Digital human semantic understanding and task execution method and system based on multi-agent collaboration

Abstract

The invention discloses a digital human semantic comprehension and task execution method and system based on multi-agent collaboration, and relates to the technical field of artificial intelligence semantic comprehension and agent collaboration execution. Furthermore, semantic task blocks are aggregated on the basis of anaphora association and action association, so that fine-grained modeling of cross-statement and cross-round semantic relationships is realized, and the problems of semantic segmentation and ambiguity accumulation caused by the fact that a traditional method only depends on whole sentence intention recognition are effectively avoided; through a multi-agent overlapping section collaborative analysis and consistency arbitration mechanism, multi-view verification and alignment can be performed on the same task semantics, and the consistency and reliability of analysis results of entity parameters, action instruction parameters and expected state parameters are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence semantic understanding and intelligent agent collaborative execution technology, and in particular to a method and system for digital human semantic understanding and task execution based on multi-agent collaboration. Background Technology

[0002] With the continuous development of natural language processing, knowledge representation, and intelligent agent collaboration technologies, digital humans and digital employees are gradually evolving from simple question-and-answer interactions to collaborative execution of complex tasks. Existing research typically uses methods such as semantic parsing, intent recognition, and process scheduling to transform users' natural language input into executable business instructions, with a background scheduling module driving execution resources to complete the corresponding tasks. However, semantic input in real-world interaction scenarios often features multi-turn, cross-sentence, and strong contextual dependencies, with complex semantic referential relationships and a wealth of implicit action information. Relying solely on single-turn, sentence-level semantic recognition is insufficient to accurately depict the complete task intent and its execution boundaries, thus affecting the accuracy of task decomposition, the rationality of the execution order, and the consistency of execution results.

[0003] CN119443112A discloses a semantic recognition and task scheduling method and system for digital employees. It achieves intent recognition through a semantic analysis model and a deep learning model, and maps user requests to business processes based on a pre-defined business rule base, thereby enabling resource scheduling. While it can achieve intent understanding and task allocation at the single-request level, the semantic processing unit primarily uses "whole sentence text" as its basic parsing granularity. It fails to perform fine-grained decomposition and reorganization of referential and action-related relationships across sentences in multi-turn interactions, and it lacks a mechanism for multi-agent collaborative alignment and parsing of the same semantic task. Therefore, when user input presents multi-sentence decomposition, implicit referencing, and continuous action instructions, it is difficult to form stable and consistent semantic results. Furthermore, it lacks a mechanism for historical feedback modeling and weighted ranking of execution reliability, easily leading to task parsing ambiguity, unreasonable execution order, and insufficient result consistency. Summary of the Invention

[0004] The purpose of this section is to outline some aspects of the embodiments of the present invention and to briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section, as well as in the abstract and title of the present application, to avoid obscuring the purpose of this section, the abstract and title of the invention. Such simplifications or omissions shall not be used to limit the scope of the present invention.

[0005] In view of the aforementioned existing problems, the present invention is proposed.

[0006] Therefore, this invention addresses the common problems in existing digital employee semantic recognition and scheduling technologies, such as coarse-grained semantic segmentation of multi-round interactions, insufficient semantic association modeling, and lack of execution credibility feedback mechanisms.

[0007] To solve the above technical problems, the present invention provides the following technical solution: S1: The interactive semantic input received by the digital human is split into atomic semantic units according to the interaction rounds and sentence order; S2: Adjacent atomic semantic units with semantic relationships are aggregated into semantic task blocks, wherein the semantic relationships include referential relationships and action relationships; S3: For each semantic task block, multiple agents generate a unique aligned semantic result based on the overlapping segment parsing method; the aligned semantic result includes entity parameters, action instruction parameters, and expected state parameters; S4: Each semantic task block is assigned a credibility weight according to historical successful execution records, and the executing agents are triggered to execute the aligned semantic result in weight order, while the historical successful execution records are updated.

[0008] As a preferred embodiment of the present invention, wherein: S1 includes: obtaining the original statements in the interactive semantic input, labeling each original statement with the interaction round to which it belongs and the sequence number within the interaction round, forming an atomic semantic unit; the data structure of each atomic semantic unit includes the original statement text, the interaction round identifier and the statement sequence number.

