Literature interpretation method and device based on multi-Agent collaboration, medium and program product

Abstract

The embodiment of the invention discloses a literature interpretation method and device based on multi-Agent collaboration, a medium and a program product. The method comprises the following steps: performing feature extraction on a target document, generating a document feature vector, screening a target processing template matched with the target processing template from a preset processing template set, and constructing a directed acyclic graph by using the target processing template; key information of the target literature is extracted and stored in a shared work area; on the basis of the execution sequence of the directed acyclic graph, calling a corresponding functional agent to read input data from the shared working area, executing a processing task on the basis of the input data, and writing an execution result back to the shared working area; and after the execution of each functional agent is completed, outputting a literature interpretation report integrated by the execution results stored in the shared working area. According to the method, dynamic adaptation of a literature interpretation strategy can be realized through feature-driven dynamic scheduling and a multi-Agent cooperation mechanism, flexible expansion of system functions is supported, and the intelligent level and comprehensive efficiency of complex literature interpretation are greatly improved.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of artificial intelligence, and particularly relates to a literature interpretation method and device based on multi-agent cooperation, a medium and a program product. BACKGROUND

[0002] With the continuous activity of global scientific research innovation, the number of scientific research literature is growing exponentially. The academic value and knowledge resources contained in the massive literature are increasingly rich, but also make researchers fall into the predicament of information overload. Efficient and accurate extraction of key information and mining of knowledge association from literature have become one of the core needs of scientific research work. In this context, literature interpretation technology emerges as the times require. The existing technology relies on computer vision, natural language processing and other technologies to build literature processing systems to realize the basic functions of structured analysis and information extraction of literature, and to help researchers simplify the literature reading and analysis process.

[0003] However, the existing literature processing system is difficult to meet the complex literature interpretation requirements. It lacks flexibility in the processing flow and cannot dynamically adjust the analysis strategy according to the type and content of the literature. Moreover, the system has insufficient scalability and is difficult to flexibly access new functions or adapt to the processing requirements of different literature types. In addition, the intelligent level is limited and lacks autonomous task planning and collaborative decision-making mechanism. SUMMARY

[0004] Therefore, the embodiments of the present disclosure provide a literature interpretation method and device based on multi-agent cooperation, a medium and a program product, which can realize dynamic adaptation of literature interpretation strategy through feature-driven dynamic scheduling and multi-agent cooperation mechanism, support flexible expansion of system functions, and greatly improve the intelligent level and comprehensive performance of complex literature interpretation.

[0005] In a first aspect, the embodiments of the present disclosure provide a literature interpretation method based on multi-agent cooperation, which adopts the following technical scheme: Obtaining a target literature, performing feature extraction on the target literature, and generating a literature feature vector; From a preset processing template set, a target processing template matching the literature feature vector is selected; Based on the target processing template, a directed acyclic graph is constructed, wherein the nodes in the directed acyclic graph represent function agents, and the edges between the nodes represent the dependency relationship between the function agents; Extracting key information of the target literature and storing it in a shared working area; Based on the execution order of the directed acyclic graph, the corresponding function agent reads the input data from the shared working area, performs a processing task based on the input data, and writes the execution result back to the shared working area; When the function agents in the directed acyclic graph are executed, an integrated document interpretation report is outputted by the execution results of the function agents stored in the shared workspace.

[0006] Optionally, the filtering of the target processing template matching the document feature vector from the preset processing template set comprises: The corresponding exclusive feature vector is preset for each processing template in the preset processing template set. The similarity between the exclusive feature vector of each processing template and the document feature vector is obtained. The processing template with the highest similarity to the document feature vector is selected as the target processing template.

[0007] Optionally, the presetting of the corresponding exclusive feature vector for each processing template in the preset processing template set comprises: The document types are divided according to the interpretation requirements, and the exclusive feature dimension system of each document type is defined, and each document type has an exclusive processing template. The representative samples of each document type are collected, and the feature data of the representative samples are extracted according to the exclusive feature dimension system. The feature data of the representative samples of each document type is statistically analyzed according to the dimension, and the standard features of each document type are obtained. The standard features of each document type are combined in a unified dimension order to generate the exclusive feature vector of the corresponding processing template.

[0008] Optionally, the construction of the directed acyclic graph based on the target processing template comprises: The processing step sequence pre-configured in the target processing template is read. The function agents corresponding to each step in the processing step sequence are parsed, and the input data constraint conditions and output data format information of each function agent are extracted. The dependency relationship between each step in the processing step sequence is determined through data flow conversion relationship analysis. The function agents are mapped to the nodes of the directed acyclic graph, and the directed edges are constructed between the nodes with data input and output association based on the dependency relationship. The input data constraint conditions and output data format information are written into the attribute field of the corresponding node as node attributes to generate the directed acyclic graph.

[0009] Optionally, the extraction of the key information of the target document and the storage in the shared workspace comprise: The key information of the target document is extracted based on the input data constraint conditions and output data format information in the node attributes of the directed acyclic graph. The extracted key information is identified by type, and based on the storage characteristics of different data types, the key information is classified and stored in the corresponding storage partition of the shared workspace. The storage paths and index identifiers of various key information in the shared work area are synchronously written into the key-value pair storage area with the task session ID as the primary key, and then associated and bound with the task information of the corresponding functional intelligent agent node.

[0010] Optionally, the step of invoking the corresponding functional agent to read input data from the shared workspace based on the execution order of the directed acyclic graph, executing processing tasks based on the input data, and writing the execution results back to the shared workspace includes: Based on the topological execution order of the directed acyclic graph, a message instruction containing data access information is generated and sent to the corresponding functional agent; After receiving a message instruction, the functional intelligent agent retrieves matching input data from the shared workspace to execute a preset processing task, and writes the execution result back to the specified storage location in the shared workspace. After the execution result is written back, the state of the corresponding directed acyclic graph node is updated to "completed". When the current node status is detected as completed, the next functional agent is determined based on the directed acyclic graph and a message instruction is sent to the next functional agent.

[0011] Optionally, the document interpretation method based on multi-agent collaboration further includes: All functional agents are divided into layers according to processing stages and data flow levels; Within each layer, groups are formed based on functional similarity and data dependencies, and functional agents in the same group share a dedicated message queue. Construct a message routing table, which stores the mapping relationship between the node identifiers of the directed acyclic graph and the hierarchy, group, and corresponding message queue of the functional agents; Based on the message routing table, the target functional agent is matched with its level and group by the target node identifier carried in the message instruction, and the message instruction is published to the corresponding message queue. The same group of functional agents listen to and consume message instructions in the corresponding message queue in parallel.

