Methods, devices, computer equipment, and storage media for user multi-intention recognition in a multi-agent framework

Abstract

This application discloses a method, apparatus, computer device, and storage medium for user multi-intent recognition in a multi-agent framework, relating to the field of Internet technology. It effectively avoids the limitations of traditional methods in understanding complex semantics and implicit logic, significantly improving the completeness and accuracy of intent transmission in multi-agent collaborative links, and reducing execution confusion and repeated interactions caused by semantic understanding biases. The method includes: receiving natural language instructions input by a user and performing semantic parsing to obtain multiple intent elements; combining the multiple intent elements and performing multiple intent decomposition steps in parallel on the natural language instructions to obtain multiple step-level structured intent units; based on the natural language instructions, identifying negative constraints and verifying the completeness of the multiple intent elements to generate additional structured intent fields; and integrating the multiple step-level structured intent units and the additional structured intent fields to obtain a multi-intent recognition result for the natural language instructions.

Description

Technical Field

[0001] This application relates to the field of Internet technology, and in particular to a method, apparatus, computer device and storage medium for recognizing multiple user intentions in a multi-agent framework. Background Technology

[0002] With the rapid development of artificial intelligence technology, multi-agent collaborative architectures are widely used in the processing of complex tasks to improve the intelligence level of systems. In practical applications, user input is usually presented in natural language, characterized by high degree of freedom, semantic ambiguity, and unclear structure. It often integrates multi-step operations, complex semantics, additional constraints, and mixed positive and negative requirements. This complexity makes it difficult for the system to accurately and completely understand the user's core needs and implicit logic, which can easily lead to deviations in subsequent task decomposition and collaborative execution, thereby affecting the overall stability of the system and the user experience.

[0003] In related technologies, when identifying complex user intents, a "slot-filling" structured extraction or classification-based intent recognition method is used. Through predefined templates or categories, key information in the user's natural language is extracted into structured data, such as identifying and filling specific slots, or classifying user intents into known intent types, thereby converting user input into basic instructions that the system can process.

[0004] However, the applicant recognizes that the relevant technology has at least the following technical problems in its implementation: It is unable to losslessly reproduce complex semantics of users and it is difficult to fully decompose the intent elements in users' natural language. Under the multi-agent collaborative architecture, it is prone to problems such as execution chaos, information loss and repeated interactions in the multi-agent link. This not only reduces the reliability and efficiency of the intelligent experience, but also affects the stability of the system and results in poor intelligence. Summary of the Invention

[0005] In view of this, this application provides a method, apparatus, computer device and storage medium for user multi-intention recognition in a multi-agent framework, the main purpose of which is to solve the problems that currently reduce the reliability and efficiency of intelligent experience, and affect the stability and intelligence of the system.

[0006] According to a first aspect of this application, a method for recognizing multiple user intents in a multi-agent framework is provided, the method comprising: It receives natural language commands from the user and performs semantic parsing to obtain multiple intent elements; Combining the multiple intent elements, multiple intent decomposition steps are executed in parallel on the natural language instruction to obtain multiple step-level structured intent units, wherein each intent decomposition step decomposes and parses an independent intent in the natural language instruction; Based on the natural language instructions, negative constraint conditions are identified and integrity checks are performed on the multiple intent elements to generate additional structured intent fields; The multiple step-level structured intent units and the additional structured intent fields are integrated to obtain a multi-intent recognition result for the natural language instruction. The multi-intent recognition result is output to a multi-agent architecture, so that each agent in the multi-agent architecture processes its assigned step-level structured intent unit in combination with the additional structured intent field, generates feedback content for the step-level structured intent unit, and the multi-agent architecture pushes the multiple feedback contents generated by the multiple agents for the multiple step-level structured intent units to the user for feedback.

[0007] According to a second aspect of this application, a user multi-intention recognition device for a multi-agent framework is provided, the device comprising: The semantic parsing module is used to receive natural language commands input by the user and perform semantic parsing to obtain multiple intent elements; The multi-step decomposition module is used to combine the multiple intent elements and execute multiple intent decomposition steps in parallel on the natural language instruction to obtain multiple step-level structured intent units, wherein each intent decomposition step decomposes and parses an independent intent in the natural language instruction. The additional field generation module is used to identify negative constraints and verify the integrity of the multiple intent elements based on the natural language instructions, so as to generate additional structured intent fields. The recognition result generation module is used to integrate the multiple step-level structured intent units and the additional structured intent fields to obtain a multi-intent recognition result for the natural language instruction. The multi-intent recognition result is then output to a multi-agent architecture, so that each agent in the multi-agent architecture processes its assigned step-level structured intent unit in conjunction with the additional structured intent fields, generates feedback content for the step-level structured intent unit, and pushes the multiple feedback contents generated by the multiple agents for the multiple step-level structured intent units to the user for feedback.

[0008] According to a third aspect of this application, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method described in any of the first aspects above.

[0009] According to a fourth aspect of this application, a readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described in any one of the first aspects above.

[0010] By employing the above technical solutions, this application provides a method, apparatus, computer device, and storage medium for user multi-intention recognition in a multi-agent framework. Through a collaborative processing mechanism of semantic parsing, parallel intention decomposition, negative constraint recognition, and integrity verification, this application can accurately extract and reconstruct multi-level intention elements of users from natural language instructions with high degrees of freedom and ambiguous structure. This effectively avoids the limitations of traditional intention recognition methods in understanding complex semantics and implicit logic, thereby significantly improving the integrity and accuracy of intention transmission in the multi-agent collaborative link, reducing execution chaos and repeated interactions caused by semantic understanding bias, and enhancing the overall stability, reliability, and intelligent response efficiency of the system.

[0011] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application 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 application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0012] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 This illustration shows a flowchart of a user multi-intention recognition method for a multi-agent framework provided in an embodiment of this application. Figure 2 This illustration shows a flowchart of another user multi-intention recognition method for a multi-agent framework provided in an embodiment of this application. Figure 3 This paper illustrates a schematic diagram of the system architecture of a human-computer interaction system provided in an embodiment of this application. Figure 4 This illustration shows a schematic diagram of a user multi-intention recognition device for a multi-agent framework provided in an embodiment of this application. Figure 5 A schematic diagram of the device structure of a computer device provided in an embodiment of this application is shown. Detailed Implementation

[0013] Exemplary embodiments of the present application will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present application are shown in the drawings, it should be understood that the present application may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this application will be thorough and complete, and will fully convey the scope of the present application to those skilled in the art.