[0009] As a preferred embodiment of the present invention, wherein: S2 includes: sequentially traversing atomic semantic units; if the subject or object in the subsequent atomic semantic unit refers to an entity that has appeared in the previous atomic semantic unit, then it is determined that there is a referential association; if the action description in the subsequent atomic semantic unit acts on an entity that has been mentioned in the previous atomic semantic unit, then it is determined that there is an action association; aggregating consecutive atomic semantic units with referential or action associations into a semantic task block, and recording the start and end numbers of the atomic semantic units contained in each semantic task block; the aggregation includes: creating a new semantic task block, and storing the interaction round identifier, start and end statement sequence number, and original statement text of the consecutive atomic semantic units in sequence.

[0010] In a preferred embodiment of the present invention, step S3 includes: for a semantic task block, the original statement texts are concatenated into task text according to the ascending order of the statement numbers of the contained atomic semantic units; the task text is divided into K text slices using a sliding window of length M, ensuring that there are M-1 overlapping statements between adjacent text slices; the K text slices are assigned to K different parsing agents, and each parsing agent extracts entities, action instructions, and state parameters from the text slices to form a slice parsing result; the slice parsing result includes all entity parameters, action instruction parameters, and state parameters identified from the statements within the text slices. The original statement number is used as the basis for determining the content of each overlapping statement. For each overlapping statement, the parsing results of all text segments containing the overlapping statement are compared. For the overlapping statements, if parameters targeting the same target are identified in different parsing results, a consistency check is performed: if all parsed parameter contents are consistent, the parameter is adopted; if they are inconsistent, arbitration is initiated: the different parameter contents and frequencies appearing in all parsing results are counted, and the parameter content with the highest frequency is adopted; if the frequencies are the same, the parameter content parsed by the parsing agent with the smallest number is adopted; based on the arbitration results of all statements, a unique aligned semantic result for the semantic task block is synthesized.

[0011] As a preferred embodiment of the present invention, the step of assigning credibility weights to each semantic task block based on historical successful execution records includes: querying historical successful execution records, which are stored in units of interaction rounds, and obtaining the total number of times each semantic task block has been successfully executed in past interaction rounds; calculating the weight of each semantic task block, where the weight is equal to the number of successful executions plus one; and sorting all semantic task blocks to be executed in descending order according to the weights to generate an execution queue.

[0012] In a preferred embodiment of the present invention, the triggering of the execution agent includes: retrieving the semantic task block with the highest weight from the execution queue, reading the entity parameters, action instruction parameters, and expected state parameters from its aligned semantic results; invoking the corresponding entity control agent according to the entity parameters, invoking the corresponding action execution agent according to the action instruction parameters, and invoking the corresponding state monitoring agent according to the expected state parameters, and sending execution and verification instructions respectively; the entity control agent, action execution agent, and state monitoring agent together form the execution agent group for this task, and respectively provide feedback on the instruction execution results; if all feedback results match the expected state parameters in the aligned semantic results, the execution is determined to be successful.

[0013] As a preferred embodiment of the present invention, updating the historical execution success record includes: if the execution is successful, recording the unique identifier of the semantic task block and the interaction round identifier to which the successful execution belongs to the historical execution success record, and incrementing the cumulative number of successful executions by one; regardless of success or failure, all feedback results of the execution are recorded.

[0014] On the other hand, the present invention also provides a digital human semantic understanding and task execution system based on multi-agent collaboration, comprising: The atomic semantic decomposition module is used to decompose the interactive semantic input received by the digital human into atomic semantic units according to the interaction rounds and sentence order; The semantic task aggregation module is used to aggregate adjacent atomic semantic units that have semantic relationships into semantic task blocks, wherein the semantic relationships include referential relationships and action relationships; The alignment semantic generation module is used to generate a unique alignment semantic result for each semantic task block by multiple agents based on the overlapping segment parsing method; the alignment semantic result includes entity parameters, action instruction parameters and expected state parameters. The trusted scheduling and execution module is used to assign a trust weight to each semantic task block based on the historical execution success record, and to trigger the execution agent to execute the aligned semantic result in order of weight, while updating the historical execution success record.