[0012] Secondly, this disclosure also provides a document interpretation system based on multi-agent collaboration, employing the following technical solution: The feature extraction module is used to acquire target documents, extract features from the target documents, and generate document feature vectors. The template filtering module is used to filter target processing templates that match the document feature vector from a preset set of processing templates; The graph construction module is used to construct a directed acyclic graph based on the target processing template. Nodes in the directed acyclic graph represent functional agents, and edges between nodes represent dependencies between functional agents. The information extraction module is used to extract key information from the target document and store it in the shared workspace; The task execution module is used to call the corresponding functional agent to read input data from the shared work area based on the execution order of the directed acyclic graph, execute processing tasks based on the input data, and write the execution results back to the shared work area. The report output module is used to output a literature interpretation report, which integrates the execution results of each functional agent stored in the shared workspace, after each functional agent in the directed acyclic graph has completed its execution.

[0013] Thirdly, this disclosure also provides a computer device, which adopts the following technical solution: The computer device includes: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, which enables the at least one processor to perform any of the above-described multi-agent collaborative document interpretation methods.

[0014] Fourthly, embodiments of this disclosure also provide a computer-readable storage medium storing computer instructions for causing a computer to execute any of the above-described multi-agent collaborative document interpretation methods.

[0015] Fifthly, embodiments of this disclosure also provide a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of any of the methods described above.

[0016] The document interpretation method based on multi-agent collaboration provided in this disclosure, through a pre-screening mechanism of feature extraction and template matching, can accurately anchor the processing logic adapted to the target document, avoiding ineffective task allocation and resource waste in the multi-agent collaboration process. Specifically, the document feature vector can quantitatively represent the core attributes of the document, such as its topic, domain, and complexity. Screening the target processing template based on this vector ensures that the subsequently constructed directed acyclic graph (DAG) highly matches the document interpretation requirements. Simultaneously, the DAG, by defining functional agents through nodes and dependencies through edges, enables orderly collaboration among multiple agents, avoiding the low interpretation efficiency caused by task conflicts and dependency chaos in traditional parallel processing, significantly improving the standardization and execution efficiency of the document interpretation process. This method, relying on a shared workspace and directed execution architecture, ensures the consistency and integrity of information flow during multi-agent collaboration, improving the accuracy and comprehensiveness of the document interpretation report. On the one hand, key information is centrally stored in a shared workspace, allowing each functional agent to read input data as needed. This avoids redundant operations of repeatedly extracting document information between agents and prevents data deviations caused by scattered information storage. On the other hand, calling agents and writing back results according to the execution order of a directed acyclic graph ensures that each agent's processing task closely aligns with the core needs of the document. The resulting integrated interpretation report covers multi-level information of the document, preserving the accuracy of key content while providing comprehensive multi-dimensional analysis, meeting the practical needs of interpreting complex documents. Furthermore, because this solution is based on feature matching to select target processing templates and using a directed acyclic graph to define agent collaboration logic, new document types and interpretation needs can be quickly adapted by adding processing templates or agent nodes without reconstructing the entire system, thus supporting flexible functional expansion. Multiple agents collaboratively execute tasks in an orderly manner according to dependencies, achieving unified information flow and reuse through a shared workspace. This replaces the traditional fixed-process processing model, improving the intelligent collaborative decision-making capability of the interpretation process, reducing redundant operations, and significantly enhancing the overall efficiency of interpreting complex documents.

[0017] The above description is merely an overview of the technical solution disclosed herein. In order to better understand the technical means of this disclosure and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this disclosure more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of this disclosure, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A flowchart illustrating the document interpretation method based on multi-agent collaboration provided in this embodiment of the disclosure; Figure 2 A flowchart illustrating the target processing template filtering method provided in this embodiment of the disclosure; Figure 3 A flowchart illustrating the proprietary feature vector generation method provided in this embodiment of the disclosure; Figure 4 A flowchart illustrating the method for obtaining the similarity between a proprietary feature vector and a document feature vector provided in this embodiment of the disclosure; Figure 5 A schematic flowchart illustrating the method for constructing a directed acyclic graph provided in this embodiment of the disclosure; Figure 6 A flowchart illustrating the key information storage method provided in the embodiments of this disclosure; Figure 7 A flowchart illustrating the functional agent invocation method provided in this embodiment of the disclosure; Figure 8 A flowchart illustrating the message instruction sending and receiving method provided in this embodiment of the disclosure; Figure 9 A flowchart illustrating the typical functional intelligent agent workflow provided in the embodiments of this disclosure; Figure 10 A schematic diagram of the principle of a document interpretation system based on multi-agent collaboration provided in this embodiment of the disclosure; Figure 11 This is a schematic diagram of the structure of a computer device provided in an embodiment of the present disclosure. Detailed Implementation

[0020] The embodiments of this disclosure will now be described in detail with reference to the accompanying drawings.

[0021] It should be understood that the following specific examples illustrate the implementation of this disclosure, and those skilled in the art can easily understand other advantages and effects of this disclosure from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of this disclosure, and not all of them. This disclosure can also be implemented or applied through other different specific implementation methods, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this disclosure. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0022] It should be noted that various aspects of embodiments within the scope of the appended claims are described below. It will be apparent that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this disclosure, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or method can be implemented using structures and / or functionalities other than one or more of the aspects set forth herein.

[0023] It should also be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of this disclosure. The drawings only show the components related to this disclosure and are not drawn according to the number, shape and size of the components in actual implementation. In actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0024] Furthermore, specific details are provided in the following description to facilitate a thorough understanding of the examples. However, those skilled in the art will understand that the described aspects can be practiced without these specific details.

[0025] Reference Figure 1 This disclosure provides a document interpretation method based on multi-agent collaboration, including the following steps: S1: Obtain the target document, extract its features, and generate a document feature vector; S2: Select target processing templates that match the document feature vectors from the preset processing template set; S3: Based on the target processing template, construct a directed acyclic graph. Nodes in the directed acyclic graph represent functional agents, and edges between nodes represent dependencies between functional agents. S4: Extract key information from the target document and save it to the shared workspace; S5: Based on the execution order of the directed acyclic graph, the corresponding functional agent is called to read input data from the shared work area, execute processing tasks based on the input data, and write the execution results back to the shared work area; S6: After each functional agent in the directed acyclic graph has completed its execution, output a literature interpretation report that integrates the execution results of each functional agent stored in the shared workspace.

[0026] The multi-agent collaborative document interpretation method disclosed herein, through a pre-screening mechanism of feature extraction and template matching, can accurately anchor the processing logic adapted to the target document, avoiding ineffective task allocation and resource waste in the multi-agent collaborative process. Specifically, the document feature vector can quantitatively represent the core attributes of the document, such as its topic, domain, and complexity. Screening the target processing template based on this vector ensures that the subsequently constructed directed acyclic graph (DAG) is highly matched with the document interpretation requirements. Simultaneously, the DAG, by defining functional agents through nodes and dependencies through edges, enables orderly collaboration among multiple agents, avoiding the low interpretation efficiency caused by task conflicts and dependency chaos in traditional parallel processing, and significantly improving the standardization and execution efficiency of the document interpretation process.