[0014] This application provides a method for recognizing multiple user intents in a multi-agent framework, such as... Figure 1 As shown, the method includes: S10: Receives natural language instructions input by the user and performs semantic parsing to obtain multiple intent elements.

[0015] The technical solutions in this application embodiment can be applied to human-computer interaction systems in various scenarios. These systems can receive and understand user input in natural language, voice, or other free-form formats, thereby driving backend services or devices to perform corresponding tasks. Specifically, the human-computer interaction system uses multi-intent recognition and parsing technology to accurately convert unstructured user input, which may contain multiple requests or complex conditions, into a series of structured, executable operation instructions. Human-computer interaction systems can take many forms in different application scenarios. For example, in smart home scenarios, they can be integrated into the home automation control hub of a central control device or smart speaker, capable of understanding complex user commands such as "turn off the living room lights and turn on the air conditioner to 25 degrees Celsius." In in-vehicle scenarios, they can be in-vehicle infotainment and control systems, capable of handling multi-task requests such as "navigate to the airport and tell me the gas stations along the way." In enterprise office scenarios, they can be intelligent meeting scheduling or process assistants, capable of parsing complex commands such as "book a large conference room for 2 pm tomorrow and notify all members of the project team," which include actions, time, objects, and notification conditions. In customer service scenarios, they can be intelligent customer service or digital employee systems, capable of accurately responding to users' mixed requests such as "I want to return the phone I bought last week, but not the black one." This application does not limit the specific form of the human-computer interaction system.

[0016] In this embodiment, the human-computer interaction system receives natural language commands input by the user and performs semantic parsing to obtain multiple intent elements. Specifically, the system analyzes the syntactic structure, dependency relationships, and context of the commands to extract the basic semantic components constituting the user's intent as intent elements. This is achieved through a deep neural network to understand the overall semantics of the language and extract discrete intent elements such as actions (predicates), target objects (objects), attributes (modifiers), and logical connectives. Intent elements are atomized semantic units that carry the user's request. For example, assuming the natural language command is "Find me a nearby Chinese restaurant with high ratings, but not Sichuan cuisine," after parsing, this command yields multiple intent elements such as "action: find," "target: Chinese restaurant," "attributes: nearby, high ratings," and the logical connective "not," which includes "attribute: Sichuan cuisine." This achieves a preliminary structural deconstruction of the user's free expression, transforming unstructured natural language into a set of elements carrying clear semantics that can be processed in subsequent steps, laying the foundation for subsequent fine-grained decomposition.

[0017] In step S10, which involves receiving natural language input from the user and performing semantic parsing to obtain multiple intent elements, the steps include: S11: When a natural language instruction is received from the user, the natural language instruction is decomposed to obtain multiple discrete semantic units.

[0018] In this embodiment, the original natural language instruction is first subjected to preliminary linguistic segmentation and parsing. Specifically, the human-computer interaction system utilizes techniques such as word segmentation, part-of-speech tagging, named entity recognition, and dependency parsing to decompose the continuous natural language sentence into a set of multiple discrete semantic units carrying independent, basic semantics. These discrete semantic units can be words, phrases, or clauses with complete syntactic functions, forming the basis for understanding the natural language instruction. For example, assuming the natural language instruction is "Help me book a Sichuan restaurant suitable for a gathering tomorrow night, near a subway station and with private rooms," this instruction can be decomposed into multiple discrete semantic units such as "help me," "book," "a," "tomorrow night," "suitable for a gathering," "of," "Sichuan restaurant," "want," "near subway station," "and," "have," and "private rooms." This achieves the transformation from continuous text to computable semantic units, providing a clear and unambiguous input basis for subsequent structured mapping and avoiding overall comprehension bias caused by complex sentence structures.

[0019] S12: Perform context understanding and entity recognition on multiple discrete language units to determine multiple levels of intent units in the multiple discrete language units, and map each level of intent unit to the corresponding level of structured intent fields to obtain multiple levels of structured intent fields.

[0020] After obtaining discrete semantic units, the human-computer interaction system identifies and extracts intent units that represent the user's core objectives from multiple discrete language units through contextual understanding and entity recognition. In this embodiment, a hierarchical structure is set for intents, so that intent units are divided into different levels, such as first-level intents (main_intent) and second-level sub-type intents (sub_intent). The human-computer interaction system will analyze the semantic relationships and domain knowledge between discrete language units to determine which unit combinations jointly define the user's main purpose and its specific classification, and fill these identified intent units into the corresponding fields of the preset structured template to obtain structured intent fields of multiple levels.

[0021] Continuing with the previous example, the human-computer interaction system can identify the core action "order" and the target object "Sichuan restaurant," thus defining the primary intent as "restaurant reservation." Based on "suitable for group dining" and "Sichuan cuisine," it further refines the secondary subtype to "group dining - Sichuan cuisine." This achieves the transformation from fragmented semantics to a hierarchical and categorized structured intent, clearly labeling vague user needs with target tags. It provides a hierarchical and standardized instruction classification basis for subsequent multi-agent task routing and division of labor, ensuring that tasks are assigned to the correct processing modules.

[0022] S13: In multiple discrete language units, identify multiple other discrete language units besides multiple levels of intent units, and map the multiple other discrete language units to corresponding fields according to the content indicated by each other discrete language unit to obtain multiple other fields.

[0023] In this embodiment, the human-computer interaction system determines multiple other discrete language units besides the multiple levels of intent units among multiple discrete language units. Based on their semantic roles, the multiple other discrete language units are mapped to various detailed fields predefined based on the JSON structure to obtain multiple other fields. The multiple other fields together constitute the specific parameters and context required to execute the intent.