[0015] The beneficial effects of this invention are as follows: By decomposing the semantics of multi-turn interactions of digital humans into atomic semantic units and further aggregating them into semantic task blocks based on referential association and action association, this invention achieves fine-grained modeling of semantic relationships across sentences and turns, effectively avoiding the semantic fragmentation and ambiguity accumulation problems caused by traditional methods that rely solely on whole-sentence intent recognition; through multi-agent overlapping segment collaborative parsing and consistency arbitration mechanism, it is possible to perform multi-perspective verification and alignment of the semantics of the same task, significantly improving the consistency and reliability of the parsing results of entity parameters, action instruction parameters, and expected state parameters.

[0016] Meanwhile, this invention dynamically assigns task credibility weights based on historical successful execution records and optimizes the execution order accordingly, enabling the system to prioritize high-success-rate tasks, reduce execution conflicts and failure rates, and achieve adaptive optimization of semantic understanding and task execution processes by combining continuous feedback of execution results and a weight self-updating mechanism. This improves the accuracy, stability and intelligence of digital human's complex interactive task processing as a whole. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein: Figure 1 This is a flowchart illustrating a digital human semantic understanding and task execution method based on multi-agent collaboration, as shown in this invention. Detailed Implementation

[0018] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0019] Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without inventive effort should fall within the scope of protection of this invention.

[0020] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0021] According to an embodiment of the present invention, in combination Figure 1 The flowchart shown illustrates a method for semantic understanding and task execution of digital humans based on multi-agent collaboration, comprising: S1: The interactive semantic input received by the digital human is broken down into atomic semantic units according to the interaction rounds and sentence order.

[0022] It should be noted that in the process of semantic understanding and task execution of digital humans based on multi-agent collaboration, the interactive semantic input is usually presented as a continuous, multi-round, context-related natural language sequence. If the overall semantics are directly used as the processing object, it is easy to cause problems such as unclear semantic boundaries, loss of sequential relationships, and difficulty in backtracking subsequent execution feedback.

[0023] Therefore, in the initial stage of semantic processing, this invention first atomically decomposes the interactive semantic input received by the digital human to construct basic semantic units with clear temporal attributes and round affiliations. Specifically: First, the original statements in the interactive semantic input are obtained, and each original statement is labeled with its corresponding interaction round and its sequence number within the interaction round, forming an atomic semantic unit.

[0024] Secondly, the data structure of each atomic semantic unit includes the original statement text, the interaction round identifier, and the statement sequence number.

[0025] Specifically, the interactive semantic input received by the digital human is segmented according to the interaction rounds. Each interaction round consists of a user input and a corresponding digital human response. In scenarios involving only user input, the time interval between two adjacent inputs or the change of the session identifier is used as the basis for dividing the interaction rounds.

[0026] After determining the boundaries of the interaction rounds, the semantic input of the interaction in each interaction round is split into statements at the statement level. The statement splitting is based on a preset set of statement terminators, which includes at least one or more of the period, question mark, exclamation mark and newline character, thereby obtaining the original statement in the semantic input of the interaction.

[0027] Furthermore, within the same interaction round, each original statement is assigned a unique statement sequence number according to the order in which the original statements appear in the interactive semantic input. The statement sequence number is used to represent the relative position of the original statement in the corresponding interaction round.

[0028] Subsequently, each original statement is combined and encapsulated with its corresponding interaction round identifier and statement sequence number to generate an atomic semantic unit. Each atomic semantic unit corresponds to only one original statement and does not cross the interaction round boundary, thereby ensuring the minimization of semantic decomposition granularity and the traceability of sequence relationship.

[0029] Each atomic semantic unit is stored in a structured data format, and the data structure includes at least the following fields: original statement text field, used to record the complete text content of the corresponding statement; interaction round identifier field, used to indicate the interaction round to which the atomic semantic unit belongs; and statement sequence number field, used to indicate the sequential position of the atomic semantic unit within the interaction round.