[0027] This method, relying on a shared workspace and targeted execution architecture, ensures the consistency and integrity of information flow during multi-agent collaboration, thereby improving the accuracy and comprehensiveness of the document interpretation report. On one hand, key information is centrally stored in the shared workspace, allowing each functional agent to read input data as needed, avoiding redundant operations of repeatedly extracting document information between agents, and preventing data deviations caused by scattered information storage. On the other hand, calling agents and writing back results according to the execution order of a directed acyclic graph ensures that each agent's processing tasks closely align with the core requirements of the document. The resulting integrated interpretation report covers multi-level information of the document, preserving the accuracy of key content while possessing comprehensive multi-dimensional analysis, meeting the practical needs of interpreting complex documents.

[0028] Furthermore, because this solution is based on a feature-matching-based approach to select target processing templates and uses a directed acyclic graph to define agent collaboration logic, it can quickly adapt to new document types and interpretation needs by adding processing templates or agent nodes without refactoring the entire system, thus supporting flexible functional expansion. Multiple agents collaboratively execute tasks in an orderly manner according to dependencies, and combined with a shared workspace, achieve unified information flow and reuse, replacing the traditional fixed-process model. This not only enhances the intelligent collaborative decision-making capabilities of the interpretation process but also reduces redundant operations, thereby significantly improving the overall efficiency of interpreting complex documents.

[0029] In S1, the dynamic task planning agent (Orchestrator Agent) serves as the system scheduling hub. It includes a document feature parsing module, which receives the input target document and preprocesses it, extracting metadata (such as title, abstract, chapter, number of figures and tables, references, etc.) from the target document, and constructing the extracted features into a document feature vector.

[0030] In S2, refer to Figure 2The flowchart illustrating the target processing template selection method shows that "selecting target processing templates that match the document feature vectors from the preset processing template set" includes the following steps: S21: For each processing template in the preset processing template set, a corresponding exclusive feature vector is preset; S22: Obtain the similarity between the unique feature vector of each processing template and the feature vector of the document; S23: Select the processing template with the highest similarity to the feature vector of the document as the target processing template.

[0031] In sections S21-S23, by introducing domain experts to deeply analyze existing literature interpretation scenarios and combining this with a survey of public user interpretation needs, a core interpretation requirement system is systematically organized and refined. Since the core objectives and information extraction focuses of different interpretation requirements vary, differentiated interpretation implementation methods are needed. Therefore, a dedicated processing template is constructed for each specific interpretation requirement. Simultaneously, based on the common characteristics of document types suitable for this interpretation requirement, representative dedicated feature vectors are preset and embedded into the corresponding processing template, achieving precise binding between "interpretation requirement - processing template - document features." Therefore, given that dedicated feature vectors possess common characteristics of a single type of document, a similarity-based filtering method can accurately obtain the most suitable processing template for the target document being interpreted.

[0032] In S21, refer to Figure 3 The flowchart illustrating the method for generating custom feature vectors shows that "for each processing template in the preset processing template set, a corresponding custom feature vector is pre-set" includes the following steps: S211: Classify document types according to interpretation needs and define a unique feature dimension system for each type of document. Each document type has its own unique processing template. S212: Collect representative samples for each document type and extract feature data of the representative samples according to the exclusive feature dimension system; S213: Perform statistical analysis on the characteristic data of representative samples of each document type according to their respective dimensions to obtain the standard characteristics of each type of document; S214: Combine the standard features of each document type in a uniform dimensional order to generate a unique feature vector for the corresponding processing template.

[0033] In S211, a hierarchical classification method is used to categorize literature based on its application scenario, interpretation goals, and domain characteristics. For example, literature is categorized by research paradigm into review articles, experimental research, theoretical derivation, and engineering application literature; and by subject area into medical, materials science, and computer science literature, ensuring clear category boundaries and coverage of mainstream interpretation needs. For each literature type, an initial feature dimension system is constructed from basic metadata, content features, and structural features. Furthermore, each dimension is further subdivided into numerical, Boolean, and categorical types, outputting a more detailed, proprietary feature dimension system. During the segmentation process, consistency between feature extraction and subsequent vector matching is ensured. The basic metadata dimension primarily uses numerical features, including directly quantifiable indicators such as title length, abstract character count, number of chapter levels, number of figures and tables, and number of references. The content feature dimension focuses on numerical features, supplemented by categorical features. These include numerical indicators extracted using natural language processing techniques, such as the density of core domain terms, the proportion of keywords in research methods, and the proportion of features in conclusion statements; as well as categorical indicators based on text content, such as the core semantic type of the title, the thematic tendency type of the abstract, and the type of core arguments in the chapters. The structural feature dimension encompasses both Boolean and categorical features. Boolean features include binary indicators reflecting the document structure, such as whether there is an experimental design chapter, whether there is a formula derivation module, and whether there is an independent discussion chapter. Categorical features include indicators with clear classification attributes, such as the type of reference citation format and the type of chapter hierarchy organization. Furthermore, the feature dimension system of each type of document must be completely aligned with the feature vector dimensions of the documents to be matched, ensuring a one-to-one correspondence between feature types and value methods.

[0034] In S212, for each category of literature, a sufficient number of representative sample documents are collected from different journals, publication times, and research teams. The sample size must meet the confidence requirements of the statistical analysis to avoid feature bias caused by sample uniformity. A unified preprocessing workflow is performed on the collected sample documents, including format standardization, noise removal, and text structure parsing. Then, based on the defined exclusive feature dimension system, feature data of the corresponding dimensions are extracted for each sample document.

[0035] In S213, the characteristic data of representative samples of each document type are organized by dimension to form a sample characteristic dataset of each document type in each dimension. Then, the type of the characteristics in each dataset (numerical, categorical, Boolean) is determined, and differential statistical analysis is performed accordingly.

[0036] The standard feature acquisition method for numerical features includes: for directly quantifiable numerical features such as title length, abstract character count, number of chapters, number of figures and tables, and number of references, the mean, median, variance, and reasonable value range of the sample feature dataset are calculated through statistical analysis. Then, the data is calibrated in combination with the experience of domain experts, such as removing the interference of abnormally short titles (such as subheadings of conference proceedings) and abnormally many references (such as extreme samples of review literature) in the sample. Finally, the standard mean / median of this dimension (as the dimension value of the exclusive feature vector) and a reasonable threshold range (as an auxiliary constraint condition when similarity matching) are determined.