[0024] Among them, several other fields include at least the action field, subject field, object field, condition field, and environment field. Continuing with the previous example, "order" is mapped to the action field "operate"; "I" is mapped to the subject field "subject" and its user ID is determined through the dialogue history and mapped to the subject_identifier field; "tomorrow night" is mapped to the condition field "condition"; "Sichuan restaurant" itself is already partially reflected in the intent as the core object, but its specific limitations "near the subway station" and "have private rooms" can be mapped as key attributes to the environment field "environment" or as the object field "object".

[0025] In this way, through the above process, all the details of the original natural language instructions can be collected in detail, and the vague natural language description can be transformed into key-value pair parameters that the machine can operate precisely. This ensures the executability of subsequent intentions and provides all the specific information required for multiple downstream intelligent agents to perform actions, fundamentally avoiding execution failures or result deviations caused by missing or misunderstanding of key parameters.

[0026] S14: Treat each field obtained as an intent element to obtain multiple intent elements.

[0027] In this embodiment, the human-computer interaction system treats all generated structured intent fields as independent intent elements to obtain multiple intent elements. These multiple intent elements together constitute a comprehensive, structured, and lossless digital representation of the natural language instruction. It is a highly structured and information-complete set of intent elements, rather than a single label or a few slots. It preserves the hierarchical relationships and execution details in the user's original instruction, providing high-quality and standardized input for subsequent parallel intent decomposition, negative constraint verification, and other processing. This is a key prerequisite for achieving accurate multi-intent recognition and eliminating information loss and execution chaos.

[0028] S20: Combining multiple intent elements, multiple intent decomposition steps are executed in parallel on natural language instructions to obtain multiple step-level structured intent units.

[0029] In this embodiment, after acquiring multiple intent elements, the human-computer interaction system initiates multiple intent decomposition steps in parallel. Each intent decomposition step decomposes and parses an independent intent within the natural language instruction, responsible for identifying and merging elements matching the independent intent it is processing from multiple intent elements, thereby constructing a step-level structured intent unit. A step-level structured intent unit is an instruction submodule with a complete internal structure, capable of independent interpretation and execution, containing the core action, operation object, and key limiting conditions of the corresponding independent intent. It should be noted that multiple intent decomposition steps are executed in parallel, allowing the parsing of multiple sub-intents to proceed synchronously without blocking each other, significantly improving processing efficiency.

[0030] For example, assuming the natural language instruction is "navigate to the company and play news on the way," the human-computer interaction system will trigger two decomposition steps in parallel: navigation intent and media playback intent. The navigation intent decomposition step will capture elements such as "action: navigation" and "goal: company" to form a navigation instruction unit. The media playback intent decomposition step will capture "action: play," "goal: news," and the condition "on the way" to form a media control unit. This efficiently and clearly extracts and structures multiple independent requests from a user's complex utterance, solving the problem that traditional serial or single intent recognition methods are prone to omitting or confusing different steps when faced with complex instructions. This provides accurate input for the parallel task distribution in multi-agent systems.

[0031] In step S20, which involves combining multiple intent elements and performing multiple intent decomposition steps in parallel on the natural language instruction to obtain multiple step-level structured intent units, the steps include: S21: Perform semantic parsing on natural language instructions, determine multiple intent boundaries in the natural language instructions, and decompose the natural language instructions according to the multiple intent boundaries to obtain multiple independent intents.

[0032] In this embodiment, the human-computer interaction system identifies the dividing points between multiple sub-intents in a natural language instruction, that is, identifies the intent boundaries. Specifically, it can use a pre-trained sequence labeling model or analysis based on dependency syntax rules to scan the natural language instruction, focusing on identifying conjunctions that indicate parallel relationships such as "and" and "moreover", sequential relationships such as "then", and adversative relationships, as well as obvious switching of semantic subject-predicate structures. The identified conjunctions or semantic structure switching points are determined as intent boundaries, thereby logically dividing the original natural language instruction into multiple semantically relatively independent segments, each segment corresponding to an independent intent.

[0033] For example, suppose the natural language instruction is "Check the weather in Beijing tomorrow and then book the earliest high-speed rail ticket to Shanghai." The human-computer interaction system will recognize the conjunction "then" as a clear intent boundary, thereby breaking down the instruction into two independent intents: "Check the weather in Beijing tomorrow" and "Book the earliest high-speed rail ticket to Shanghai." This captures the implicit task sequence or parallel relationship in the user's speech, clearly isolating mixed instructions from the source. This provides a clear input division for subsequent parallel and specialized processing, effectively preventing the logical entanglement and information mismatch caused by traditional methods that mix multiple intents together for parsing.

[0034] S22: Set an intent decomposition step for each independent intent, and combine multiple intent elements to execute multiple intent decomposition steps corresponding to multiple independent intents in parallel, so as to generate step-level structured intent units of the corresponding independent intents by executing each intent decomposition step, and obtain multiple step-level structured intent units.

[0035] In this embodiment of the application, after obtaining multiple independent intentions, the human-computer interaction system dynamically assigns an intention decomposition step to each independent intention, so that multiple independent intentions correspond to multiple intention decomposition steps, and the multiple intention decomposition steps are executed in parallel and work independently.

[0036] Specifically, for each intent decomposition step, when executing the intent decomposition step, the human-computer interaction system will first determine the specified independent intent to be processed by the intent decomposition step, and extract multiple specified intent elements related to the specified independent intent from multiple intent elements.

[0037] Next, the multiple specified intent elements are normalized for actions and entities, resulting in processed specified intent elements. Specifically, during normalization, action intent elements related to actions and entity intent elements related to entities are extracted from the multiple specified intent elements. Then, a pre-built standardized description library is used to query the standardized action descriptions corresponding to the action intent elements and the standardized entity descriptions corresponding to the entity intent elements, and to identify whether the standardized action descriptions and standardized entity descriptions match. If a mismatch is found, disambiguation analysis is performed on the standardized action descriptions and standardized entity descriptions, combining contextual information from natural language instructions and historical interaction data, to determine the standard semantic form. Referring to the standard semantic form, the standardized action descriptions and standardized entity descriptions are normalized to ensure that the actions and entities in the multiple specified intent elements are consistent and unambiguous, thus obtaining the processed specified intent elements. In other words, normalization requires the human-computer interaction system to extract the parts of elements related to actions such as "search" and "book" and core entities such as "weather" and "high-speed rail tickets," and map them against a standardized description library. This transforms diverse user expressions into unified standardized expressions within the system; for example, "search" and "query" are both standardized as "query." Moreover, if a logical mismatch is found between the standardized action and the entity during this process—for example, if "book" might be incorrectly associated with "book a hotel" instead of "book a ticket" from some historical ambiguous data—the human-computer interaction system will initiate disambiguation analysis. Combining the context of the original natural language command and the user's historical interaction data, it will determine a more reasonable standard semantic form to ensure that the relationship between the action and the entity is unambiguous.

