Literature retrieval and reading method based on multi-agent

Abstract

The present application discloses a method for document retrieval and reading based on multi-agents. The method comprises: using a meta-agent to decompose the task to be retrieved into target sub-tasks, and determining the collaborative work strategy between the target sub-tasks; according to the collaborative work strategy, calling the sub-agent corresponding to the target sub-task to execute the target sub-task, and determining the execution result of the target sub-task; according to the execution result, determining the retrieval result of the task to be retrieved. The present application uses a meta-agent to deeply understand user needs, decompose the task to be retrieved into sub-tasks, and plan the collaborative work strategy between the target sub-tasks, and conduct document retrieval and reading through multi-agent collaboration, thereby improving the efficiency of document study, solving the illusion problem existing in the processing of academic tasks, and improving the relevance of retrieval information and the accuracy of retrieval results.

Description

Technical Field

[0001] The present application relates to the technical field of artificial intelligence and large language models, and in particular to a document retrieval and reading method based on multi-agents. Background Art

[0002] With the rapid development of information technology, academic resources have exploded. When researchers use various academic search engines, they need to spend a lot of energy to screen, read and organize literature materials. Faced with the needs of interdisciplinary research, it is difficult for researchers to quickly grasp key information in unfamiliar fields, which further increases the difficulty of scientific research.

[0003] In related technologies, some academic search engines and recommendation systems have applied artificial intelligence technology to keyword matching and information recommendation. However, these tools can usually only complete simple search tasks and conduct simple questions and answers. They cannot directly understand the user's deep-level needs and accurately retrieve relevant information from documents. They cannot conduct in-depth reading of documents or conduct comprehensive analysis among multiple documents. In addition, there are also hallucination problems and single-perspective problems when dealing with academic tasks. Summary of the invention

[0004] In view of this, the present application provides a multi-agent-based document retrieval and reading method, which uses meta-agents to deeply understand user needs, decomposes the task to be searched into sub-tasks, and plans collaborative work strategies between target sub-tasks. Through multi-agent collaboration, document retrieval and reading are performed to improve the efficiency of document reading, solve the illusion problem that exists when processing academic tasks, and improve the relevance of retrieval information and the accuracy of retrieval results.

[0005] According to one aspect of the present application, a multi-agent-based document retrieval and reading method is provided, comprising: using a meta-agent to decompose a task to be retrieved into target sub-tasks, and determining a collaborative working strategy between the target sub-tasks, wherein the meta-agent is fine-tuned based on historical retrieval tasks and a preset document library; according to the collaborative working strategy, calling a sub-agent corresponding to the target sub-task to execute the target sub-task, and determining an execution result of the target sub-task, wherein the sub-agent is constructed based on a large language model, and is fine-tuned based on historical target sub-tasks and the preset document library; and determining a retrieval result of the task to be retrieved based on the execution result.

[0006] According to another aspect of the present application, a multi-agent-based document retrieval and reading device is provided, comprising:

[0007] A decomposition module, used to decompose the task to be searched into target subtasks by using a meta-agent, and determine the collaborative working strategy between the target subtasks, wherein the meta-agent is fine-tuned according to historical search tasks and a preset document library;

[0008] A calling module, used for calling the sub-agent corresponding to the target sub-task to execute the target sub-task according to the collaborative work strategy, and determining the execution result of the target sub-task, wherein the sub-agent is constructed based on a large language model and is fine-tuned according to the historical target sub-task and the preset document library;

[0009] A determination module is used to determine the search result of the task to be searched according to the execution result.

[0010] According to another aspect of the present application, a readable storage medium is provided, on which a program or instruction is stored. When the program or instruction is executed by a processor, the steps of the above-mentioned multi-agent-based document retrieval and reading method are implemented.

[0011] According to another aspect of the present application, a computer device is provided, including a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, and the processor implements the steps of the above-mentioned multi-agent-based document retrieval and reading method when executing the program.

[0012] Through the above technical solution, this application provides a multi-agent-based document retrieval and reading method, which uses meta-agents to deeply understand user needs, decomposes the task to be retrieved into sub-tasks, and performs document retrieval and reading according to the decomposed target sub-tasks to ensure the relevance of the retrieval information and the accuracy of the retrieval results. At the same time, the meta-agent is used to plan the collaborative work strategy between the target sub-tasks, and the document retrieval and reading are performed through multi-agent collaboration, which improves the efficiency of document reading and solves the illusion problem in handling academic tasks. Researchers can quickly obtain high-quality retrieval results without spending a lot of time reading all the documents.

[0013] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0015] Figure 1 A schematic diagram of the process of a multi-agent-based document retrieval and reading method provided in an embodiment of the present application is shown;

[0016] Figure 2 A structural block diagram of a multi-agent-based document retrieval and reading device provided in an embodiment of the present application is shown. DETAILED DESCRIPTION

[0017] The present application will be described in detail below with reference to the accompanying drawings and in combination with embodiments. It should be noted that the embodiments and features in the embodiments of the present application can be combined with each other without conflict.

[0018] The embodiments of the present application are described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present application, and cannot be interpreted as limiting the present application.

[0019] Those skilled in the art will appreciate that, unless expressly stated, the singular forms "a", "an", "said" and "the" used herein may also include plural forms. It should be further understood that the term "comprising" used in the specification of the present application refers to the presence of the features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or groups thereof. It should be understood that when we refer to an element as being "connected" or "connected" to another element, it may be directly connected or connected to the other element, or there may be intermediate elements. In addition, the "connection" or "connection" used herein may include wireless connection or wireless fusion. The term "and / or" used herein includes all or any unit and all combinations of one or more associated listed items.

[0020] Now, exemplary embodiments according to the present application will be described in more detail with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in a variety of different forms and should not be interpreted as being limited to the embodiments set forth herein. It should be understood that these embodiments are provided to make the disclosure of the present application thorough and complete, and to fully convey the concepts of these exemplary embodiments to those of ordinary skill in the art.

[0021] In this embodiment, a multi-agent based document retrieval and reading method is provided. Figure 1 As shown, the method includes:

[0022] Step 101, using a meta-agent to decompose the task to be retrieved into target subtasks, and determining the collaborative working strategy between the target subtasks.

[0023] Among them, the meta-agent is built based on a large language model and fine-tuned according to historical retrieval tasks and preset literature library.

[0024] Step 102: According to the collaborative work strategy, the sub-agent corresponding to the target sub-task is called to execute the target sub-task and determine the execution result of the target sub-task.

[0025] Among them, the sub-agent is built based on the large language model and fine-tuned according to the historical target sub-tasks and the preset literature library.