[0037] The standard feature acquisition method for categorical features includes: for categorical features such as chapter structure type (e.g., "Abstract-Introduction-Experimental Methods-Results-Discussion-Conclusion" or "Abstract-Theoretical Derivation-Simulation Verification-Conclusion"), reference citation format type (e.g., APA format, GB / T7714 format), and "core semantic / topic type" of title and abstract (e.g., research object type in title, research conclusion type in abstract), the frequency and proportion of each category in the sample are statistically analyzed, and the core category with the highest proportion is determined as the standard feature value; if there are multiple high-frequency categories, weights are assigned according to their proportions to construct a composite standard feature of "category + weight", ensuring that it is completely consistent with the category / weight expression of this dimension in the input document feature vector.

[0038] The standard feature acquisition method for Boolean features includes: for Boolean features such as whether an experimental data section, whether a formula derivation module, or whether a figure or chart annotation is included, the percentage of "yes" or "no" in the sample is counted. When the percentage of a certain state exceeds a preset threshold (such as 70%), that state is determined as the standard feature value. If the percentages of two states are close, the state with the higher percentage is taken as the core value, and the compatibility weight of the other state is marked to ensure a one-to-one correspondence with the Boolean values ​​of the input document feature vector.

[0039] In S214, the standard features of various types of documents are arranged in an orderly manner according to the preset dimensional order, generating a standardized exclusive feature vector for that type of document. The dimensional order and dimensional type of the exclusive feature vector are exactly the same as the feature vector of the input document. For example, the exclusive feature vector of experimental research documents can be defined as [mean title length, mean number of characters in the abstract, median number of chapters, whether it contains experimental chapters (Boolean value), lower limit of the range of number of figures and tables, mean number of references, core category of citation format], which corresponds one-to-one with the dimensions of the feature vector of the input document, ensuring the accuracy of subsequent similarity calculations.

[0040] In S22, refer to Figure 4The flowchart illustrating the method for obtaining the similarity between the proprietary feature vector and the document feature vector shows that "obtaining the similarity between the proprietary feature vector of each processing template and the document feature vector" includes the following steps: S221: Based on the unique feature dimension system of each processing template, obtain the local similarity between the unique feature vector and the document feature vector in each dimension; S222: Based on the preset dimension weight allocation rules, the local similarities of each dimension are weighted and summed to obtain the similarity between the specific feature vector and the document feature vector.

[0041] In S221 and S222, the weights are initially configured based on the importance of different dimensions of various types of documents, and can be dynamically adjusted according to changes in user needs and technological development trends. For numerical features, if the numerical difference between the corresponding dimension of the specific feature vector and the document feature vector exceeds a preset reasonable threshold range, the weight of that dimension is significantly reduced, or even reset to zero. This method can accurately capture the degree of matching between the target document and the processing template in different dimensions such as basic metadata, content features, and structural features, avoiding the neglect of local key features in traditional single-vector matching. At the same time, it can assign higher weights to key feature dimensions based on different document interpretation needs, making the matching results more in line with actual application scenarios. This effectively solves the shortcomings of existing technologies, such as fixed processing procedures and inability to dynamically adapt to document types and interpretation needs, ultimately improving the accuracy and flexibility of processing template matching.

[0042] In S3, refer to Figure 5 The flowchart illustrating the method for constructing a directed acyclic graph (DAG) shows the steps involved in "Constructing a DAG based on a target processing template": S31: Read the pre-configured sequence of processing steps in the target processing template; S32: Analyze the functional agents corresponding to each step in the processing step sequence, and extract the input data constraints and output data format information of each functional agent; S33: Through data flow relationship analysis, determine the dependencies between each step in the processing step sequence; S34: Map the functional agents to nodes in a directed acyclic graph, and construct directed edges between nodes with data input-output associations based on dependencies; S35: Write the input data constraints and output data format information as node attributes into the attribute fields of the corresponding nodes to generate a directed acyclic graph.

[0043] In the above steps, each processing template is pre-configured with an ordered sequence of processing steps. After selecting the target processing template, the dynamic task planning agent reads the sequence of processing steps, performs structured parsing, identifies the functional agent identifier corresponding to each processing step, and extracts the core attribute information of each functional agent, including input data constraints (such as data format, data precision, and data source range) and output data format information (such as data structure type, field definition, and encoding rules). Based on this, by comparing the output data format of the functional agents in the preceding steps with the input data constraints of the functional agents in the subsequent steps, a unidirectional data flow link between steps is established, thereby determining the upstream and downstream dependencies of each processing step.

[0044] The dynamic task planning agent maps each functional agent to an independent node in a directed acyclic graph (DAG). Based on the established step dependencies, directed edges are constructed between nodes with data input-output relationships, with the edge direction consistent with the data flow direction to ensure the graph structure is free of closed loops. Subsequently, the dynamic task planning agent writes the input data constraints and output data format information corresponding to each functional agent into the attribute storage area of ​​the corresponding node, completing the encapsulation and configuration of node attributes, and finally generating a DAG that can be used to guide the execution of the literature interpretation process.

[0045] The dynamic task planning agent performs topology parsing on the constructed directed acyclic graph (DAG), extracting the topological structure information of the DAG, as well as the task description, inter-node dependencies, and task status information (including pending, processing, and completed) for each node. This information is stored in a dedicated storage area within the shared workspace, employing a key-value pair data structure with the task session ID as the primary key, enabling precise association and rapid retrieval of various information types.

[0046] In S4, refer to Figure 6 The flowchart illustrating the key information storage method shows that "extracting key information from the target document and storing it in the shared workspace" includes the following steps: S41: Extract key information from the target document based on the input data constraints and output data format information in the node attributes of the directed acyclic graph; S42: Identify the type of the extracted key information, and classify and store the key information into the corresponding storage partition of the shared work area according to the storage characteristics of different data types. S43: Synchronously write the storage path and index identifier of various key information in the shared work area into the key-value pair storage area with the task session ID as the primary key, and associate and bind it with the task information of the corresponding functional intelligent agent node.

[0047] In S41, the dynamic task planning agent retrieves the full set of node attribute information of the directed acyclic graph (DAG) from the key-value pair area with the task session ID as the primary key in the shared workspace. It then filters out the input data constraints (such as data format, field requirements, precision threshold, source range, etc.) and output data format information (such as data structure, encoding rules, field definitions, format specifications, etc.) corresponding to each functional agent node, and organizes them into a mapping table of "node identifier - extraction rule" according to the node association relationship.