[0038] Finally, the human-computer interaction system obtains a preset structured template and integrates multiple specified intent elements after processing, referring to the structured template, to obtain a step-level structured intent unit with a specified independent intent. The structured template can be similar to JSON, containing slots for operations, objects, time, location, etc., thereby generating a well-formatted and complete step-level structured intent unit. For example, for the independent intent "book the earliest high-speed rail ticket to Shanghai", the corresponding decomposition steps would extract elements such as {operate: "book", object: "train ticket", object_identifier: {destination: "Shanghai", train number requirement: "earliest"}, condition: "tomorrow"}. After normalization and disambiguation, such as ensuring that "book" is mapped to "book" and the object is "train ticket" and not something else, it is integrated into a complete structured instruction unit. Thus, while improving the processing efficiency of complex instructions through parallelization, by focusing on the deep parsing of individual intents, strict normalization and context disambiguation, it ensures that each sub-intent is accurately and unambiguously understood and structured, generating high-quality instructions that can directly drive specific intelligent agents to execute. This fundamentally solves the problems of action object mismatch and semantic ambiguity that are easy to occur when multiple intents are mixed, providing accurate, reliable and non-interfering task input for downstream multi-agent collaboration, and ensuring the accuracy and stability of the entire chain execution.

[0039] S30: Based on natural language instructions, perform negative constraint identification and integrity verification on multiple intent elements to generate additional structured intent fields.

[0040] In this embodiment, the human-computer interaction system performs preliminary element extraction and decomposition supplementation and correction to obtain implicit and negative information in natural language instructions that are easily overlooked or insufficiently processed, thereby ensuring the complete reproduction of user intent. Specifically, the human-computer interaction system performs negative constraint condition identification and integrity verification on multiple intent elements based on natural language instructions to generate additional structured intent fields. Negative constraint condition identification refers to using specific semantic analysis, such as focusing on negative words and exclusionary phrases, to identify conditions that the user does not want to occur or needs to exclude from the original natural language instructions and parsed intent elements, such as content guided by "don't," "avoid," or "except for," and transforming them into explicit constraint rules. Integrity verification, based on domain logic, verifies whether the combination of all identified positive intent elements constitutes a complete and executable instruction, checking for logical contradictions or missing necessary information.

[0041] The human-computer interaction system encodes the identified negative constraints and supplementary information discovered during validation, such as default value filling and explicit implicit conditions, to form an additional structured intent field. This additional structured intent field is independent of specific step units and is used to globally modify or limit the entire task. For example, assuming the natural language instruction is "Book the earliest high-speed rail ticket for tomorrow, not a through train," in addition to breaking it down into the "book ticket" unit, the human-computer interaction system will also identify the negative constraint "not a through train" and validate it to ensure the integrity of key information such as time (i.e., tomorrow) and train frequency requirement (i.e., the earliest train). Finally, it generates an additional field containing {excluded train type: through train}, thereby enhancing the ability to capture complex and implicit user demands, accurately modeling negative and exclusionary semantics that are difficult to handle with traditional methods, ensuring the rigor and comprehensiveness of the final intent representation, and reducing execution errors caused by misunderstanding what the user does not want.

[0042] In step S30, which involves identifying negative constraints and verifying the integrity of multiple intent elements based on natural language instructions to generate additional structured intent fields, the following steps are included: S31: By identifying negative semantic elements included in multiple intent elements, determine negative constraints, and generate negative constraint fields by combining natural language instructions and negative constraints.

[0043] In this embodiment, the human-computer interaction system recognizes explicit negation or exclusion intents in natural language commands, thereby avoiding the omission of key information. The specific process is as follows: First, the human-computer interaction system will identify multiple intent elements and determine negative semantic elements. These negative semantic elements include negation words, negative expressions, or exclusionary semantic expressions, such as words or phrases like "don't," "don't," "avoid," and "except for."

[0044] Once a negative semantic element is identified, the human-computer interaction system will locate the semantic segment containing the negative semantic element in the natural language instruction as a negative constraint condition, and identify the constrained action and the constrained object of the constrained action in the negative constraint condition. For example, the negated behavior is "release", and the constrained object is "chili".

[0045] Then, to ensure the accuracy of the constraints, the human-computer interaction system will identify the contextual information of the constrained action and the constrained object in the natural language instructions, and determine the constrained scenario and the scope of the constraint. For example, if the constrained scenario is "ordering food", the scope of the constraint is global exclusion or local exception.

[0046] Finally, the human-computer interaction system performs structured processing on the constrained actions, constrained objects, constrained scenarios, and constraint ranges, converting these contents into standardized key-value pairs that can be processed internally, thus obtaining the negative constraint field.

[0047] For example, suppose a user's natural language instruction is "I want to order Kung Pao Chicken, but without chili peppers." The human-computer interaction system identifies the negative semantic element "no," locates the constraint "no chili peppers," analyzes the constrained action "add," the constrained object "chili peppers," the scenario is "ordering Kung Pao Chicken," and the scope is the entire dish. The final result is a structured field such as {negative_constraint:{action: "add," object: "chili peppers," scope: "all," context: "dish: Kung Pao Chicken"}}. In this way, through the above process, the colloquial and easily overlooked exclusionary requirement is transformed into a prohibition rule that the machine can clearly understand and strictly enforce. This fundamentally avoids the problem of execution results deviating significantly from user expectations due to misunderstanding what the user doesn't want, significantly improving the rigor of intent understanding and service satisfaction.

[0048] S32: Combine natural language instructions to perform integrity verification on the overall intent chain composed of multiple intent elements and generate an integrity verification field.