[0026] Step 103: Determine the search result of the task to be searched according to the execution result.

[0027] In this embodiment, a task to be retrieved is generated based on the query information input by the user, and a pre-fine-tuned meta-agent is used to understand the user's deep-seated needs in the task to be retrieved. The meta-agent is used to decompose the task to be retrieved into sub-tasks based on the user's real needs, and a collaborative working strategy between target sub-tasks is planned. Thus, document retrieval and reading are performed through multi-agent collaboration to ensure the relevance of retrieval information and the accuracy of retrieval results, improve the efficiency of document reading, prevent hallucinations when processing academic tasks, and ensure the reliability of retrieval.

[0028] In one embodiment, the multi-agent based document retrieval and reading method also includes: marking according to the historical intent of the historical retrieval task to determine the first label of the historical retrieval task; marking according to the historical collaborative work strategy of the historical retrieval task to determine the second label of the historical retrieval task; and fine-tuning the preset large language model according to the preset document library and the historical retrieval tasks carrying the first label and the second label to obtain a meta-agent.

[0029] In this embodiment, a large amount of academic literature data and historical retrieval tasks related to the academic field are collected in advance to build a high-quality literature retrieval reading database, and the constructed literature retrieval reading database is used to fine-tune the preset large language model, so as to use the academic literature data to help the preset large language model learn how to decompose the retrieval tasks related to the academic field into sub-tasks, and learn how to plan the collaborative work strategies between sub-tasks, so as to obtain a meta-agent based on the fine-tuned preset large language model.

[0030] Exemplarily, the collected historical retrieval tasks are annotated. On the one hand, according to the specific intent of the user query in the historical retrieval task, the preset subtask corresponding to the historical retrieval task is annotated to determine the first label of the historical retrieval task. On the other hand, the historical collaborative work strategy of the preset subtask corresponding to the historical retrieval task is annotated to record the specific methods and strategies adopted by the preset large language model in planning the historical collaborative work strategy.

[0031] Among them, the collaborative work strategy includes three strategies: serial, parallel, and hybrid operation. Specifically, under the serial strategy, the sub-agents corresponding to the subtasks execute the subtasks one by one in the order under the serial strategy. The execution of each subtask must wait for the completion of the previous subtask before it can start, and the subsequent subtask needs to rely on the output result of the previous subtask. For example, if there is a clear dependency relationship between subtasks, the serial strategy can be selected. In parallel mode, multiple subtasks can be executed simultaneously, and different subtasks are independent of each other, which can make full use of computing resources to accelerate the completion of the overall task. The hybrid mode combines the advantages of serial and parallel strategies, and flexibly adjusts the execution mode of subtasks according to the relationship and dependency between different subtasks. In the hybrid mode, some subtasks will be executed serially, while other subtasks can be executed in parallel, thereby maximizing resource utilization and improving efficiency. It should be noted that here, the subtask refers to the preset subtask corresponding to the historical retrieval task in the scenario of fine-tuning the preset large language model, and refers to the target subtask in the scenario of applying the meta-agent.

[0032] For example, if the task types of the preset subtasks include retrieval, summary and comparison, and a historical retrieval task is "the latest progress in artificial intelligence in the medical field", then the first label of the historical retrieval task will be marked as a retrieval subtask and a summary subtask, so that the relevant literature on artificial intelligence in the medical field is retrieved through the retrieval subtask, and then the latest progress in artificial intelligence in the medical field is obtained through the summary subtask based on the relevant literature on artificial intelligence in the medical field. Further, the second label of the historical retrieval task will be marked as a serial strategy, so that the retrieval subtask is executed first, and then the execution results of the retrieval subtask are summarized through the summary subtask. For another example, if a historical retrieval task is "find papers on artificial intelligence and biomedicine, and list their main research directions", firstly, two retrieval subtasks are marked for the historical retrieval task to retrieve papers on artificial intelligence and biomedicine respectively. However, the two retrieval subtasks of retrieving papers on artificial intelligence and biomedicine can be performed at the same time, so the parallel strategy is selected. For another example, if a historical retrieval task is "find papers on big data analysis, summarize the content of one of the important papers, and compare the methods of the other two papers", firstly, the historical retrieval task is marked with a retrieval subtask, a summary subtask and a comparison subtask. After the retrieval subtask retrieves the papers on big data analysis, the summary subtask and the comparison subtask can be performed in parallel, so the hybrid mode is selected.

[0033] Furthermore, the preset large language model is fine-tuned using the annotated literature retrieval reading database so that the preset large language model can learn the annotated data. After repeated iterative fine-tuning, the preset large language model can identify subtasks of different task types in the retrieval tasks in the academic field, and accurately plan the collaborative work strategies between subtasks, realize the coordination and management of multiple sub-agents, and thus obtain a meta-agent.

[0034] Exemplarily, the preset large language model analyzes the annotated historical retrieval tasks and extracts various features. For example, the features such as part of speech, word frequency, semantic vector, etc. of the keywords of the historical retrieval task can be extracted, and the grammatical structure, length, and occurrence of academic field terms of the historical retrieval task can also be analyzed. For each historical retrieval task, these features are combined into a feature vector representation. During the fine-tuning process, the preset large language model learns to associate the input feature vector with the corresponding preset subtask. Through a large amount of training data and repeated iterative fine-tuning, the historical retrieval task gradually masters the mapping relationship between different feature combinations and preset subtasks. At the same time, through repeated iterative fine-tuning, the preset large language model continuously optimizes its own parameters to improve the accuracy of identifying subtasks and planning collaborative work strategies in academic retrieval reading tasks, thereby obtaining a meta-agent.

[0035] It should be noted that the large language model is the core of the meta-agent, but the meta-agent is more powerful and can execute different and complex commands.

[0036] Furthermore, as a refinement and extension of the specific implementation methods of the above-mentioned embodiment, in order to fully illustrate the specific implementation process of this embodiment, a meta-agent is used to decompose the task to be retrieved into target subtasks, specifically including: performing semantic recognition processing on the task to be retrieved to determine the semantic feature information of the task to be retrieved; matching the semantic feature information with preset subtasks of different task types, and determining the preset subtask with matching semantic feature information as the target subtask of the task to be retrieved.

[0037] In this embodiment, a meta-agent is used to perform semantic understanding of the task to be retrieved and identify the specific intention of the task to be retrieved, thereby decomposing the complex task to be retrieved into target subtasks to focus on the specific needs of each target subtask, and then combining the context of the task to be retrieved and the knowledge of the academic field to provide more in-depth and professional retrieval results.