[0048] The extraction rules in the mapping table are structured and parsed into executable extraction instructions. For example, if the input data constraint of a node is "experimental data table (including independent variable, dependent variable, and experiment number fields)," it is parsed as "identify table-type data from the target literature and filter tables containing the fields 'independent variable', 'dependent variable', and 'experiment number'." If the constraint is "abstract text with a density of core domain terms ≥ 3%", it is parsed as "extract the literature abstract text, count the frequency of core domain terms and calculate the density, and filter abstract fragments that meet the density standard and are semantically complete."

[0049] The document parsing engine is invoked to perform the extraction operation, employing appropriate extraction methods for different types of key information: For structured data (such as experimental parameters and knowledge triples), rule-based field matching and entity recognition algorithms are used, combined with a domain dictionary to verify the extraction results; for unstructured text (such as body paragraphs and abstracts), topic modeling and keyword extraction algorithms from natural language processing are used, combined with semantic constraints in the extraction rules to filter effective information; for image data (such as experimental setup diagrams and data curves), image type recognition algorithms from computer vision are used, matching image features specified in the rules (such as whether it contains the outline of experimental equipment or whether it is a coordinate curve) to complete the extraction. During the extraction process, the DAG node identifier corresponding to each piece of key information is recorded synchronously to ensure the traceability of the correspondence between information and nodes.

[0050] In S42, a data type identification module is constructed to determine the type of each extracted key piece of information and clarify its category. The determination rules are as follows: image data (such as experimental diagrams, schematic diagrams, and other visualization files in .png and .jpg formats); unstructured intermediate text data (such as unformatted text fragments, abstract texts, discussion texts, and other text information without fixed field structures); and structured data (such as experimental data tables with fixed fields, knowledge triples (entity-relation-entity), entity lists, quantification parameters, and other information with standardized structures). During the determination process, a type label is added to each piece of information (such as "image - experimental setup diagram", "unstructured text - abstract", "structured data - experimental parameter table").

[0051] The shared workspace is pre-configured with three main storage partitions, each adapted to the storage characteristics of different data types. In the unstructured data partition, an object storage service or document database is deployed to store unstructured text and image data, supporting large file storage and fuzzy retrieval. In the structured data partition, a relational database (such as MySQL) or a graph database (such as Neo4j) is deployed. The relational database is used to store tabular structured data, while the graph database is used to store structured data with relationships, such as knowledge triples. The index management partition stores index information and storage path mapping tables for various types of data.

[0052] Based on data type tags, key information is written to the corresponding storage partition. For unstructured text and image data, it is uploaded to the unstructured data partition, where a unique storage address is assigned by the object storage service or document database. Metadata tags containing task session ID, node identifier, data type, extraction time, and data source (such as document page number and chapter) are generated. An inverted index is built based on the metadata tags to ensure that subsequent retrieval is possible through multi-dimensional conditions. For structured data, it is imported into the corresponding database according to its data structure. For example, tabular data is written to the corresponding data table in a relational database, and knowledge triples are written to the node and relation tables in a graph database. Field indexes are built for the core fields of structured data (such as the "parameter name" field in the experimental parameter table and the "entity" field in the triples), and relation indexes are built for related fields to improve query efficiency.

[0053] In S43, the core storage information of each key piece of information is aggregated from the storage partitions of the shared workspace. This includes the storage path (such as the unique access address of unstructured data, the table name and row number of structured data in the database), index identifiers (such as metadata index ID, field index ID, relation index ID), and the corresponding task session ID, DAG node identifier, and data type label, forming a detailed list of "storage information - task / node association".

[0054] The dynamic task planning agent invokes the shared workspace key-value pair storage service. Using the current task's session ID as the primary key, it encapsulates the information in the detailed list into a value corresponding to this primary key, following the structure of "node identifier - storage path - index identifier - data type," and writes it to the dedicated key-value pair area. During the writing process, it ensures that each DAG node identifier is precisely associated with the storage path and index identifier of its corresponding key information. That is, by using the combination of "task session ID + node identifier," the storage location and retrieval index of the input data required by the agent for that node's function can be directly queried.

[0055] After the write operation is complete, a synchronization verification mechanism is triggered. The information just written is read from the dedicated key-value pair area and compared with the actual storage path and index identifier in each storage partition to confirm that the information is consistent and complete. If any inconsistencies or omissions are found, a correction process is automatically triggered to rewrite the correct information and verify it again until the verification passes. After successful verification, a "association binding complete" feedback signal is returned to the dynamic task planning agent, providing the prerequisite for the subsequent message queue triggering function agent to complete data preparation.

[0056] In S5, refer to Figure 7 The flowchart illustrating the functional agent invocation method demonstrates that "based on the execution order of the directed acyclic graph, the corresponding functional agent is invoked to read input data from the shared workspace, executes processing tasks based on the input data, and writes the execution results back to the shared workspace," which includes the following steps: S51: Based on the topological execution order of the directed acyclic graph, generate message instructions containing data access information and send them to the corresponding functional agents; S52: After receiving a message instruction, the functional agent retrieves matching input data from the shared workspace to execute a preset processing task and writes the execution result back to the specified storage location in the shared workspace. S53: After the execution result is written back, update the state of the corresponding directed acyclic graph node to "completed"; S54: When the current node state is detected as completed, determine the next functional agent based on the directed acyclic graph and send a message instruction to the next functional agent.

[0057] In S51-S54, the dynamic task planning agent generates message instructions based on the validated DAG topology, following the node dependency order. Each message instruction encapsulates the task session ID, target node identifier, and the storage path and index of the corresponding input data in the shared workspace. The data access information, such as the storage path and index of the corresponding input data in the shared workspace, is used by the agent to retrieve the data. Instructions are precisely pushed to the corresponding functional agents via a message queue, ensuring the orderly and reliable transmission of instructions.

[0058] After listening to the message queue and obtaining a matching instruction, the functional agent uses the task session ID and node identifier in the instruction to query the storage location of the input data in the key-value pair area of ​​the shared workspace. For unstructured data, it retrieves data files from object storage or document databases; for structured data, it directly retrieves the target data from relational or graph databases using indexes. After obtaining the data, the functional agent executes preset processing tasks, such as text semantic parsing, image feature extraction, and structured data association analysis. Upon completion of the task, the agent writes the results to the designated storage partition of the shared workspace according to data type and generates a new storage path and index identifier.

[0059] After the execution result is written back, the functional agent sends a status update notification to the dynamic task planning agent. Upon receiving the notification, the dynamic task planning agent immediately updates the status field of the corresponding DAG node in the dedicated area of ​​the shared workspace key-value pair from "pending" or "processing" to "completed". At the same time, it records the result storage information and the task completion timestamp to form a complete task execution log.