[0049] In this embodiment, the human-computer interaction system proactively detects and completes missing or ambiguous information in user commands to ensure that the intent can be executed. The specific process is as follows: First, the human-computer interaction system uses multiple intent elements to form an overall intent chain to represent a complete user request. At the same time, it determines a preset core element chain, which is a template of key information items that must be included to successfully execute this type of intent. For example, the core element chain of the intent "book a restaurant" includes action, restaurant type, time, and number of people.

[0050] By comparing these two chains, the human-computer interaction system assesses the completeness of the overall intent chain based on the core element chain to identify anomalous intent elements. Anomalous intent elements are those missing from the overall intent chain relative to the core element chain, or those with ambiguity within the overall intent chain, such as the user not specifying the number of diners or saying "evening" without specifying a date. For missing items, the system generates a follow-up question field containing specific questions that need to be clarified with the user.

[0051] In addition, the human-computer interaction system will combine natural language instructions to understand the contextual information of each intent element in order to identify fuzzy intent elements among multiple intent elements and generate fuzzy element fields for fuzzy intent elements. Fuzzy intent elements are intent elements that exist but are vaguely expressed, such as "good" in "find a good restaurant". By generating fuzzy element fields, its uncertainty and possible interpretation range are marked.

[0052] Finally, the human-computer interaction system uses the follow-up question content field and the ambiguous element field as integrity verification fields. For example, the natural language instruction is "Reserve a seat for tomorrow night." The overall intent chain includes the action (reservation), the object (seat), and the time (tomorrow night). However, compared with the core element chain of "restaurant reservation," it is found that the two key objects "restaurant" and "number of people" are missing. At the same time, the instruction does not specify a specific restaurant, and "seat" is an ambiguous object. Therefore, the human-computer interaction system will generate integrity verification fields, which include the follow-up question content {missing_info:["restaurant name or type", "number of diners"]} and the ambiguous element {"object":"seat","note":"must be associated with a specific restaurant"}}.

[0053] In this way, the above process enables pre-diagnosis of the feasibility of the intent, which can proactively discover information gaps and ambiguities and record them through structured fields. This provides a clear basis for subsequent multi-agent judgments on whether to initiate intelligent follow-up questions, enable default values, or conduct probabilistic reasoning. This effectively reduces the awkward interactions caused by incomplete or ambiguous information, such as execution failures, inaccurate results, or multiple confirmations, and improves interaction efficiency.

[0054] S33: Add negative constraint fields and integrity check fields as additional structured intent fields.

[0055] In this embodiment, the human-computer interaction system uses negative constraint fields and integrity verification fields as additional structured intent fields. These additional structured intent fields are independent of multiple step-level structured intent units and are used to carry global, modifying information about "what cannot be done" and "what conditions are needed to do what." For example, for the natural language instruction "Reserve a seat for tomorrow night, no chili peppers in the dish," the final generated additional structured intent field may include: {constraints: {dietary_restriction: "No chili peppers"}, validation:{completeness_check:{missing:["restaurant", "number of people"],ambiguous:{"object": "seat"}}}}. This encapsulates non-core operational processes, but crucial for the correct execution of the task, with supplementary explanations and verification results based on the additional structured intent fields. This separation design makes the core task flow clear while ensuring that all auxiliary and constraint information is transmitted losslessly to the downstream decision engine or various execution agents. This ensures that they fully consider all explicit and implicit user requirements during execution or planning, achieving highly reliable and accurate multi-agent collaboration and decision-making.

[0056] S40: Integrate multiple step-level structured intent units and additional structured intent fields to obtain multi-intent recognition results for natural language instructions. Output the multi-intent recognition results to a multi-agent architecture so that each agent in the multi-agent architecture can process its assigned step-level structured intent unit in combination with the additional structured intent field, generate feedback content for the step-level structured intent unit, and push multiple feedback contents generated by multiple agents for multiple step-level structured intent units to the user for feedback.

[0057] In this embodiment, the human-computer interaction system logically associates and formats multiple step-level structured intent units with additional structured intent fields to generate a unified and complete structured representation as the multi-intent recognition result. The integration process is not a simple stacking; rather, it establishes sequential, parallel, or conditional dependencies between units based on a domain model and correctly applies global constraints from the additional fields to all relevant step units. For example, integrating the "navigate to the company" unit, the "play news" unit, and the additional field "{volume: moderate}" can form a structured task package. This package explicitly includes two parallel executable task sub-items, their respective parameters, and shared preference settings. This transforms a user's free and ambiguous natural language command into a standardized task that the system can accurately understand and seamlessly distribute to multiple agents for collaborative execution.

[0058] Thus, through the above process, the technical solution in this application embodiment outputs a lossless structured instruction set that retains all key details of the original semantics, fundamentally avoiding information loss and execution chaos, providing clear, reliable, and complete input for downstream multi-agent collaborative work, and improving the stability of the entire system and the smoothness of the intelligent experience.

[0059] In this embodiment, the multi-intent recognition result generated through the above steps drives the downstream system to complete the user command. Therefore, optionally, this embodiment may also include the following method: The multi-intent recognition results are output to a multi-agent architecture, which is a system in which multiple agents work collaboratively. Each agent specializes in handling a specific type of task, such as querying, booking, payment, and notification. When the multi-intent recognition results are input, the scheduling module in the multi-agent architecture will accurately assign them to the agents with expertise based on the core intent of each step unit, such as main_intent: restaurant booking, thus achieving parallel task distribution and specialized execution.

[0060] In this process, each agent receiving a task, when processing its assigned step-level structured intent unit, makes a comprehensive judgment by combining it with the additional structured intent fields. That is, the agent not only executes the explicit operation instructions within the step-level structured intent unit, but also proactively adopts global negative constraints and refers to integrity check prompts. For example, when processing a restaurant reservation request, the agent simultaneously reads the negative constraints in the additional fields, such as {dietary_restriction: "No chili"}, and uses this as a mandatory filtering condition when selecting a restaurant or placing an order. At the same time, it also checks for missing information marked in the integrity check fields, such as missing: ["Number of diners"]. If it cannot infer this from the context or historical data, it can proactively generate a clarifying question based on the follow-up question field. Furthermore, based on its expertise, external interfaces connecting to restaurant databases, and adherence to all constraints, the agent generates specific feedback content for that step unit. This feedback might be a successful execution result, such as "Restaurant A has been reserved for you at 7 pm tomorrow for a table of four," or it might be the reason for execution failure or a clarifying question.