[0030] By limiting the data structure as described above, any atomic semantic unit can be independently reconstructed in terms of temporal position and round affiliation without relying on context inference.

[0031] As can be seen, compared with the existing method of directly processing the entire interactive semantics, this invention introduces an atomic semantic unit decomposition mechanism based on the interaction rounds and statement order, which clearly defines the smallest operation unit for semantic processing and avoids parsing ambiguity caused by excessive semantic span; through interaction round identifiers and statement order numbers, complete time sequence information is preserved, which facilitates the subsequent construction of semantic closed loop and the backtracking of execution results.

[0032] S2: Aggregate adjacent atomic semantic units that have semantic relationships into semantic task blocks, wherein the semantic relationships include referential relationships and action relationships.

[0033] After the atomic semantic units are decomposed, each atomic semantic unit has a clear round and sequence attribute. However, in actual interaction, there are usually continuous semantic dependencies between multiple adjacent atomic semantic units, such as entity reference continuation or inheritance of the object of action.

[0034] If such semantic dependencies are not explicitly modeled, semantic fragmentation or redundant processing can easily occur during subsequent task execution. Therefore, this invention further performs ordered aggregation of semantically related atomic semantic units. The specific operation is as follows: S2.1: Iterate through the atomic semantic units sequentially. If the subject or object in the later atomic semantic unit refers to an entity that has appeared in the previous atomic semantic unit, then it is determined that there is a referential relationship.

[0035] Specifically, atomic semantic units are sequentially traversed according to their interaction round identifiers and statement sequence numbers. During the traversal, for each atomic semantic unit, subject and object entities are extracted from the original statement text (using a rule-based extraction method based on dependency parsing). The extracted entities are then recorded as the current entity set using text matching. Matching can be performed by extracting nominal subjects and objects as entity texts using dependency parsing, and storing the original word forms in a list. When performing referential judgment, pronouns are lemmatized and then precisely matched against the list.

[0036] When traversing to the next atomic semantic unit, the entity set corresponding to the previous atomic semantic unit is first retrieved for subsequent semantic association determination. The entity set only includes the entity text that appears explicitly in the previous atomic semantic unit and does not introduce external inference or semantic extension.

[0037] For the next atomic semantic unit, pronoun terms or elliptical subjects are extracted from the original sentence text and compared with the entity set of the previous atomic semantic unit item by item. When the pronoun term appears as a subject or object in the grammatical position and there is a unique matching entity in the entity set of the previous atomic semantic unit, it is determined that there is a referential relationship between the next atomic semantic unit and the previous atomic semantic unit.

[0038] The pronouns include at least personal pronouns, demonstrative pronouns, or zero-subject placeholders, and the determination of referential association is performed only between adjacent atomic semantic units and does not cross the sentence sequence number.

[0039] S2.2: If the action description in the subsequent atomic semantic unit applies to an entity that has been mentioned in the previous atomic semantic unit, then an action association is determined to exist.

[0040] Specifically, if the referential association determination condition is not met, the action description in the subsequent atomic semantic unit is further analyzed to extract the action name and the corresponding target text; when the target text is completely consistent with any entity text in the entity set of the previous atomic semantic unit, it is determined that there is an action association between the subsequent atomic semantic unit and the previous atomic semantic unit.

[0041] The determination of action association does not rely on action semantic reasoning, but is confirmed solely based on the consistency between the action's target and the entity text, in order to ensure the certainty and repeatability of the determination process.

[0042] S2.3: Aggregate consecutive atomic semantic units with referential or action associations into a semantic task block, and record the start and end numbers of the atomic semantic units contained in each semantic task block.

[0043] That is, when it is determined that there is a referential association or an action association between adjacent atomic semantic units, the consecutive atomic semantic units are grouped into the same semantic task block; if there is no such association, the current semantic task block is terminated and a new semantic task block is created.

[0044] The aggregation operation includes: creating a new semantic task block and storing the interaction round identifiers, start and end statement sequence numbers, and original statement text of the continuous atomic semantic units in sequence.