[0060] The dynamic task planning agent polls the status information of DAG nodes in the shared workspace in real time. When it detects that the current node's status is "completed," it resolves the node's direct successor node based on the DAG topology. According to the aforementioned instruction generation rules, it encapsulates the task session ID, node identifier, and new input data (i.e., the execution result of the current node) access information corresponding to the successor node, generates a new message instruction, and sends it to the message queue to trigger the next functional agent to execute the task. This process is repeated until the task execution of all nodes in the DAG is completed.

[0061] Furthermore, referring to Figure 8 The flowchart illustrating the message command sending and receiving method shows the following steps: S71: All functional agents are hierarchically divided according to processing stages and data flow levels; S72: Within each layer, groups are formed based on functional similarity and data dependencies, and functional agents in the same group share a dedicated message queue; S73: Construct a message routing table, which stores the mapping relationship between the node identifiers of the directed acyclic graph and the hierarchy, group, and corresponding message queue of the functional agents; S74: Based on the message routing table, the message instruction is published to the corresponding message queue by matching the target functional agent to the level and group by the target node identifier carried in the message instruction; S75: Functional agents in the same group listen to and consume message instructions in the corresponding message queue in parallel.

[0062] In S71-S75, combining the business logic and data flow sequence of the entire document interpretation process, all functional agents are divided into multiple core layers according to the "data processing progression relationship." For example, in complex scenarios, they are divided into 5 core layers. The responsibilities of each layer and the types of agents it contains are clearly defined as follows: ① Data preprocessing layer: responsible for basic preprocessing tasks such as document format conversion (e.g., PDF to text), noise data filtering (e.g., cleaning up redundant spaces / invalid characters), and data fragmentation (e.g., splitting documents by chapter), including format conversion agents, noise filtering agents, and data fragmentation agents. The layer consists of five layers: ① **Core Parsing Layer:** This layer is responsible for extracting core information, including text entity recognition agents, multimodal information extraction agents, and table / formula parsing agents. ② **Knowledge Enhancement Layer:** This layer enables external knowledge association and data completion, including knowledge base matching agents, knowledge completion agents, and cross-source data fusion agents. ③ **Inference and Analysis Layer:** This layer performs deep logical analysis, including association reasoning agents, conclusion derivation agents, and abnormal data verification agents. ④ **Report Generation Layer:** This layer integrates and outputs results, including result summarization agents, formatting agents, and report export agents. Each layer is allocated an independent message queue namespace (e.g., Preprocessing_NS, CoreParsing_NS) to ensure message transmission isolation between layers and avoid cross-layer interference.

[0063] For functional agents within each layer, a "functional clustering + dependency analysis" approach is used for grouping. For example: ① Functional similarity clustering: Agents performing similar subtasks are grouped together. In the core parsing layer, text entity recognition and multimodal information extraction agents are grouped into the "Information Extraction Group," while table parsing and formula recognition agents are grouped into the "Structured Data Parsing Group." ② Data dependency analysis: If multiple agents need to process the same batch of data or have sequential processing dependencies (e.g., completing text parsing before entity recognition), they are grouped together to ensure smooth data flow. A dedicated shared message queue is configured for each group (e.g., CoreParsing_InfoExtract_Queue for the Information Extraction Group, and Preprocessing_Queue or CoreParsing_Queue for other groups), and a message persistence mechanism is configured for the queue to prevent message loss. Simultaneously, message consumption priority rules are set within the group, marking priority for task messages with sequential dependencies to ensure ordered execution; task messages without dependencies are processed according to the default priority, supporting parallel consumption.

[0064] The message routing table is constructed primarily by a dynamic task planning agent, employing a hybrid structure of key-value pairs and relational tables to ensure accurate and traceable mapping relationships. The core mapping logic of the routing table is as follows: using the "DAG node identifier" as the primary key, the corresponding value is structured data containing "agent's hierarchy, agent's group, corresponding message queue address, and queue access permission token." During construction, all nodes in the DAG are traversed one by one, matching the corresponding functional agent through the node's task description field, and then associating the agent's hierarchy, group information, and corresponding message queue network address to complete the mapping relationship entry. After construction is complete, the message routing table is stored in the configuration partition of the shared workspace, and a real-time synchronization mechanism is enabled. When the functional agent's hierarchy / group is adjusted or the message queue address changes, the routing table information is automatically updated to ensure routing accuracy.

[0065] After receiving the "association and binding complete" feedback signal from the preceding process, the dynamic task planning agent generates message instructions based on the DAG topology execution order. The core information encapsulated in the message instructions includes the task session ID, the target node identifier (i.e., the DAG node identifier), and the input data storage path and index identifier (data access information) retrieved from the key-value pair area using the "task session ID + target node identifier". The task planning agent retrieves the message routing table, using the "target node identifier" in the instruction as the query key to match the corresponding target agent's level, group, and message queue address. After confirming a correct match, the agent pushes the message instructions to the corresponding message queue according to the message queue's communication protocol (such as AMQP or Kafka), simultaneously recording a message publication log (including the task session ID, target node identifier, target queue address, and publication timestamp), and writing the log to the shared workspace's log partition for subsequent task tracing and anomaly investigation.

[0066] After multiple functional agent instances within the same group start up, they monitor the message status of their corresponding shared message queues in real time. When a new message instruction enters the queue, the agent instance preempts message consumption rights based on a preset load balancing strategy (such as round-robin or weighted allocation) to prevent the same message from being consumed repeatedly. After receiving the message instruction, the agent parses the task session ID, target node identifier, and data access information. It then reads the input data from the corresponding storage partition in the shared workspace using the data access information and executes a preset processing task (such as text entity recognition or table parsing). After the task is completed, the agent writes the processing result back to the designated storage partition in the shared workspace according to the data type. Simultaneously, it sends a "task completed + node status updated" notification to the dynamic task planning agent, which updates the status of the corresponding DAG node to "completed." This triggers the next round of message instruction generation and publishing, until all DAG node tasks are completed.

[0067] In S6, the visualization report generation agent is typically the last functional agent to execute in the entire process. This agent retrieves the execution results of all other agents from the shared workspace, integrates the outputs of each stage to generate an interactive and interpretable report, and finally, the report exporting agent completes the report output. For example, refer to... Figure 9 The flowcharts showcasing typical functional agents and their outputs are as follows: The text parsing agent outputs structured document breakdown results, including independent modules such as title, author, main text paragraphs, and references; the multimodal information extraction agent outputs structured text description data for images and tables, generated using OCR and image recognition technologies; the domain knowledge enhancement agent outputs terminology-enhanced text data after entity linking and semantic disambiguation, along with related entries from external knowledge graphs or professional databases; the semantic relation extraction agent outputs triplet relationship data between concepts, and a local knowledge graph model built based on this data; the questioning and discovery agent outputs comparison results between document conclusions and existing research, including innovation point annotations, contradiction analysis, and a list of supporting evidence. The visualization report generation agent integrates and generates an interactive document interpretation report, which includes text summaries, knowledge graphs, and evidence analysis.

