional knowledge graph, and a preconfigured tool library.

3. The unit of claim 2, wherein, The first agent is also configured to: obtain the form of the scientific research question input by the user; wherein the form of the scientific research question includes: text form, image form, and mixed form of text and image; match the target preprocessing means corresponding to the form from the configured different forms and different preprocessing means configuration table; preprocess the scientific research question using the target preprocessing means to obtain the preprocessed scientific research question.

4. The machine of claim 3, wherein The first agent is specifically configured to: perform semantic understanding on the preprocessed scientific research question to obtain a semantic understanding result; wherein the semantic understanding result includes: a to-be-queried entity contained in the scientific research question, a keyword, a first query intention corresponding to the scientific research question, and a semantic relationship between the to-be-queried entity and the query intention; split the first query intention according to the semantic relationship to obtain a plurality of second query intentions; generate a plurality of first query questions according to the plurality of second query intentions and the to-be-queried entity; filter an association relationship list associated with the to-be-queried entity and the keyword from the three-dimensional knowledge graph; generate an expansion list according to the association relationship list and the to-be-queried entity; expand the plurality of first query questions according to the expansion list to obtain a plurality of second query questions; perform transformation adjustment on the plurality of first query questions and the plurality of second query questions according to the configured chain transformation rule to obtain a plurality of third query questions; filter the plurality of first query questions, the plurality of second query questions, and the plurality of third query questions according to the semantic understanding result to obtain at least one target query question; generate a first query request corresponding to the scientific research question according to the obtained target query question.

5. The unit of claim 4, wherein, The second agent is specifically configured to: filter the summary vector, the compound, the comprehensive characteristics, and the image description associated with the query request sent by the first agent from the three-dimensional knowledge graph; obtain the associated data associated with the query request using the tool library; determine the scientific research literature corresponding to the summary vector as a target scientific research literature; search for text blocks associated with the query request from the target scientific research literature; calculate the cosine similarity between each text block; determine the reference weight corresponding to each text block according to the reference relationship of the target scientific research literature corresponding to each text block; sort each text block according to the cosine similarity between the text blocks and the corresponding reference weight; screen the target text block from the sorted text blocks according to the configured first screening rule; screen the compound, the comprehensive characteristic, the associated data and the image according to the configured second screening rule respectively, to obtain a target compound, a target comprehensive characteristic, target associated data and a target image; generate response information associated with the query request based on the target text block, the target compound, target comprehensive characteristic, target associated data and target image.

6. The unit of claim 1, wherein, The third agent is further configured to: generate a configured response deviation prompt if the evaluation result does not satisfy the configured evaluation rule, and send the response deviation prompt to the first agent; The first agent is further configured to: generate a new query request based on the response deviation prompt and the corresponding query request and response information in response to the received response deviation prompt.

7. A scientific research question and answer reasoning control method based on multi-agent cooperation, characterized in that, The method is applied to scientific research question and answer in the fields of natural products of microorganisms, drug research and development, biology and chemistry, and the method comprises: obtaining a scientific research question input by a user; controlling the first agent to perform semantic understanding on the scientific research question to obtain a semantic understanding result, performing semantic expansion on the semantic understanding result according to a configured expansion rule to obtain a first query request corresponding to the scientific research question, and sending the first query request to the second agent; controlling the second agent to screen first response information associated with the first query request from a pre-constructed scientific research knowledge base, and sending the first response information to the first agent; controlling the first agent to receive the first response information, and generating a second query request according to the first response information and the semantic understanding result if the richness of the first response information does not satisfy a configured richness threshold, and sending the second query request to the second agent; controlling the second agent to screen second response information associated with the second query request from the pre-constructed scientific research knowledge base, and sending the second response information to the first agent; controlling the first agent to receive the second response information, and taking the second response information as target response information of the scientific research question if the lexical richness of the second response information satisfies a configured lexical richness threshold, generating an evaluation request, and sending the evaluation request to the third agent; wherein the evaluation request comprises the semantic understanding result, the first query request, the second query request and the target response information; controlling the third agent to evaluate whether the target response information and the corresponding semantic understanding result, first query request and second query request satisfy a configured evaluation rule to obtain an evaluation result; controlling the first agent to receive the evaluation result, and outputting the target response information if the evaluation result satisfies the configured evaluation rule. The method for determining the richness of the response information comprises: extracting all key information points from the response information; the key information points comprise entities, attributes and relationships; and counting the number of the key information points to obtain a first score result; classifying the response information according to types; counting the number of information of each type; and scoring according to the diversity and quantity of the information types to obtain a second score result; The vocabulary quantity of the words in the statistical response information is determined according to the association degree of each word with the entity to be queried, the keyword and the query intention, and the weight of each word is determined; The third evaluation result is calculated according to the frequency, quantity, distribution in the sentence and weight of each word; The richness of the response information is obtained by weighted summation of the first score result, the second score result and the third score result.

8. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, and the computer program is executed by the processor to realize the method in claim 7.

Images