[0061] Ultimately, the coordination module in the multi-agent architecture is responsible for collecting multiple feedback responses generated by various agents in response to structured intent units at multiple step levels. This feedback is then sequentially or logically organized and merged to form a coherent and complete response, which is then uniformly pushed to the user for feedback. For example, for the natural language command "Book a restaurant for tomorrow night and check the weather," the feedback from the weather query agent and the restaurant reservation agent will be integrated into "Tomorrow night, Beijing weather will be sunny, 18-25 degrees Celsius; a reservation for restaurant B at 7 pm has been successfully made for you."

[0062] In this way, through the above process, by connecting the structured and lossless intent understanding results with the multi-agent architecture, the user's complex and mixed intents can be decomposed and executed simultaneously and accurately. Each agent works under the guidance of global constraints, ensuring that even if tasks are carried out in parallel, no user exclusionary requirements are violated or key information gaps are ignored. This improves the coordination, accuracy and success rate of task execution, providing users with an efficient, reliable and intelligent service experience that better meets their real needs.

[0063] In summary, the logical process of the technical solution in the embodiments of this application is summarized as follows: Figure 2 As shown, the process includes steps S1 to S6. In step S1, a natural language query input from the user is received. Then, step S2 performs deep semantic parsing, capturing the natural language instruction globally and specifically performing action verb recognition, subject / object extraction, condition / background recognition, and detection of multiple intents / complex semantics. After parsing, step S3 performs instruction decomposition, breaking down the complex natural language instruction into multiple executable steps. This process includes progressively identifying the intention graph and performing semantic normalization on actions and entities. Next, step S4 performs negative constraint identification, locating exclusionary and negative requirements in the user instruction and mapping them to the `negative_constraints` field. Simultaneously, step S5 performs information integrity verification, determining whether the current parsing result lacks necessary elements for execution. If missing, it is marked as requiring further inquiry, i.e., the `need_ask` field is set to `true`, and a specific follow-up inquiry field `ask_content` is generated. Finally, step S6 is executed to integrate the outputs of all the above steps and output a unified, structured intent model (IntentResult), thereby completing the complete logical transformation from free text to a multi-intent structured representation that can be accurately understood and executed by machines.

[0064] In practical applications, an intent recognition layer can be set up in the human-computer interaction system to implement the technical solution of this application. The system architecture of the human-computer interaction system can be as follows: Figure 3As shown, the system architecture of the human-computer interaction system includes a user input layer, an intent recognition layer, a multi-agent execution layer, and an execution result integration layer. The user input layer receives user requests (User Queries) submitted in natural language. These requests are then sent to the intent recognition layer, which includes a deep semantic parsing module, a multi-step intent parsing module, a semantic normalization module, a negative constraint recognition module, an integrity verification and regression module, and an IntentResult generation module. These modules work together to recognize and structure the user's multi-intents. The generated standardized intent results are output to the multi-agent execution layer, which consists of various intelligent agents such as TaskPlanner, Tool Agent, and Code Agent. These agents execute specific tasks in parallel or collaboratively based on the structured intents. The execution outputs of all agents are aggregated, sorted, and formatted by the execution result integration layer, ultimately forming a complete and coherent response returned to the user, thus completing a full-link intelligent interaction from natural language instructions to task results.

[0065] The method provided in this application, through a collaborative processing mechanism of semantic parsing, parallel intent decomposition, negative constraint recognition, and integrity verification, can accurately extract and reconstruct the user's multi-level intent elements from natural language instructions with high degrees of freedom and ambiguous structure. This effectively avoids the limitations of traditional intent recognition methods in understanding complex semantics and implicit logic, thereby significantly improving the integrity and accuracy of intent transmission in multi-agent collaborative links, reducing execution confusion and repeated interactions caused by semantic understanding biases, and enhancing the overall stability, reliability, and intelligent response efficiency of the system.

[0066] Furthermore, as Figure 1 In a specific implementation of the method, this application provides a user multi-intention recognition device for a multi-agent framework, such as... Figure 4 As shown, the device includes: a semantic parsing module 401, a multi-step decomposition module 402, an additional field generation module 403, and a recognition result generation module 404.

[0067] The semantic parsing module 401 is used to receive natural language instructions input by the user and perform semantic parsing to obtain multiple intent elements; The multi-step decomposition module 402 is used to combine the multiple intent elements and execute multiple intent decomposition steps in parallel on the natural language instruction to obtain multiple step-level structured intent units, wherein each intent decomposition step decomposes and parses an independent intent in the natural language instruction. The additional field generation module 403 is used to identify negative constraints and verify the integrity of the multiple intent elements based on the natural language instructions, so as to generate additional structured intent fields. The recognition result generation module 404 is used to integrate the multiple step-level structured intent units and the additional structured intent fields to obtain a multi-intent recognition result for the natural language instruction, and output the multi-intent recognition result to a multi-agent architecture, so that each agent in the multi-agent architecture processes its assigned step-level structured intent unit in combination with the additional structured intent fields, generates feedback content for the step-level structured intent unit, and pushes the multiple feedback contents generated by the multiple agents for the multiple step-level structured intent units to the user for feedback by the multi-agent architecture.

[0068] In a specific application scenario, the semantic parsing module 401 is used to decompose the natural language instruction input by the user upon receiving the natural language instruction, obtaining multiple discrete semantic units; perform context understanding and entity recognition on the multiple discrete language units to determine multiple levels of intent units among the multiple discrete language units, and map each level of intent unit to a corresponding level of structured intent field to obtain multiple levels of structured intent fields; determine multiple other discrete language units besides the multiple levels of intent units among the multiple discrete language units, and map the multiple other discrete language units to corresponding fields according to the content indicated by each of the other discrete language units to obtain multiple other fields, wherein the multiple other fields include at least an action field, a subject field, an object field, a condition field, and an environment field; and use each of the obtained fields as an intent element to obtain the multiple intent elements.