[0068] This solution achieves efficient collaboration and optimized resource allocation among intelligent agents through a hierarchical grouping design of functional agents, precise mapping of message routing tables, and hierarchical deployment of message queues. Independent namespaces between layers ensure message isolation at different processing stages, avoiding cross-layer interference. Intra-layer grouped shared queues, combined with priority rules, guarantee the orderly execution of dependent tasks while supporting the parallel advancement of independent tasks, significantly improving the overall execution efficiency of the document interpretation process. Simultaneously, the precise matching mechanism based on message routing tables enables targeted delivery of task instructions. Combined with a unified data management model in the shared workspace, this completely breaks down the coupling between modules, giving the system strong scalability; adding intelligent agents or adjusting processing flows requires no changes to the overall architecture.

[0069] Furthermore, the parallel listening and load-balanced consumption strategy of the same group of intelligent agent instances on the shared queue effectively achieves resource reuse and load balancing, avoiding the problems of single intelligent agent overload or resource idleness, and significantly improving the system's stability and resource utilization. Combined with the dynamically generated directed acyclic graph processing flow, the system can flexibly adapt the optimal execution path according to the document content. Compared with the fixed pipeline mode, it can better cope with the in-depth interpretation needs of multi-type and multi-structured documents, and further enhance the system's intelligent decision-making ability and scenario adaptability while ensuring processing accuracy.

[0070] Reference Figure 10 This disclosure provides a document interpretation system based on multi-agent collaboration, including: Feature extraction module 101 is used to acquire target documents, extract features from target documents, and generate document feature vectors; The template filtering module 102 is used to filter target processing templates that match the document feature vectors from a preset set of processing templates; Graph construction module 103 is used to construct a directed acyclic graph based on the target processing template. Nodes in the directed acyclic graph represent functional agents, and edges between nodes represent dependencies between functional agents. Information extraction module 104 is used to extract key information from the target document and store it in the shared workspace; The task execution module 105 is used to call the corresponding functional agent to read input data from the shared work area based on the execution order of the directed acyclic graph, execute processing tasks based on the input data, and write the execution results back to the shared work area. The report output module 106 is used to output a literature interpretation report that integrates the execution results of each functional agent stored in the shared workspace after each functional agent in the directed acyclic graph has completed its execution.

[0071] The various variations and specific examples of the document interpretation method based on multi-agent collaboration provided above are also applicable to the document interpretation system based on multi-agent collaboration provided in this disclosure. Through the foregoing detailed description of the document interpretation method based on multi-agent collaboration, those skilled in the art can clearly understand the implementation method of the document interpretation system based on multi-agent collaboration. For the sake of brevity, it will not be described in detail here.

[0072] A computer device according to embodiments of the present disclosure includes a memory and a processor. The memory is used to store non-transitory computer-readable instructions. Specifically, the memory may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may, for example, include random access memory (RAM) and / or cache memory. The non-volatile memory may, for example, include read-only memory (ROM), hard disk, flash memory, etc.

[0073] The processor may be a central processing unit (CPU) or other form of processing unit with data processing capabilities and / or instruction execution capabilities, and may control other components in the computer device to perform desired functions. In one embodiment of this disclosure, the processor is used to run computer-readable instructions stored in the memory, causing the computer device to perform all or part of the steps of the multi-agent collaborative document interpretation method of the foregoing embodiments of this disclosure.

[0074] Those skilled in the art will understand that, in order to solve the technical problem of how to achieve a good user experience, this embodiment may also include well-known structures such as communication buses and interfaces, and these well-known structures should also be included within the protection scope of this disclosure.

[0075] like Figure 11 This is a schematic diagram of a computer device provided for an embodiment of the present disclosure. It illustrates a structural schematic diagram suitable for implementing the computer device in the embodiments of the present disclosure. Figure 11 The computer device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments disclosed herein.

[0076] like Figure 11 As shown, a computer device may include a processor (such as a central processing unit, graphics processing unit, etc.), which can perform various appropriate actions and processes based on programs stored in read-only memory (ROM) or programs loaded from storage devices into random access memory (RAM). The RAM also stores various programs and data required for the operation of the computer device. The processor, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus.

[0077] Typically, the following devices can be connected to the I / O interface: input devices, such as sensors or visual information acquisition devices; output devices, such as displays; storage devices, such as magnetic tapes or hard drives; and communication devices. Communication devices allow the computer device to communicate wirelessly or wiredly with other devices (such as edge computing devices) to exchange data. Although Figure 11 A computer apparatus with various devices is shown, but it should be understood that it is not required to implement or have all of the devices shown. More or fewer devices may be implemented or included alternatively.

[0078] In particular, according to embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from a storage device, or installed from ROM. When the computer program is executed by a processor, all or part of the steps of the multi-agent collaborative document interpretation method of embodiments of this disclosure are performed.

[0079] For a detailed description of this embodiment, please refer to the corresponding descriptions in the foregoing embodiments, which will not be repeated here.

[0080] A computer-readable storage medium according to embodiments of the present disclosure stores non-transitory computer-readable instructions. When these non-transitory computer-readable instructions are executed by a processor, all or part of the steps of the multi-agent collaborative document interpretation methods described in the foregoing embodiments of the present disclosure are performed.

[0081] The aforementioned computer-readable storage media include, but are not limited to: optical storage media (e.g., CD-ROM and DVD), magneto-optical storage media (e.g., MO), magnetic storage media (e.g., magnetic tape or portable hard drive), media with built-in rewritable non-volatile memory (e.g., memory card), and media with built-in ROM (e.g., ROM cartridge).

[0082] For a detailed description of this embodiment, please refer to the corresponding descriptions in the foregoing embodiments, which will not be repeated here.

[0083] The basic principles of this disclosure have been described above with reference to specific embodiments. However, it should be noted that the advantages, benefits, and effects mentioned in this disclosure are merely examples and not limitations, and should not be considered as essential features of each embodiment of this disclosure. Furthermore, the specific details disclosed above are for illustrative and facilitative purposes only, and are not limitations. These details do not limit the scope of this disclosure to the necessity of employing the aforementioned specific details for implementation.

[0084] In this disclosure, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. The block diagrams of devices, apparatuses, devices, and systems involved in this disclosure are merely illustrative examples and are not intended to require or imply that they must be connected, arranged, or configured in the manner shown in the block diagrams. As those skilled in the art will recognize, these devices, apparatuses, devices, and systems can be connected, arranged, and configured in any manner. Words such as "comprising," "including," "having," etc., are open-ended terms meaning "including but not limited to," and are used interchangeably with them. The terms "or" and "and" as used herein refer to the terms "and / or," and are used interchangeably with them unless the context clearly indicates otherwise. The term "such as" as used herein refers to the phrase "such as but not limited to," and is used interchangeably with it.