[0045] Each semantic task block includes at least the following information: semantic task block identifier, the starting statement sequence number and the ending statement sequence number of the contained atomic semantic units, the original statement text sequence arranged in statement order, and the corresponding interaction round identifier.

[0046] The above structure enables the semantic task blocks to logically form a continuous and independently processable semantic segment.

[0047] S3: For each semantic task block, multiple agents generate a unique aligned semantic result based on the overlapping segment parsing method; the aligned semantic result includes entity parameters, action instruction parameters, and expected state parameters.

[0048] S3.1: For a semantic task block, the original statement texts are concatenated into the task text according to the ascending order of the statement numbers of the contained atomic semantic units.

[0049] The task text serves as a unified input benchmark for subsequent parsing, and it still retains statement-level boundary markers to support subsequent parameter backtracking and consistency judgment.

[0050] S3.2: Divide the task text into K text segments using a sliding window of length M, ensuring that there are M-1 overlapping statements between adjacent text segments, so that each statement except for the first and last statements is contained in at least two different text segments. The window length M is empirically set to a fixed value between 3 and 5, or is set to no more than 1 / 3 of the total number of statements N in the semantic task block.

[0051] This overlapping fragmentation method allows the same statement to be parsed repeatedly in different contexts, providing a foundation for eliminating contextual dependency differences in the future.

[0052] S3.3: Assign K text segments to K different parsing agents. Each parsing agent extracts entities, action instructions, and state parameters from the text segments to form segment parsing results. The segment parsing results include all entity parameters, action instruction parameters, and state parameters identified from the statements in the text segments, as well as their corresponding original statement numbers.

[0053] Specifically, after text segmentation is completed, each text segment is assigned to a different parsing agent for independent parsing processing. These parsing agents are given a unified parsing output structure specification during system initialization, but they do not share intermediate states during the parsing process, ensuring the independence of the parsing results.

[0054] Each parsing agent performs entity recognition, action command recognition, and state change recognition operations sentence by sentence for the received text segment (entity recognition can be performed using a pre-trained BERT-CRF sequence labeling model, action command recognition can be performed using a BERT-based intent classification model, and state change recognition can be performed by combining dependency parsing and state keyword matching rules), and records the parsing results in a structured manner. Each parsed entity parameter, action command parameter, or state parameter is associated with at least the following information: the parameter content itself, the original sentence sequence number from which the parameter originates, and the contextual position identifier of the parameter within the current text segment.

[0055] The resulting fragmented parsing results not only contain parameter content but also clearly define the source at the statement level, providing a direct comparison basis for subsequent cross-fragment alignment.

[0056] S3.4: For each overlapping statement, compare the parsing results of all text segments containing the overlapping statement.

[0057] S3.5: For the overlapping statements, if parameters targeting the same target (same entity, same action, or same state) are identified in different fragmented parsing results, a consistency check is performed: If all the parsed parameters are consistent, then the result is accepted. If there is a disagreement, arbitration will be initiated: Analyze the different parameters and their frequencies in all fragmented parsing results, and adopt the parameter with the highest frequency as the final adopted result; If the frequency is the same, the parameter content parsed by the parsing agent with the smallest number is adopted according to the fixed numbering order of the parsing agents, thereby avoiding uncertainty.

[0058] Through the above-mentioned comparison and arbitration process, sentence by sentence and parameter by parameter, each sentence is compressed into a unique and definite parameter expression under multiple context parsing conditions.

[0059] S3.5: Based on the arbitration results of all statements, synthesize a unique aligned semantic result for the semantic task block.

[0060] Specifically, after arbitrating the parameters of all statement sequence numbers, the final adopted parameters corresponding to each statement are summarized in ascending order of statement sequence number to form a unified semantic expression covering the entire semantic task block.

[0061] The unified semantic expression is the aligned semantic result that constitutes the semantic task block. It clearly distinguishes the entity parameter set, the action instruction parameter set, and the expected state parameter set, and retains the mapping relationship between the parameters and the original statement sequence number.

[0062] The final generated alignment semantic result is the only parsed output of the semantic task block, written into the corresponding semantic task block record, and used as the direct input for subsequent semantic credibility gradient calculation and closed-loop execution sorting.