[0069] In a specific application scenario, the multi-step decomposition module 402 is used to perform semantic parsing on the natural language instruction, determine multiple intent boundaries in the natural language instruction, and decompose the natural language instruction according to the multiple intent boundaries to obtain multiple independent intents; set an intent decomposition step for each independent intent, and combine the multiple intent elements to execute the multiple intent decomposition steps corresponding to the multiple independent intents in parallel, so as to generate step-level structured intent units of the corresponding independent intents by executing each intent decomposition step, thereby obtaining the multiple step-level structured intent units.

[0070] In a specific application scenario, the multi-step decomposition module 402 is used to, for each intent decomposition step, determine the specified independent intent processed by the intent decomposition step when executing the intent decomposition step, extract multiple specified intent elements related to the specified independent intent from the multiple intent elements; perform action and entity normalization processing on the multiple specified intent elements to obtain the processed multiple specified intent elements; obtain a preset structured template, and integrate the processed multiple specified intent elements with reference to the structured template to obtain the step-level structured intent unit of the specified independent intent.

[0071] In specific application scenarios, the multi-step decomposition module 402 is used to extract action-related intention elements and entity-related intention elements from the multiple specified intention elements; query the standardized action descriptions corresponding to the action intention elements and the standardized entity descriptions corresponding to the entity intention elements in a pre-built standardized description library, and identify whether the standardized action descriptions and standardized entity descriptions match; if the standardized action descriptions and standardized entity descriptions do not match, combine the context information and historical interaction data in the natural language instructions to perform disambiguation analysis on the standardized action descriptions and standardized entity descriptions to determine the standard semantic form; refer to the standard semantic form, normalize the standardized action descriptions and standardized entity descriptions to ensure that the actions and entities in the multiple specified intention elements are consistent and unambiguous, so as to obtain the processed multiple specified intention elements.

[0072] In specific application scenarios, the additional field generation module 403 includes: The negative constraint recognition submodule is used to determine the negative constraint conditions by recognizing the negative semantic elements included in the multiple intent elements, and to generate a negative constraint field by combining the natural language instruction and the negative constraint conditions. The integrity verification submodule is used to perform integrity verification on the overall intent chain composed of the multiple intent elements in conjunction with the natural language instructions, and generate an integrity verification field. The additional field generation module 403 is further configured to use the negative constraint field and the integrity verification field as the additional structured intent field.

[0073] In specific application scenarios, the negative constraint recognition submodule is used to identify the multiple intent elements, determine the negative semantic elements, wherein the negative semantic elements include negation words, negative expressions, or exclusionary semantic expressions; locate the semantic segment where the negative semantic element is located in the natural language instruction as the negative constraint condition; identify the constrained action and the constrained object of the constrained action in the negative constraint condition; perform context information recognition on the constrained action and the constrained object in the natural language instruction to determine the constrained scenario and the constraint range; and perform structured processing on the constrained action, the constrained object, the constrained scenario, and the constraint range to obtain the negative constraint field.

[0074] In a specific application scenario, the integrity verification submodule is used to construct the overall intent chain using the multiple intent elements and determine a preset core element chain; to perform an integrity assessment on the overall intent chain based on the core element chain to identify abnormal intent elements, and to generate a follow-up question field for the abnormal intent elements. The abnormal intent elements are intent elements that are missing from the overall intent chain relative to the core element chain or intent elements that are ambiguous in the overall intent chain; simultaneously, in conjunction with the natural language instructions, to perform contextual information understanding on each intent element to identify fuzzy intent elements among the multiple intent elements, and to generate a fuzzy element field for the fuzzy intent elements; and to use the follow-up question field and the fuzzy element field as the integrity verification field.

[0075] The apparatus provided in this application, through a collaborative processing mechanism of semantic parsing, parallel intent decomposition, negative constraint recognition, and integrity verification, can accurately extract and reconstruct multi-level intent elements of users from natural language instructions with high degrees of freedom and ambiguous structure. This effectively avoids the limitations of traditional intent recognition methods in understanding complex semantics and implicit logic, thereby significantly improving the integrity and accuracy of intent transmission in multi-agent collaborative links, reducing execution chaos and repeated interactions caused by semantic understanding deviations, and enhancing the overall stability, reliability, and intelligent response efficiency of the system.

[0076] It should be noted that other corresponding descriptions of the functional units involved in the user multi-intention recognition device for a multi-agent framework provided in this application embodiment can be found in the following references. Figures 1 to 3 The corresponding descriptions in [the document] will not be repeated here.

[0077] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.

[0078] The above embodiments and the technical features in the embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0079] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

[0080] In an exemplary embodiment, see Figure 5 The invention also provides a computer device including a bus, a processor, a memory, and a communication interface. It may also include an input / output interface and a display device, wherein the various functional units can communicate with each other via the bus. The memory stores a computer program, and the processor executes the program stored in the memory to perform the user multi-intention recognition method for a multi-agent framework described in the above embodiments.

[0081] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the user multi-intention recognition method for a multi-agent framework.

[0082] Through the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented in hardware or by using software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of this application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, external hard drive, etc.) and includes several instructions to cause a computer device (such as a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0083] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of a preferred embodiment, and the modules or processes shown in the drawings are not necessarily essential for implementing this application.

[0084] Those skilled in the art will understand that the modules in the apparatus of the implementation scenario can be distributed within the apparatus of the implementation scenario as described, or they can be located in one or more apparatuses different from this implementation scenario, with corresponding changes. The modules of the above-described implementation scenario can be combined into one module, or they can be further divided into multiple sub-modules.

[0085] The serial numbers in this application are for descriptive purposes only and do not represent the superiority or inferiority of the implementation scenario.

[0086] The above disclosures are only a few specific implementation scenarios of this application. However, this application is not limited to these. Any variations that can be conceived by those skilled in the art should fall within the protection scope of this application.