[0038] Exemplarily, the meta-agent can use natural language processing technology to perform lexical analysis, syntactic analysis, and semantic role labeling on the task to be retrieved, so as to understand the meaning of the task to be retrieved and determine the language feature information of the task to be retrieved. For example, for the task to be retrieved "compare the performance of different machine learning algorithms in image classification and list their advantages and disadvantages", the meta-agent can identify the two key actions of "compare" and "list", as well as the two key topics of "machine learning algorithm" and "image classification". In addition, the meta-agent can split the task to be retrieved into individual words or vocabulary units, and identify the part of speech of each word (such as noun, verb, adjective, etc.). For example, for the task to be retrieved "compare the performance of different machine learning algorithms in image classification and list their advantages and disadvantages", it is identified that "compare" is a verb and "machine learning algorithm" is a noun phrase. In addition, the meta-agent analyzes the sentence structure of the task to be retrieved, determines the subject, predicate, object and other components and the relationship between them. For example, in the above-mentioned task to be retrieved, "different machine learning algorithms" are the objects of comparison, "performance" is the content of comparison, and the object of "listing" is "advantages and disadvantages". At the same time, the meta-agent determines the semantic roles played by each component in the sentence of the task to be retrieved, such as agent, patient, time, place, etc., which helps to understand the meaning of the query more accurately. For example, in the above task to be retrieved, "different machine learning algorithms" are the agents, and "performance" and "advantages and disadvantages" are the patients. Then, the obtained semantic feature information is matched with the preset subtasks of different task types, and the preset subtasks with matching semantic feature information are determined as the target subtasks of the task to be retrieved.

[0039] It should be noted that, in this embodiment, the task types of the preset subtasks include retrieval (first task type), summary (second task type), comparison (third task type), enumeration (fourth task type) and sequence (fifth task type). Among them, the retrieval subtask focuses on finding and locating relevant documents or data, and the main goal is to effectively find document resources and related information that meet user needs. The summary subtask focuses on extracting and summarizing the core ideas, main viewpoints and key information of academic documents or information, and the main goal is to simplify long texts or complex content into concise and easy-to-understand overviews. The comparison subtask focuses on comparing and analyzing the similarities and differences between different academic documents or information to extract valuable insights or conclusions. The main goal is to help users understand the characteristics, advantages and shortcomings of each document by systematically comparing different academic information. The enumeration subtask focuses on listing relevant documents, viewpoints, concepts and other information in academic documents or information. The main goal is to systematically organize and present information to facilitate users to quickly obtain and browse related content. The sequence subtask focuses on identifying the order of events, steps or processes mentioned in academic literature or information, and arranging the information in chronological or logical order, highlighting the order in which events, steps or processes occur. The main goal is to describe events, steps or processes in sequence.

[0040] Furthermore, as a refinement and extension of the specific implementation methods of the above-mentioned embodiments, in order to fully illustrate the specific implementation process of this embodiment, the semantic feature information is matched with the preset subtasks, including: identifying the first target data in the preset document library that meets the first preset condition; determining the semantic relationship between the first target data based on the preset document library; constructing a document semantic network diagram based on the first target data and the semantic relationship; mapping the semantic feature information to the nodes in the document semantic network diagram, and matching the semantic feature information with the preset subtasks corresponding to the nodes through a semantic reasoning algorithm.

[0041] In this embodiment, a document semantic network diagram is constructed based on a preset document library, and the potential information between academic document data is deeply mined. The document semantic network diagram is used to match the preset subtasks that best meet user needs through semantic reasoning, thereby improving the accuracy and efficiency of document retrieval and analysis.

[0042] Exemplarily, natural language processing technology is used to identify first target data that meets the first preset condition from a large amount of document data contained in a preset document library. The first preset condition is used to identify key concepts and entities in the academic field from a large amount of document data, thereby determining the first target data, and then using the first target data as a node of the document semantic network. For example, from a document on the application of artificial intelligence in medical imaging diagnosis, entities such as "artificial intelligence" and "medical imaging diagnosis" can be identified.

[0043] Next, the semantic relationship between the first target data is extracted based on the preset document library to establish edges between nodes. For example, semantic relationships can include causal relationships, association relationships, inclusion relationships, parallel relationships, etc. Language relationship types can also include "proposed" (a certain scholar proposed a certain theory), "applied to" (a certain method is applied to a certain research object), "published in" (a certain paper is published in a certain journal), etc. Thus, a document semantic network diagram is constructed based on the first target data and the semantic relationship.

[0044] Furthermore, the key concepts and entities in the semantic feature information of the task to be retrieved are matched and mapped with the nodes and edges in the document semantic network diagram. If a completely matching node is found, the semantic feature information is associated with the node; if there is no completely matching node, the closest node can be found through approximate matching or semantic similarity calculation for association.

[0045] Furthermore, the important concepts and entities in the semantic feature information are used as starting nodes. For example, if the semantic feature information is "Comparing the performance of different artificial intelligence algorithms in image recognition", the two entities "artificial intelligence algorithm" and "image recognition" can be determined as starting nodes. Next, the shortest path algorithm is used to find the shortest path from the starting node to the node associated with the preset subtask in the document semantic network diagram. Then, based on the semantic relationship between the nodes on the shortest path, the preset subtask matching the semantic feature information is determined. For example, if there are nodes such as "performance indicators" and "experimental data sets" on the shortest path from the starting node to the node associated with the comparison subtask, and the language feature information is closer to the node associated with the comparison subtask than the shortest path of other preset subtasks, then the possibility of the comparison subtask can be further confirmed.

[0046] Furthermore, as a refinement and extension of the specific implementation methods of the above-mentioned embodiments, in order to fully illustrate the specific implementation process of this embodiment, the sub-agent corresponding to the target subtask is called to execute the target subtask and determine the execution result of the target subtask, including: determining the target sub-agent corresponding to the target subtask and the target execution strategy of the target sub-agent according to the task type of the target subtask; calling the target sub-agent to execute the target subtask according to the target execution strategy and determining the execution result of the target subtask.

[0047] In this embodiment, the sub-agent is constructed based on the large language model and is fine-tuned according to the historical target sub-task and the preset document library. Specifically, for each task type, the historical target sub-task corresponding to the task type is collected, and the historical execution strategy and the final historical execution result corresponding to the historical target sub-task are recorded. Then, a target sub-task database is constructed according to the historical target sub-task, historical execution strategy, historical execution result of the historical target sub-task and the preset document library corresponding to the task type, and the preset large language model is fine-tuned using the target sub-task database to help the preset large language model fully learn the historical execution strategy through the academic field literature data, and better adapt to the sub-task, improve the efficiency and accuracy of task processing, so as to determine the preset large language model after fine-tuning the target sub-task database as the sub-agent corresponding to the task type, and the execution strategy learned by the sub-agent corresponding to the task type is determined as the execution strategy corresponding to the task type.