[0085] Additionally, as used herein, the "or" used in a list of items beginning with "at least one" indicates a separate list, such that a list of, for example, "at least one of A, B, or C" means A or B or C, or AB or AC or BC, or ABC (i.e., A and B and C). Furthermore, the word "exemplary" does not imply that the described example is preferred or better than other examples.

[0086] It should also be noted that in the systems and methods of this disclosure, the components or steps can be decomposed and / or recombined. These decompositions and / or recombinations should be considered as equivalent solutions to this disclosure.

[0087] Various changes, substitutions, and modifications can be made to the technology described herein without departing from the teachings defined by the appended claims. Furthermore, the scope of the claims of this disclosure is not limited to the specific aspects of the processes, machines, manufactures, events, means, methods, and actions described above. Currently existing or later-developed processes, machines, manufactures, events, means, methods, or actions that perform substantially the same function or achieve substantially the same result as the corresponding aspects described herein can be utilized. Therefore, the appended claims include such processes, machines, manufactures, events, means, methods, or actions within their scope.

[0088] The above description of the disclosed aspects is provided to enable any person skilled in the art to make or use this disclosure. Various modifications to these aspects will be readily apparent to those skilled in the art, and the general principles defined herein may be applied to other aspects without departing from the scope of this disclosure. Therefore, this disclosure is not intended to be limited to the aspects shown herein, but rather to be carried out within the widest scope consistent with the principles and novel features disclosed herein.

[0089] The above description has been given for purposes of illustration and description. Furthermore, this description is not intended to limit the embodiments of this disclosure to the forms disclosed herein. Although numerous exemplary aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, alterations, additions, and sub-combinations therein.

Claims

1. A document interpretation method based on multi-agent collaboration, characterized in that, include: Obtain the target document, extract its features, and generate a document feature vector; From the preset set of processing templates, select target processing templates that match the document feature vector; Based on the target processing template, a directed acyclic graph is constructed, where nodes in the directed acyclic graph represent functional agents, and edges between nodes represent dependencies between functional agents. Extract the key information from the target document and store it in the shared workspace; Based on the execution order of the directed acyclic graph, the corresponding functional agent is invoked to read input data from the shared work area, executes processing tasks based on the input data, and writes the execution results back to the shared work area; After each functional agent in the directed acyclic graph has completed its execution, a literature interpretation report integrating the execution results of each functional agent stored in the shared workspace is output.

2. The document interpretation method based on multi-agent collaboration according to claim 1, characterized in that, The step of selecting target processing templates that match the document feature vector from a preset processing template set includes: For each processing template in the preset processing template set, a corresponding exclusive feature vector is preset; Obtain the similarity between the unique feature vector of each processing template and the feature vector of the document; The processing template with the highest similarity to the feature vector of the document is selected as the target processing template.

3. The document interpretation method based on multi-agent collaboration according to claim 2, characterized in that, The step of presetting a corresponding unique feature vector for each processing template in the preset processing template set includes: Documents are categorized according to interpretation needs, and a unique feature dimension system is defined for each type of document. Each document type has its own unique processing template. Collect representative samples for each document type, and extract feature data from the representative samples according to the exclusive feature dimension system; The characteristic data of representative samples of each document type are statistically analyzed according to their respective dimensions to obtain the standard characteristics of each type of document; The standard features of each document type are combined in a uniform dimensional order to generate a unique feature vector for the corresponding processing template.

4. The document interpretation method based on multi-agent collaboration according to claim 1, characterized in that, The construction of a directed acyclic graph based on the target processing template includes: Read the pre-configured sequence of processing steps from the target processing template; The functional agents corresponding to each step in the processing step sequence are analyzed, and the input data constraints and output data format information of each functional agent are extracted. By analyzing data flow relationships, the dependencies between each step in the processing step sequence are determined; The functional agent is mapped to nodes in a directed acyclic graph, and directed edges are constructed between nodes with data input-output associations based on the dependencies. The input data constraints and output data format information are written as node attributes into the attribute fields of the corresponding nodes to generate a directed acyclic graph.

5. The document interpretation method based on multi-agent collaboration according to claim 4, characterized in that, The process of extracting key information from the target document, segmenting it, and storing it in a shared workspace includes: Based on the input data constraints and output data format information in the node attributes of the directed acyclic graph, the key information of the target document is extracted; The extracted key information is identified by type, and based on the storage characteristics of different data types, the key information is classified and stored in the corresponding storage partition of the shared workspace. The storage paths and index identifiers of various key information in the shared work area are synchronously written into the key-value pair storage area with the task session ID as the primary key, and then associated and bound with the task information of the corresponding functional intelligent agent node.

6. The document interpretation method based on multi-agent collaboration according to claim 1, characterized in that, The process of invoking the corresponding functional agent to read input data from the shared workspace based on the execution order of the directed acyclic graph, executing processing tasks based on the input data, and writing the execution results back to the shared workspace includes: Based on the topological execution order of the directed acyclic graph, a message instruction containing data access information is generated and sent to the corresponding functional agent; After receiving a message instruction, the functional intelligent agent retrieves matching input data from the shared workspace to execute a preset processing task, and writes the execution result back to the specified storage location in the shared workspace. After the execution result is written back, the state of the corresponding directed acyclic graph node is updated to "completed". When the current node status is detected as completed, the next functional agent is determined based on the directed acyclic graph and a message instruction is sent to the next functional agent.

7. The document interpretation method based on multi-agent collaboration according to claim 6, characterized in that, Also includes: All functional agents are divided into layers according to processing stages and data flow levels; Within each layer, groups are formed based on functional similarity and data dependencies, and functional agents in the same group share a dedicated message queue. Construct a message routing table, which stores the mapping relationship between the node identifiers of the directed acyclic graph and the hierarchy, group, and corresponding message queue of the functional agents; Based on the message routing table, the target functional agent is matched with its level and group by the target node identifier carried in the message instruction, and the message instruction is published to the corresponding message queue. The same group of functional agents listen to and consume message instructions in the corresponding message queue in parallel.

8. A computer device, characterized in that, The computer device includes: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the document interpretation method based on multi-agent collaboration as described in any one of claims 1-7.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing a computer to perform the document interpretation method based on multi-agent collaboration as described in any one of claims 1-7.

10. A computer program product comprising computer instructions, characterized in that, When executed by a processor, the computer instructions implement the steps of the document interpretation method based on multi-agent collaboration as described in any one of claims 1-7.

Images