[0063] As can be seen, compared with existing single-model, single-context parsing methods, this invention effectively reduces semantic omissions and referential errors caused by single context truncation through overlapping fragmentation and independent parsing by multiple agents; through the statement number-based item-by-item consistency arbitration mechanism, the parsing results are deterministic and reproducible, avoiding the impact of model randomness on the execution results; and by preserving the mapping relationship between parameters and original statements, it provides a structured basis for subsequent historical consistency record updates and execution feedback backtracking.

[0064] S4: Assign credibility weights to each semantic task block based on historical successful execution records, and trigger the execution agent to execute the aligned semantic results in weight order, while updating the historical successful execution records.

[0065] S4.1: Assign credibility weights to each semantic task block based on historical successful execution records, including: Query historical execution success records, which are stored in units of interaction rounds, and obtain the total number of times each semantic task block has been successfully executed in past interaction rounds.

[0066] Calculate the weight of each semantic task block, where the weight equals the number of successful executions plus one, to ensure that new task blocks have a base weight. Sort all semantic task blocks to be executed in descending order according to their weights to generate an execution queue. This execution queue remains unchanged within the same interaction round to constrain the subsequent execution triggering order.

[0067] To prevent the first semantic task block from being permanently delayed due to a lack of historical data, a fixed base bias is introduced for each semantic task block during the calculation process, so that the weight is equal to the sum of the number of successful executions and the preset base value, thereby ensuring that the new task block has an initial weight that can participate in the sorting.

[0068] S4.2: Triggering the execution of the intelligent agent includes: The system retrieves the semantic task block with the highest weight from the execution queue and reads the entity parameters, action instruction parameters, and expected state parameters from the aligned semantic results. Based on this, the system generates instruction descriptions for different types of intelligent agents.

[0069] The corresponding entity control agent is invoked based on the entity parameters, the corresponding action execution agent is invoked based on the action instruction parameters, and the corresponding state monitoring agent is invoked based on the expected state parameters. Execution and verification instructions are then sent respectively.

[0070] The entity control agent, action execution agent, and state monitoring agent together form the execution agent group for this task, and each provides feedback on the instruction execution results.

[0071] After receiving all execution feedback results, the system compares each feedback result with the expected state parameters. If all feedback results match the expected state parameters in the aligned semantic results, the execution is considered successful.

[0072] S4.3: Update the historical execution success records, including: If the execution is successful, the unique identifier of the semantic task block and the interaction round identifier to which it belongs will be recorded in the historical execution success record, and the number of successful executions will be incremented by one. Specifically, when the execution is determined to be successful, the system writes the unique identifier of the semantic task block and the corresponding interaction round identifier into the historical execution success record, and updates the record based on the original number of successful executions, so that the semantic task block will obtain a higher credibility weight in subsequent interaction rounds.

[0073] Simultaneously, regardless of success or failure, all feedback results for this execution are recorded. The system packages the complete context of each execution into a case. A case includes: the text of the original semantic task block, the aligned semantic result generated by S3, the actual execution feedback result, and the final execution status (success / failure). This can be used to iteratively improve the accuracy of parameter extraction for the parsing agent in S3.

[0074] The present invention also provides a digital human semantic understanding and task execution system based on multi-agent collaboration, including: an atomic semantic decomposition module, used to decompose the interactive semantic input received by the digital human into atomic semantic units according to the interaction rounds and sentence order; The semantic task aggregation module is used to aggregate adjacent atomic semantic units that have semantic relationships into semantic task blocks, wherein the semantic relationships include referential relationships and action relationships; The alignment semantic generation module is used to generate a unique alignment semantic result for each semantic task block by multiple agents based on the overlapping segment parsing method; the alignment semantic result includes entity parameters, action instruction parameters and expected state parameters. The trusted scheduling and execution module is used to assign a trust weight to each semantic task block based on the historical execution success record, and to trigger the execution agent to execute the aligned semantic result in order of weight, while updating the historical execution success record.

[0075] The system also includes one or more processors and memory.