[0048] Thus, through the above-mentioned fine-tuning process, the correspondence between different task types and sub-agents and execution strategies is pre-established, so that different task types correspond to different sub-agents and execution strategies, that is, the task types are pre-correlated with the sub-agents and execution strategies. For each target sub-task decomposed from the task to be retrieved, the target sub-agent corresponding to the target sub-task and the target execution strategy of the target sub-agent are determined according to the task type of the target sub-task, so as to call the target sub-agent to execute the target sub-task according to the target execution strategy and determine the execution result of the target sub-task.

[0049] The execution strategy includes the specific methods and steps used by the sub-agent to achieve the goal of the sub-task. The historical target sub-task can be collected from the historical retrieval task processing records of various task types, or other data that can complete the fine-tuning of the sub-agent.

[0050] For example, the retrieval subtask corresponds to a retrieval sub-agent and a retrieval execution strategy, the comparison subtask corresponds to a comparison sub-agent and a comparison execution strategy, the summary subtask corresponds to a summary sub-agent and a summary execution strategy, the listing subtask corresponds to a listing sub-agent and a listing execution strategy, and the sequence subtask corresponds to a sequence sub-agent and a sequence execution strategy.

[0051] It should be noted that the sub-agent has the ability to learn and make decisions. Through learning algorithms and data analysis, the sub-agent can extract useful information from massive data to form its own knowledge base. In the decision-making process, it comprehensively considers various factors and uses logical reasoning, probability statistics and other methods to make the best decision.

[0052] Furthermore, as a refinement and extension of the specific implementation methods of the above-mentioned embodiment, in order to fully illustrate the specific implementation process of this embodiment, the target sub-agent is called to execute the target sub-task according to the target execution strategy to determine the execution result of the target sub-task, including: if the task type of the target sub-task is the first task type, performing semantic recognition processing on the target sub-task to determine the retrieval requirement information of the target sub-task; determining the target document data that meets the retrieval requirement information in the preset document library; determining the second target data in the target document data that meets the second preset condition; sorting the target document data to determine the document order; based on the document order, determining the execution result of the target sub-task according to the target document data and the second target data.

[0053] In this embodiment, for the retrieval subtask, that is, the target subtask of the first task type, massive real document data is used as the retrieval library of the retrieval sub-agent corresponding to the retrieval subtask, and real academic documents are cited to effectively reduce the risk of hallucinations that may be caused by large language models, thereby ensuring the reliability and traceability of the sub-agent execution results.

[0054] Specifically, if the target subtask is a retrieval subtask, that is, the task type of the target subtask is the first task type, the retrieval sub-agent (target sub-agent) executes the retrieval subtask according to the retrieval execution strategy (target execution strategy) to determine the execution result.

[0055] Exemplarily, in the search sub-agent, semantic recognition processing is performed on the search sub-task, key information in the search sub-task is extracted, and the search subject of the search sub-task, such as the document subject, document title, and keywords, is analyzed. Other related limiting conditions, such as year, author, journal, subject, research field, etc., are analyzed to determine the search requirement information of the search sub-task.

[0056] Next, the target document data is determined by searching the preset document library for documents that meet the search requirement information. For example, the document can be screened and filtered in multiple dimensions by the relevant limiting conditions given by the user in the search requirement information, such as filtering the search information by the year, author, journal, subject, research field, etc. specified by the user, to help users find documents that meet specific requirements more accurately.

[0057] Then, the second target data meeting the second preset condition is extracted from the target document data, for example, the author, publication date, journal name, keywords, abstract, etc. of the document are extracted, so that the user can quickly understand the main information of the target document data through the second target data.

[0058] Furthermore, the retrieved target document data is sorted according to the user's sorting intention in the search requirement information. For example, the documents are sorted by publication time. If there is no sorting intention in the search requirement information, the vector similarity between each document in the target document data and the search requirement information is calculated, and the target document data is sorted according to the vector similarity to determine the document order. Thus, according to the document order, the target document data and the second target data corresponding to the target document data are determined as the execution result of the search subtask to ensure that the most relevant and valuable documents are ranked first.

[0059] Furthermore, if the retrieval requirement information contains a need to obtain the full text of a document, the retrieval sub-agent can obtain the full text so that the target sub-task after the retrieval sub-task in the collaborative work strategy can perform reading questions and answers based on the specific content of the document.

[0060] Further, as a refinement and extension of the specific implementation methods of the above-mentioned embodiments, in order to fully illustrate the specific implementation process of this embodiment, the target sub-agent is called to execute the target sub-task according to the target execution strategy to determine the execution result of the target sub-task, including: if the task type of the target sub-task is the second task type, the input data is segmented, and the input data is determined according to the collaborative work strategy; based on the input data after the word segmentation processing, the keyword information of the input data is determined; based on the input data after the word segmentation processing, the topic information of the input data is determined using a preset topic model; based on the keyword information and topic information of the input data, the importance score of each sentence in the input data is determined; based on the importance score, the key sentence information of the input data is determined;

[0061] In this embodiment, if the target subtask is a summary subtask, that is, the task type of the target subtask is the second task type, the summary sub-agent (target sub-agent) executes the summary subtask according to the summary execution strategy (target execution strategy), converts the input data into concise and easy-to-understand core content, enhances the accuracy and comprehensibility of the information, helps users quickly grasp the core ideas and main points, and thus determines the execution results of the summary subtask.

[0062] It should be noted that the input data of the target sub-agent is the data input into the target sub-agent, which is determined according to the collaborative work strategy. For example, if the collaborative work strategy is a serial strategy, the input data of the target sub-agent is determined according to the execution result of the previous subtask. If the collaborative work strategy is a parallel strategy or a hybrid operation strategy, the input data is determined according to the data input into the target sub-agent. Here, the input data may be the execution result of the previous task or the target subtask itself.

[0063] Exemplarily, the input data is segmented into individual words or vocabulary units, and each word or vocabulary unit is labeled with a part of speech, such as noun, verb, adjective, etc.