Claims

1. A method for recognizing multiple user intents in a multi-agent framework, characterized in that, include: It receives natural language commands from the user and performs semantic parsing to obtain multiple intent elements; Combining the multiple intent elements, multiple intent decomposition steps are executed in parallel on the natural language instruction to obtain multiple step-level structured intent units, wherein each intent decomposition step decomposes and parses an independent intent in the natural language instruction; Based on the natural language instructions, negative constraint conditions are identified and integrity checks are performed on the multiple intent elements to generate additional structured intent fields; The multiple step-level structured intent units and the additional structured intent fields are integrated to obtain a multi-intent recognition result for the natural language instruction. The multi-intent recognition result is output to a multi-agent architecture, so that each agent in the multi-agent architecture processes its assigned step-level structured intent unit in combination with the additional structured intent field, generates feedback content for the step-level structured intent unit, and the multi-agent architecture pushes the multiple feedback contents generated by the multiple agents for the multiple step-level structured intent units to the user for feedback.

2. The method according to claim 1, characterized in that, The process involves receiving natural language commands input from the user and performing semantic parsing to obtain multiple intent elements, including: When the natural language instruction input by the user is received, the natural language instruction is decomposed to obtain multiple discrete semantic units; Context understanding and entity recognition are performed on the multiple discrete language units to determine multiple levels of intent units in the multiple discrete language units, and each level of intent unit is mapped to the corresponding level of structured intent field to obtain multiple levels of structured intent fields. Among the plurality of discrete language units, a plurality of other discrete language units besides the plurality of intent units at the plurality of levels are determined. Based on the content indicated by each of the other discrete language units, the plurality of other discrete language units are mapped to corresponding fields to obtain a plurality of other fields. The plurality of other fields include at least an action field, a subject field, an object field, a condition field, and an environment field. Each field obtained is used as an intent element to obtain the multiple intent elements.

3. The method according to claim 1, characterized in that, The method involves combining the multiple intent elements and performing multiple intent decomposition steps in parallel on the natural language instruction to obtain multiple step-level structured intent units, including: The natural language instruction is semantically parsed to determine multiple intent boundaries in the natural language instruction, and the natural language instruction is decomposed according to the multiple intent boundaries to obtain multiple independent intents; For each independent intent, an intent decomposition step is set, and the multiple intent elements are combined and the multiple intent decomposition steps corresponding to the multiple independent intents are executed in parallel, so as to generate the step-level structured intent unit of the corresponding independent intent by executing each intent decomposition step, thereby obtaining the multiple step-level structured intent units.

4. The method according to claim 3, characterized in that, The step of combining the multiple intent elements and executing the multiple intent decomposition steps corresponding to the multiple independent intents in parallel, so as to generate step-level structured intent units of corresponding independent intents by executing each intent decomposition step, includes: For each intent decomposition step, when performing the intent decomposition step, the specified independent intent processed by the intent decomposition step is determined, and multiple specified intent elements related to the specified independent intent are extracted from the multiple intent elements. The multiple specified intent elements are normalized in terms of action and entity to obtain the processed multiple specified intent elements. Obtain a preset structured template, and integrate the processed multiple specified intent elements with reference to the structured template to obtain the step-level structured intent unit of the specified independent intent.

5. The method according to claim 4, characterized in that, The normalization process of the multiple specified intent elements for actions and entities, to obtain the processed multiple specified intent elements, includes: Extract the action intent elements related to the action and the entity intent elements related to the entity from the multiple specified intent elements; In a pre-built standardized description library, query the standardized action description corresponding to the action intent element, the standardized entity description corresponding to the entity intent element, and identify whether the standardized action description and the standardized entity description match. If a mismatch is detected between the standardized action description and the standardized entity description, disambiguation analysis is performed on the standardized action description and the standardized entity description in combination with the context information and historical interaction data in the natural language instruction to determine the standard semantic form. Referring to the standard semantic form, the standardized action description and the standardized entity description are normalized to ensure that the actions and entities in the multiple specified intent elements are consistent and unambiguous, so as to obtain the processed multiple specified intent elements.

6. The method according to claim 1, characterized in that, The step of identifying negative constraints and verifying completeness of the multiple intent elements based on the natural language instructions to generate additional structured intent fields includes: By identifying the negative semantic elements included in the plurality of intent elements, the negative constraint conditions are determined, and a negative constraint field is generated by combining the natural language instruction and the negative constraint conditions. In conjunction with the natural language instructions, the integrity of the overall intent chain composed of the multiple intent elements is verified, and an integrity verification field is generated. The negative constraint field and the integrity check field are used as the additional structured intent field.

7. The method according to claim 6, characterized in that, The step of identifying the negative semantic elements included in the plurality of intent elements, determining the negative constraint conditions, and generating a negative constraint field by combining the natural language instruction and the negative constraint conditions includes: The multiple intent elements are identified to determine the negative semantic elements, wherein the negative semantic elements include negation words, negative expressions, or exclusionary semantic expressions; In the natural language instruction, the semantic segment where the negative semantic element is located is located is used as the negative constraint condition, and the constrained action and the constrained object to which the constrained action acts are identified in the negative constraint condition. In the natural language instructions, the context information of the constrained action and the constrained object is identified to determine the constrained scenario and the scope of the constraint. The constrained action, the constrained object, the constrained scenario, and the constraint range are structured to obtain the negative constraint field.

8. A user multi-intention recognition device for a multi-agent framework, characterized in that, include: The semantic parsing module is used to receive natural language commands input by the user and perform semantic parsing to obtain multiple intent elements; The multi-step decomposition module is used to combine the multiple intent elements and execute multiple intent decomposition steps in parallel on the natural language instruction to obtain multiple step-level structured intent units, wherein each intent decomposition step decomposes and parses an independent intent in the natural language instruction. The additional field generation module is used to identify negative constraints and verify the integrity of the multiple intent elements based on the natural language instructions, so as to generate additional structured intent fields. The recognition result generation module is used to integrate the multiple step-level structured intent units and the additional structured intent fields to obtain a multi-intent recognition result for the natural language instruction. The multi-intent recognition result is then output to a multi-agent architecture, so that each agent in the multi-agent architecture processes its assigned step-level structured intent unit in conjunction with the additional structured intent fields, generates feedback content for the step-level structured intent unit, and pushes the multiple feedback contents generated by the multiple agents for the multiple step-level structured intent units to the user for feedback.

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 method according to any one of claims 1 to 7.

10. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images