[0076] The memory is used to store operable instructions that, when executed by the one or more processors, cause the one or more processors to perform operations, including the flow of a multi-agent collaborative digital human semantic understanding and task execution method according to the foregoing embodiments, especially... Figure 1 The flowchart of the method is shown.

[0077] Other aspects disclosed in the embodiments of the present invention also propose a computer-readable medium for storing software including instructions executable by one or more computers, which, upon execution, cause the one or more computers to perform operations including the flow of a digital human semantic understanding and task execution method based on multi-agent collaboration as described in the foregoing embodiments, particularly... Figure 1 The flowchart of the method is shown.

[0078] It should be recognized that embodiments of the present invention may be implemented or carried out by computer hardware, a combination of hardware and software, or by computer instructions stored in a non-transitory computer-readable storage medium.

[0079] The method can be implemented using standard programming techniques, including a non-transitory computer-readable storage medium configured with a computer program in the computer program, wherein the storage medium is configured such that the computer operates in a specific and predefined manner.

[0080] Each program can be implemented in a high-level procedural or object-oriented programming language to communicate with the computer system; however, if required, the program can be implemented in assembly or machine language.

[0081] In any case, the language can be either compiled or interpreted.

[0082] Furthermore, for this purpose, the program can run on programmed application-specific integrated circuits.

[0083] The processes described herein (or variations and / or combinations thereof) can be executed under the control of one or more computer systems configured with executable instructions, and can be implemented by hardware or a combination thereof as code (e.g., executable instructions, one or more computer programs, or one or more applications) that commonly executes on one or more processors. The computer program includes a plurality of instructions executable by one or more processors.

[0084] Furthermore, the method can be implemented in any suitable computing platform, including but not limited to personal computers, minicomputers, mainframes, workstations, networked or distributed computing environments, standalone or integrated computer platforms, or in communication with charged particle tools or other imaging devices.

[0085] Various aspects of the present invention can be implemented in machine-readable code stored on a non-transitory storage medium or device, whether portable or integrated into a computing platform, such as a hard disk, optical read and / or write storage medium, RAM, ROM, etc., such that it can be read by a programmable computer, and when the storage medium or device is read by the computer, it can be used to configure and operate the computer to perform the processes described herein.

[0086] Furthermore, machine-readable code, or parts thereof, can be transmitted via wired or wireless networks.

[0087] When such media includes instructions or programs that combine with a microprocessor or other data processor to implement the steps described above, the invention described herein includes these and other different types of non-transitory computer-readable storage media.

[0088] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for semantic understanding and task execution of digital humans based on multi-agent collaboration, characterized in that: include: S1: The interactive semantic input received by the digital human is broken down into atomic semantic units according to the interaction rounds and sentence order; S2: Aggregate adjacent atomic semantic units that have semantic relationships into semantic task blocks, wherein the semantic relationships include referential relationships and action relationships; S3: For each semantic task block, multiple agents generate a unique aligned semantic result based on the overlapping segment parsing method; the aligned semantic result includes entity parameters, action instruction parameters, and expected state parameters. S4: Assign credibility weights to each semantic task block based on historical successful execution records, and trigger the execution agent to execute the aligned semantic results in weight order, while updating the historical successful execution records.

2. The digital human semantic understanding and task execution method based on multi-agent collaboration as described in claim 1, characterized in that: S1 includes: Obtain the original statements from the interactive semantic input, label each original statement with its corresponding interaction round and its sequence number within the interaction round, forming an atomic semantic unit; The data structure of each atomic semantic unit contains the original statement text, the interaction round identifier, and the statement sequence number.

3. The digital human semantic understanding and task execution method based on multi-agent collaboration as described in claim 2, characterized in that: S2 includes: The atomic semantic units are sequentially traversed. If the subject or object in the later atomic semantic unit refers to an entity that has appeared in the previous atomic semantic unit, then a referential relationship is determined to exist. If the action description in the subsequent atomic semantic unit applies to an entity that has been mentioned in the previous atomic semantic unit, then an action association is determined to exist. Aggregate consecutive atomic semantic units with referential or action associations into a semantic task block, and record the start and end numbers of the atomic semantic units contained in each semantic task block. The aggregation includes: creating a new semantic task block and storing the interaction round identifiers, start and end statement sequence numbers, and original statement text of the continuous atomic semantic units in sequence.