[0064] Then, extract the keywords in the input data after word segmentation. Exemplarily, the TF-IDF (TermFrequency-InverseDocumentFrequency) algorithm is used to extract keywords. Specifically, for each word or vocabulary unit in the input data, the number of times it appears in the input data is counted to determine the first word frequency of each word or vocabulary unit. For example, if the word "artificial intelligence" appears once in the input data, assuming that the input data has a total of 10 words, then the first word frequency TF of "artificial intelligence" is 1 / 10=0.1. Next, retrieve multiple documents related to the input data in the preset document library to determine the inverse document data of the input data. Then, for each word or vocabulary unit in the input data, count how many documents in the inverse document data include the word or vocabulary unit, so as to calculate the second word frequency IDF of each word or vocabulary unit. Among them, the calculation formula of the second word frequency IDF is: IDF=log (the total number of documents included in the inverse document data / the number of documents containing the word or vocabulary unit in the inverse document data). For example, assuming that there are 100 documents in the inverse document data, 10 of which contain the word "artificial intelligence". In this example, the second word frequency IDF of "artificial intelligence" = log(100 / 10) = log(10). Therefore, the first word frequency of each word or vocabulary unit is multiplied by the second word frequency to obtain the target word frequency TF-IDF of each word or vocabulary unit. For example, the TF-IDF value of "artificial intelligence" = TF×IDF=0.1×log(10).

[0065] Sort all words or vocabulary units in the input data in descending order according to the target word frequency TF-IDF. Determine the keyword information of the input data based on the set threshold or by selecting the first several words or vocabulary units as keywords. Among them, words or vocabulary units with larger TF-IDF values ​​are more important in the input data, and they appear relatively less in the entire input data, so they have higher discrimination and can reflect the main content and characteristics of the input data.

[0066] Furthermore, based on the input data after word segmentation processing, a preset topic model is used to determine the topic information of the input data. Exemplarily, the LDA (Latent Dirichlet Allocation) model is used to identify potential topics from the input data. For example, before the LDA model is used, the LDA model is trained using a preset document library so that the LDA model is suitable for the academic field and the topic-vocabulary distribution in the academic field is obtained. The input data after word segmentation processing is input into the LDA model, and the LDA model randomly initializes a topic distribution for the input data. Assuming that the trained LDA model has K topics, a random probability value can be assigned to each topic for the LDA model during initialization so that the sum of these probability values ​​is 1.

[0067] Then, the topic distribution of the input data is iteratively updated, including E-step and M-step.

[0068] Among them, step E includes: for each word or vocabulary unit in the input data, calculate the posterior probability of the word or vocabulary unit under each topic, and the posterior probability is calculated based on the topic distribution of the input data and the trained topic-vocabulary distribution. Specifically, according to Bayes' theorem, the probability P(topic k|vocabulary w) that the word w in the input data belongs to the topic k can be calculated using the formula P(topic k|vocabulary w) ∝ P(vocabulary w|topic k) × P(topic k). Among them, P(vocabulary w|topic k) is obtained from the trained topic-vocabulary distribution, which represents the probability of the word w appearing under the topic k; P(topic k) is the probability distribution of the input data on the topic k. By summing or averaging the posterior probabilities of all words or vocabulary units in the input data, a new probability distribution of the input data on each topic can be obtained.

[0069] Step M includes: updating the topic assignment of each word or vocabulary unit in the input data according to the new topic distribution of the input data obtained in step E. Specifically, for each word or vocabulary unit in the input data, the word or vocabulary unit is assigned to the topic with the highest probability according to its posterior probability under each topic.

[0070] Repeat steps E and M until the topic distribution of the input data converges or reaches the preset number of iterations, and obtain the probability distribution of the input data on each topic, reflecting the degree of correlation between the input data and different topics, so as to determine the topic information of the input data according to the topics with higher correlation.

[0071] Furthermore, according to the keyword information and topic information of the input data, combined with the position of each sentence in the input data (such as the sentences in the first and last paragraphs are more important), the number of keywords contained, the relevance to the topic and other factors, the importance of each sentence in the input data is evaluated, and the importance score of each sentence in the input data is determined. Then, according to the importance score of the sentence, several most important sentences are selected as the basis for the summary to determine the key sentence information. Next, natural language processing technology, such as syntactic analysis and semantic analysis, is used to identify the core components in the key sentence information, remove modifying words and phrases, make the sentences more concise and clear, and thus update the key sentence information. The updated key sentence information is combined and sorted according to the logical structure and topic of the input data to form a coherent summary and determine the initial execution result. Finally, the grammatical errors in the initial execution result are checked, the use of vocabulary is adjusted, and the expression of the sentence is improved to make the summary language more fluent and natural, so as to obtain a concise and easy-to-understand execution result.

[0072] Further, as a refinement and extension of the specific implementation methods of the above-mentioned embodiment, in order to fully illustrate the specific implementation process of this embodiment, the target sub-agent is called to execute the target sub-task according to the target execution strategy, and the execution result of the target sub-task is determined, including: if the task type of the target sub-task is the third task type, the target sub-task is semantically recognized and processed to determine the first task intention information of the target sub-task; according to the first task intention information, the comparison sub-data in the input data is determined, and the input data is determined according to the collaborative work strategy; the comparison sub-data is word-segmented, and the keyword information of the comparison sub-data is determined according to the comparison sub-data after the word segmentation processing; according to the keyword information of the comparison sub-data, the first comparison information between the comparison sub-data is determined; according to the comparison sub-data after the word segmentation processing, the topic information of the comparison sub-data is determined using a preset topic model; according to the topic information of the comparison sub-data, the second comparison information between the comparison sub-data is determined; based on the first task intention information, the target comparison information between the comparison sub-data is determined according to the first comparison information and the second comparison information; the execution result of the target sub-task is determined according to the target comparison information.

[0073] In this embodiment, if the target subtask is a comparison subtask, that is, the task type of the target subtask is the third task type, the comparison sub-agent (target sub-agent) performs the comparison subtask according to the comparison execution strategy (target execution strategy), compares the input data containing multiple articles or paragraphs, and systematically compares different academic information to help users understand the characteristics, advantages and disadvantages of the input data, and find out the similarities and differences between them.

[0074] Exemplarily, in the comparison sub-agent, semantic recognition processing is performed on the comparison sub-task, the first task intention information of the target sub-task is determined, and the input data is divided into comparison sub-data according to the first task intention information. For example, if the first task intention information is to compare different documents, each document is determined in the input data. If the first task intention information is to compare different paragraphs, each paragraph is determined in the input data.

[0075] Next, the keyword information and theme information of the comparison sub-data are extracted in the comparison sub-agent in the same manner as the process of determining the keyword information and theme information in the summary sub-task in the above embodiment. Thus, the keywords of different comparison sub-data are compared to find out the common keywords and unique keywords, wherein the common keywords represent the similarities between the comparison sub-data and the unique keywords reflect the characteristics of the comparison sub-data, thereby determining the first comparison information.

[0076] At the same time, the topic distributions of different comparison sub-data are compared, and the similarities between the topic distributions of different comparison sub-data are calculated using methods such as cosine similarity to determine their similarities and differences in topics, thereby obtaining the second comparison information.

[0077] You can also use text summarization algorithms, such as extractive summarization or generative summarization methods, to generate a concise summary for each comparison sub-data, summarizing the main content of the comparison sub-data. Then, compare the summaries of different comparison sub-data sentence by sentence, use text similarity algorithms to measure the similarity between the summaries, find similar expressions and different viewpoints, and obtain the third comparison information.

[0078] Then, based on the first task intention information, the purpose of comparison is determined, for example, the characteristics, advantages or disadvantages of each comparison sub-data are compared, or the differences between the comparison sub-data are compared. Therefore, for the purpose of comparison, the first comparison information, the second comparison information and the third comparison information are integrated to extract valuable insights or conclusions, determine the target comparison information between the comparison sub-data, and then determine the target comparison information as the execution result of the comparison sub-task. For example, if a document has more unique keywords and in-depth discussions on a specific topic, it can be considered that it has an advantage on the topic; if the keywords and topics of a document are relatively simple compared with other documents, then there may be deficiencies. Alternatively, the differences and commonalities of different documents in research methods, experimental results, theoretical viewpoints, etc. can also be obtained.

[0079] Furthermore, as a refinement and expansion of the specific implementation methods of the above-mentioned embodiments, in order to fully illustrate the specific implementation process of this embodiment, the target sub-agent is called to execute the target sub-task according to the target execution strategy to determine the execution result of the target sub-task, including: if the task type of the target sub-task is the fourth task type, performing semantic recognition processing on the target sub-task to determine the second task intention information of the target sub-task; determining the execution result of the target sub-task based on the data in the input data that conforms to the second task intention information, and the input data is determined according to the collaborative work strategy.

[0080] In this embodiment, if the target subtask is a listing subtask, that is, the task type of the target subtask is the fourth task type, the listing sub-agent (target sub-agent) executes the listing subtask according to the listing execution strategy (target execution strategy), accurately identifies the user's intention, and systematically organizes and presents the required information for the user, so as to facilitate the user to quickly obtain and browse related content.

[0081] For example, in the enumeration sub-agent, semantic recognition processing is performed on the enumeration sub-task to identify the relevant documents, viewpoints, and concepts required by the user, and determine the second task intention information of the target sub-task. Then, in the input data, the data that meets the second task intention information is filtered and irrelevant data is eliminated, so as to output concise and clear enumeration information according to the target sub-task, avoid redundant content, and enable users to quickly understand and obtain the required information.

[0082] For example, if the task to be searched is "requires related concepts of artificial intelligence", the task to be searched will be decomposed into a search subtask and a list subtask. The search sub-agent searches for documents related to artificial intelligence in the preset document library according to the search subtask, and inputs the retrieved documents into the list sub-agent. The list sub-agent inputs the concepts of artificial intelligence in a list.

[0083] Furthermore, as a refinement and extension of the specific implementation methods of the above-mentioned embodiments, in order to fully illustrate the specific implementation process of this embodiment, the target sub-agent is called to execute the target subtask according to the target execution strategy to determine the execution result of the target subtask, including: if the task type of the target subtask is the fifth task type, semantic recognition processing is performed on the target subtask to determine the third task intention information of the target subtask and the sorting information corresponding to the third task intention information; according to the sorting information, the data in the input data that meets the third task intention information is sorted, and the input data is determined according to the collaborative work strategy; based on the sorted data that meets the third task intention information, the execution result of the target subtask is determined.

[0084] In this embodiment, if the target subtask is a sequence subtask, that is, the task type of the target subtask is the fifth task type, the sequence sub-agent (target sub-agent) executes the sequence subtask according to the sequence execution strategy (target execution strategy), accurately identifies the user's intention, and presents information on the events, steps or processes mentioned in the user's intention in a time or logical order, highlighting the sequence of events, steps or processes.

[0085] Exemplarily, in the sequence sub-agent, semantic recognition processing is performed on the sequence sub-task, the purpose of the sequence sub-task is understood, and the third task intention information of the sequence sub-task is determined. In addition, the sorting order required by the sequence sub-task is clarified, such as time series events, step order, or logical relationship order, and the sorting information corresponding to the third task intention information is determined.

[0086] Furthermore, natural language processing technology is used to identify various events, actions or steps in the input data that meet the third task intention information. For example, the sentence structure is analyzed to identify which parts of the input data describe the occurrence of actions or events that meet the third task intention information. At the same time, the relationship between various events, actions or steps is established. For example, if the description of time or date is involved, the time information is identified and a basis for sorting is provided. For example, time nodes such as "2020" or "last year" are identified, and the time sequence of events, actions or steps is inferred based on this information. Alternatively, the causal relationship between events, actions or steps is analyzed to determine which event, action or step is the antecedent and which event, action or step is the consequence. For example, it is identified that "first, the researcher proposed a hypothesis; then, the hypothesis was verified through experiments; finally, a conclusion was drawn". Alternatively, the order of events, actions or steps is inferred by analyzing the sequential words described in the text (such as "first", "second", "next", "finally", etc.).

[0087] Then, according to the sorting information, the data in the input data that meets the third task intention information is sorted to obtain the execution result of the sequence subtask. For example, if the sorting information requires chronological arrangement, the sequence sub-agent will sort the events, actions or steps according to the time nodes when the events, actions or steps occur. For example, different historical events or research progress mentioned in the literature are arranged in a "timeline" manner. If the sorting information requires arrangement according to logical relationships, the sequence sub-agent will sort the information according to cause and effect relationships, step sequence, etc. For example, in the process of describing an experiment, it is arranged in the order of "hypothesis proposal → experimental design → data collection → result analysis". This ensures that the execution results are concise and clear, and highlights the order in which events, actions or steps occur, helping users quickly understand the evolution of events, actions or steps.

[0088] It should be noted that in the sub-agent, the input data will also be pre-processed, including cleaning, removal of irrelevant information, removal of duplicate text, etc., to ensure that the input data can be recognized and applied.

[0089] Furthermore, as a refinement and expansion of the specific implementation methods of the above-mentioned embodiment, in order to fully illustrate the specific implementation process of this embodiment, step 103, that is, determining the retrieval results of the task to be retrieved based on the execution results, specifically includes: post-processing the execution results of each target subtask, optimizing readability, optimizing grammar, adjusting language fluency, ensuring that the natural language expression of the final output is clear and understandable, removing repeated or irrelevant information generated during the generation process, and ensuring that the output is concise and clear.