4. The digital human semantic understanding and task execution method based on multi-agent collaboration as described in claim 3, characterized in that: S3 includes: For a semantic task block, the original statement texts are concatenated into the task text according to the ascending order of the statement numbers of the atomic semantic units contained therein. The task text is divided into K text segments using a sliding window of length M, ensuring that there are M-1 overlapping sentences between adjacent text segments; K text segments are assigned to K different parsing agents. Each parsing agent extracts entities, action instructions, and state parameters from the text segments to form segment parsing results. The segment parsing results include all entity parameters, action instruction parameters, and state parameters identified from the statements in the text segments, as well as their corresponding original statement numbers. For each overlapping statement, compare the parsing results of all text segments containing the overlapping statement. For the overlapping statements, if parameters targeting the same target are identified in different fragment parsing results, a consistency check is performed: If all the parsed parameters are consistent, then the result is accepted. If there is a disagreement, arbitration will be initiated: Analyze the different parameter contents and frequencies that appear in all fragmented parsing results, and adopt the parameter contents that appear the most frequently; If the frequency is the same, the parameter content parsed by the parsing agent with the smallest number shall be adopted; Based on the arbitration results of all statements, a unique aligned semantic result is synthesized for the semantic task block.

5. The digital human semantic understanding and task execution method based on multi-agent collaboration as described in claim 4, characterized in that: The process of assigning credibility weights to each semantic task block based on historical successful execution records includes: Query historical execution success records, which are stored in units of interaction rounds, and obtain the total number of times each semantic task block has been successfully executed in past interaction rounds; Calculate the weight of each semantic task block, where the weight is equal to the number of successful executions plus one; All semantic task blocks to be executed are sorted in descending order according to their weights to generate an execution queue.

6. The digital human semantic understanding and task execution method based on multi-agent collaboration as described in claim 5, characterized in that: The triggering of the intelligent agent execution includes: Retrieve the semantic task block with the highest weight from the execution queue, and read the entity parameters, action instruction parameters, and expected state parameters from its aligned semantic results; The corresponding entity control agent is invoked based on the entity parameters, the corresponding action execution agent is invoked based on the action instruction parameters, and the corresponding state monitoring agent is invoked based on the expected state parameters, and execution and verification instructions are sent respectively. The entity control agent, action execution agent, and state monitoring agent together form the execution agent group for this task, and respectively report the instruction execution results; If all feedback results match the expected state parameters in the alignment semantic results, the execution is considered successful.

7. The digital human semantic understanding and task execution method based on multi-agent collaboration as described in claim 6, characterized in that: The updated historical execution success records include: If the execution is successful, the unique identifier of the semantic task block and the interaction round identifier to which the successful execution belongs will be recorded in the historical execution success record, and the number of successful executions will be incremented by one. Regardless of success or failure, all feedback results from this execution were recorded.

8. A digital human semantic understanding and task execution system based on multi-agent collaboration, based on the digital human semantic understanding and task execution method based on multi-agent collaboration as described in any one of claims 1 to 7, characterized in that: Also includes: The atomic semantic decomposition module is used to decompose the interactive semantic input received by the digital human into atomic semantic units according to the interaction rounds and sentence order; The semantic task aggregation module is used to aggregate adjacent atomic semantic units that have semantic relationships into semantic task blocks, wherein the semantic relationships include referential relationships and action relationships; The alignment semantic generation module is used to generate a unique alignment semantic result for each semantic task block by multiple agents based on the overlapping segment parsing method; the alignment semantic result includes entity parameters, action instruction parameters and expected state parameters. The trusted scheduling and execution module is used to assign a trust weight to each semantic task block based on the historical execution success record, and to trigger the execution agent to execute the aligned semantic result in order of weight, while updating the historical execution success record.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the digital human semantic understanding and task execution method based on multi-agent collaboration as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the digital human semantic understanding and task execution method based on multi-agent collaboration as described in any one of claims 1 to 7.

Images