[0090] It should be noted that the size of the serial numbers of the steps in the above embodiments does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0091] Furthermore, if Figure 2 As shown, as a specific implementation of the above-mentioned multi-agent-based document retrieval and reading method, an embodiment of the present application provides a multi-agent-based document retrieval and reading device 200, and the multi-agent-based document retrieval and reading device 200 includes: a decomposition module 201, a calling module 202 and a determination module 203.

[0092] The decomposition module 201 is used to decompose the task to be searched into target subtasks by using a meta-agent and determine the collaborative working strategy between the target subtasks, wherein the meta-agent is constructed based on a large language model and fine-tuned according to historical search tasks and a preset document library;

[0093] The calling module 202 is used to call the sub-agent corresponding to the target sub-task to execute the target sub-task according to the collaborative work strategy, and determine the execution result of the target sub-task, wherein the sub-agent is constructed based on the large language model and fine-tuned according to the historical target sub-task and the preset document library;

[0094] The determination module 203 is used to determine the search result of the task to be searched according to the execution result.

[0095] In one embodiment, the multi-agent-based document retrieval and reading device 200 further includes:

[0096] The fine-tuning module is used to label the historical retrieval task according to the historical intention and determine the first label of the historical retrieval task; label the historical retrieval task according to the historical collaborative work strategy and determine the second label of the historical retrieval task; and fine-tune the preset large language model according to the preset document library and the historical retrieval task carrying the first label and the second label to obtain a meta-agent.

[0097] The decomposition module 201 is specifically used to perform semantic recognition processing on the task to be retrieved, determine the semantic feature information of the task to be retrieved; match the semantic feature information with preset subtasks of different task types, and determine the preset subtask with matching semantic feature information as the target subtask of the task to be retrieved.

[0098] Decomposition module 201 is specifically used to identify first target data that meets first preset conditions in a preset document library; determine the semantic relationship between the first target data based on the preset document library; construct a document semantic network diagram based on the first target data and the semantic relationship; map the semantic feature information to the nodes in the document semantic network diagram, and match the semantic feature information with the preset subtasks corresponding to the nodes through a semantic reasoning algorithm.

[0099] The calling module 202 is specifically used to determine the target sub-agent corresponding to the target sub-task and the target execution strategy of the target sub-agent according to the task type of the target sub-task; call the target sub-agent to execute the target sub-task according to the target execution strategy, and determine the execution result of the target sub-task.

[0100] Calling module 202 is specifically used to perform semantic recognition processing on the target subtask if the task type of the target subtask is the first task type, and determine the retrieval requirement information of the target subtask; determine the target document data that meets the retrieval requirement information in the preset document library; determine the second target data that meets the second preset condition in the target document data; sort the target document data to determine the document order; based on the document order, determine the execution result of the target subtask according to the target document data and the second target data.

[0101] Calling module 202 is specifically used to, if the task type of the target subtask is the second task type, perform word segmentation on the input data, and the input data is determined according to the collaborative work strategy; determine the keyword information of the input data according to the input data after the word segmentation; determine the topic information of the input data using a preset topic model according to the input data after the word segmentation; determine the importance score of each sentence in the input data according to the keyword information and topic information of the input data; determine the key sentence information of the input data according to the importance score; perform semantic recognition processing on the key sentence information to determine the core information of the key sentence information; update the key sentence information according to the core information; optimize the updated key sentence information to determine the execution result of the target subtask.

[0102] The calling module 202 is specifically used to perform semantic recognition processing on the target subtask if the task type of the target subtask is the third task type, and determine the first task intention information of the target subtask; determine the comparison sub-data in the input data according to the first task intention information, and the input data is determined according to the collaborative work strategy; perform word segmentation processing on the comparison sub-data, and determine the keyword information of the comparison sub-data according to the comparison sub-data after the word segmentation processing; determine the first comparison information between the comparison sub-data according to the keyword information of the comparison sub-data; determine the topic information of the comparison sub-data according to the comparison sub-data after the word segmentation processing using a preset topic model; determine the second comparison information between the comparison sub-data according to the topic information of the comparison sub-data; based on the first task intention information, determine the target comparison information between the comparison sub-data according to the first comparison information and the second comparison information; determine the execution result of the target subtask according to the target comparison information.

[0103] Calling module 202 is specifically used to perform semantic recognition processing on the target subtask if the task type of the target subtask is the fourth task type, and determine the second task intention information of the target subtask; determine the execution result of the target subtask based on the data in the input data that conforms to the second task intention information, and the input data is determined according to the collaborative work strategy.

[0104] Calling module 202 is specifically used to perform semantic recognition processing on the target subtask if the task type of the target subtask is the fourth task type, and determine the second task intention information of the target subtask; determine the execution result of the target subtask based on the data in the input data that conforms to the second task intention information, and the input data is determined according to the collaborative work strategy.

[0105] The specific definition of the document retrieval and reading device based on multi-agents can be found in the definition of the document retrieval and reading method based on multi-agents above, which will not be repeated here. Each module in the above-mentioned document retrieval and reading device based on multi-agents can be implemented in whole or in part by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory in the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0106] Based on the above Figure 1 The method shown in the embodiment of the present application accordingly provides a readable storage medium on which a computer program is stored, and when the program is executed by a processor, the above-mentioned Figure 1 The multi-agent based literature retrieval and reading method shown.

[0107] Based on this understanding, the technical solution of the present application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which can be a CD-ROM, a USB flash drive, a mobile hard disk, etc.), and includes a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each implementation scenario of the present application.

[0108] Based on the above Figure 1 The method shown, and Figure 2 In order to achieve the above-mentioned purpose, the embodiment of the present application further provides a computer device, which can be a personal computer, a server, a network device, etc. The computer device includes a storage medium and a processor; the storage medium is used to store a computer program; the processor is used to execute the computer program to achieve the above-mentioned Figure 1 The multi-agent based literature retrieval and reading method shown.

[0109] Optionally, the computer device may further include a user interface, a network interface, a camera, a radio frequency (RF) circuit, a sensor, an audio circuit, a WI-FI module, etc. The user interface may include a display, an input unit such as a keyboard, etc., and the optional user interface may also include a USB interface, a card reader interface, etc. The network interface may optionally include a standard wired interface, a wireless interface (such as a Bluetooth interface, a WI-FI interface), etc.

[0110] Those skilled in the art will appreciate that the computer device structure provided in this embodiment does not limit the computer device, and may include more or fewer components, or a combination of certain components, or different component arrangements.

[0111] The storage medium may also include an operating system and a network communication module. The operating system is a program that manages and saves the hardware and software resources of the computer device, and supports the operation of information processing programs and other software and / or programs. The network communication module is used to realize communication between the components inside the storage medium, and communication with other hardware and software in the physical device.

[0112] Through the description of the above implementation methods, those skilled in the art can clearly understand that the present application can be implemented by means of software plus a necessary general hardware platform, or the embodiments of the present application can be implemented by hardware.

[0113] Those skilled in the art will appreciate that the accompanying drawings are only schematic diagrams of a preferred implementation scenario, and the modules or processes in the accompanying drawings are not necessarily necessary for implementing the present application. Those skilled in the art will appreciate that the modules in the devices in the implementation scenario can be distributed in the devices of the implementation scenario according to the description of the implementation scenario, or can be changed accordingly and located in one or more devices different from the present implementation scenario. The modules of the above-mentioned implementation scenario can be combined into one module, or can be further split into multiple submodules.

[0114] The above serial numbers of this application are only for description and do not represent the advantages and disadvantages of the implementation scenarios. The above disclosure is only a few specific implementation scenarios of this application, but this application is not limited to them, and any changes that can be thought of by technicians in this field should fall within the scope of protection of this application.

Claims

1. A multi-agent based literature retrieval and reading method, characterized in that: The method comprises: Using a meta-agent to decompose the task to be searched into target sub-tasks and determine the collaborative working strategy between the target sub-tasks, wherein the meta-agent is constructed based on a large language model and fine-tuned according to historical search tasks and a preset document library; According to the collaborative work strategy, calling the sub-agent corresponding to the target sub-task to execute the target sub-task, and determining the execution result of the target sub-task, wherein the sub-agent is constructed based on a large language model and fine-tuned according to the historical target sub-task and the preset document library; According to the execution result, the search result of the task to be searched is determined.

2. The multi-agent based document retrieval and reading method according to claim 1, characterized in that: The method comprises: Marking according to the historical intent of the historical retrieval task to determine a first label of the historical retrieval task; Marking according to the historical collaborative work strategy of the historical retrieval task to determine a second label of the historical retrieval task; According to the preset document library and the historical retrieval task carrying the first label and the second label, the preset large language model is fine-tuned to obtain the meta-agent.

3. The multi-agent based document retrieval and reading method according to claim 1, characterized in that: The method of using a meta-agent to decompose the task to be retrieved into target subtasks includes: Performing semantic recognition processing on the task to be retrieved to determine semantic feature information of the task to be retrieved; The semantic feature information is matched with preset subtasks of different task types, and the preset subtask that matches the semantic feature information is determined as the target subtask of the task to be retrieved.

4. The multi-agent based document retrieval and reading method according to claim 3, characterized in that: The matching of the semantic feature information with a preset subtask includes: Identifying first target data in the preset document library that meets a first preset condition; Determining the semantic relationship between the first target data based on the preset document library; Constructing a document semantic network diagram according to the first target data and the semantic relationship; The semantic feature information is mapped to nodes in the document semantic network diagram, and the semantic feature information is matched with the preset subtask corresponding to the node through a semantic reasoning algorithm.

5. The multi-agent based document retrieval and reading method according to claim 1, characterized in that: The calling of the sub-agent corresponding to the target sub-task to execute the target sub-task and determining the execution result of the target sub-task includes: Determine, according to the task type of the target subtask, the target sub-agent corresponding to the target subtask and the target execution strategy of the target sub-agent; The target sub-agent is called to execute the target sub-task according to the target execution strategy, and the execution result of the target sub-task is determined.

6. The multi-agent-based document retrieval and reading method according to claim 5, characterized in that: The calling of the target sub-agent to execute the target sub-task according to the target execution strategy and determining the execution result of the target sub-task includes: If the task type of the target subtask is the first task type, performing semantic recognition processing on the target subtask to determine retrieval requirement information of the target subtask; Determining target document data that meets the search requirement information in the preset document library; Determining second target data in the target document data that meets a second preset condition; Sorting the target document data to determine the document sequence; Based on the document sequence, the execution result of the target subtask is determined according to the target document data and the second target data.

7. The multi-agent-based document retrieval and reading method according to claim 5, characterized in that: The calling of the target sub-agent to execute the target sub-task according to the target execution strategy and determining the execution result of the target sub-task includes: If the task type of the target subtask is the second task type, performing word segmentation processing on the input data, the input data being determined according to the collaborative work strategy; Determining keyword information of the input data according to the input data after word segmentation processing; Determine the topic information of the input data using a preset topic model according to the input data after word segmentation processing; Determining the importance score of each sentence in the input data according to the keyword information and the subject information of the input data; Determining key sentence information of the input data according to the importance score; Performing semantic recognition processing on the key sentence information to determine core information of the key sentence information; Update the key sentence information according to the core information; The updated key sentence information is optimized to determine the execution result of the target subtask.

8. The multi-agent based document retrieval and reading method according to claim 5, characterized in that: The calling of the target sub-agent to execute the target sub-task according to the target execution strategy and determining the execution result of the target sub-task includes: If the task type of the target subtask is the third task type, performing semantic recognition processing on the target subtask to determine the first task intention information of the target subtask; Determine, according to the first task intention information, comparison sub-data in the input data, wherein the input data is determined according to the collaborative work strategy; Performing word segmentation processing on the comparison sub-data, and determining keyword information of the comparison sub-data according to the comparison sub-data after the word segmentation processing; determining first comparison information between the comparison sub-data according to the keyword information of the comparison sub-data; Determining topic information of the comparison sub-data using a preset topic model according to the comparison sub-data after word segmentation processing; determining second comparison information between the comparison sub-data according to the subject information of the comparison sub-data; Based on the first task intention information, determining target comparison information between the comparison sub-data according to the first comparison information and the second comparison information; The execution result of the target subtask is determined according to the target comparison information.

9. The multi-agent based document retrieval and reading method according to claim 5, characterized in that: The calling of the target sub-agent to execute the target sub-task according to the target execution strategy and determining the execution result of the target sub-task includes: If the task type of the target subtask is the fourth task type, performing semantic recognition processing on the target subtask to determine the second task intention information of the target subtask; The execution result of the target subtask is determined according to the data in the input data that conforms to the second task intention information, and the input data is determined according to the collaborative work strategy.

10. The multi-agent based document retrieval and reading method according to claim 5, characterized in that: The calling of the target sub-agent to execute the target sub-task according to the target execution strategy and determining the execution result of the target sub-task includes: If the task type of the target subtask is the fifth task type, performing semantic recognition processing on the target subtask to determine the third task intent information of the target subtask and the sorting information corresponding to the third task intent information; sorting, according to the sorting information, data in the input data that conforms to the third task intention information, the input data being determined according to the collaborative work strategy; The execution result of the target subtask is determined according to the sorted data that meets the third task intention information.

Images