Intelligent task arrangement system and method based on intention recognition

Abstract

The invention relates to an intelligent task arrangement system and method based on intention recognition, and aims to solve the problems of insufficient semantic understanding fineness, poor task arrangement dynamic adaptability, lack of fault-tolerant backspacing mechanisms and the like in the prior art. Comprising an intention analysis module, a task decomposition module, a parallel scheduling module and a result aggregation module. According to the design, the intention understanding accuracy and fineness are improved, intelligent dynamic arrangement of tasks is achieved, the fault-tolerant capability under uncertainty is enhanced, and the parallel processing efficiency and stability are improved.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and natural language processing, and in particular to an intelligent task orchestration system and method based on intent recognition. Background Technology

[0002] Currently, artificial intelligence and natural language processing technologies are increasingly permeating many aspects of daily life, especially in intelligent customer service, intelligent assistants, intelligent recommendation systems, and emerging vertical applications such as AI in clothing, where their importance is growing daily. These systems aim to understand user intent and automate complex tasks by simulating human intelligent interaction, thereby greatly improving user experience and operational efficiency. For example, patent document CN119313076A, entitled "A Satellite Internet Task Planning Method Based on Intent Understanding and Large Model," discloses a satellite internet task planning method based on intent understanding and a large model. By parsing and understanding communication intent based on a large model, the method decomposes the communication intent into sub-tasks and further plans the tasks. This solves the problem of low efficiency in traditional task planning for large-scale constellation management, effectively improving total resource utilization and service capacity while achieving load balancing. It avoids waste and conflicts of on-board resources in the satellite internet system, thus improving the user service quality (QoS). Furthermore, research is being conducted on intent-based satellite internet mission planning to achieve intelligent mission resolution. This upgrades the traditional mission planning method, which involves manually converting communication requirements or manually specifying the network type required for a mission, to intelligent and automatic mission planning based on communication intent.

[0003] However, the aforementioned technical solutions and other existing technologies still suffer from several shortcomings. They cannot effectively handle multi-layered and complex user intents in terms of the depth and granularity of semantic understanding; they cannot intelligently decompose and schedule tasks according to complex needs in terms of the dynamics and adaptability of task orchestration; they lack reliable fault tolerance and rollback strategies to cope with information uncertainty; and they fail to fully utilize parallel computing capabilities to improve task execution efficiency. Therefore, it is necessary to develop a task orchestration mechanism that employs newer technical designs, can intelligently decompose and dynamically schedule tasks, and possesses efficient parallel processing capabilities, while ensuring that the system can still provide reliable fault tolerance and rollback strategies under conditions of information uncertainty, thereby solving the aforementioned problems. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent task orchestration system and method based on intent recognition, in order to solve the problems in the background art, such as the inability to effectively handle multi-level and complex user intents in terms of the depth and fineness of semantic understanding; the inability to intelligently decompose and schedule tasks according to complex needs in terms of the dynamics and adaptability of task orchestration; the lack of reliable fault tolerance and rollback strategies in dealing with information uncertainty; and the failure to fully utilize parallel computing capabilities to improve task execution efficiency.

[0005] This invention provides the following technical solution: an intelligent task orchestration system based on intent recognition, comprising an intent parsing module, a task decomposition module, a parallel scheduling module, and a result aggregation module. The graph parsing module is used to receive user input in natural language form and perform natural language understanding processing, entity and slot extraction, and confidence calculation. The task decomposition module is used to receive the intent tags and structured slot information output by the graph parsing module, and perform hierarchical task network planning, task directed acyclic graph generation, and constraint propagation processing based on the received intent tags and structured slot information. The parallel scheduling module is used to receive the directed acyclic graph of tasks and the constraint information output by the task decomposition module, and to execute critical path priority scheduling, confidence gating mechanism, resource upper limit control and execution layer protocol implementation. The result aggregation module is used to receive the execution results of the parallel scheduling module and perform multi-branch result collection, consistency verification, and execution report generation.

[0006] A preferred intelligent task orchestration method based on intent recognition includes the following steps: S1. Intent parsing: Receives user input in natural language form, performs natural language understanding processing using a BERT-like pre-trained language model, extracts entities and slots using a sequence labeling model, and calculates confidence scores for the identified intent and extracted slot information. S2. Task decomposition: Based on the intent tags and structured slot information output by the intent parsing module, the complex task is recursively decomposed into atomic tasks using hierarchical task network planning, and a directed acyclic graph of the task is constructed to clarify the execution order and dependencies between tasks, while propagating and processing various constraints. S3. Parallel Scheduling: Based on the directed acyclic graph of the tasks and the constraint information, the critical path priority strategy is adopted to schedule task execution. At the same time, the confidence of slot position is dynamically monitored through a confidence gating mechanism. When the confidence is lower than a preset threshold, a user confirmation or system rollback process is triggered, and the system resource upper limit is controlled. The execution layer protocol is completed through idempotent actions and transaction compensation strategies. S4. Result Aggregation: Collect the results of all parallel execution branches, standardize the data format and verify logical consistency, and generate a detailed task execution report, including execution trajectory, key decisions and any exception handling records.

[0007] Preferably, step S1. Intent parsing, which utilizes a BERT-like pre-trained language model for natural language understanding processing, includes the following steps: This method achieves bidirectional contextual encoding of text by using a self-attention mechanism. Its internal architecture consists of multiple stacked Transformer encoders, each composed of a multi-head self-attention mechanism and a feedforward neural network. Residual connections and layer normalization techniques are used to improve training stability. The masked language model and next-sentence prediction task are pre-trained on a large text corpus, and fine-tuned on a labeled dataset containing domain-specific terms and expressions when applied to specific domains. Task-specific prediction is achieved by introducing one or more linear classification layers.

[0008] Preferably, the entity and slot extraction in step S1. intent parsing includes the following steps: A multi-task learning framework is adopted, wherein the task-specific prediction includes an intent classification task and a slot filling task. The intent classification task and the slot filling task share the parameters of the underlying BERT-like pre-trained language model and perform prediction through their respective independent output layers. The slot filling task specifically adopts the Span-BERT architecture, which can directly predict continuous spans in the text as entities or slot values, effectively handling overlapping and nested entities. For implicit slots, the semantic reasoning module infers and fills logically necessary slots that are not explicitly expressed based on explicit information provided by the user and the domain knowledge graph. Preferably, step S1. Calculating the confidence score for the identified intent and extracted slot information in intent parsing includes the following steps: The identified intent and extracted slot information are subjected to posterior probability and prediction uncertainty estimation based on Monte Carlo Dropout. For each slot, the softmax probability value of the predicted slot label is calculated as an initial confidence level. By running the model with the Dropout layer multiple times during the inference phase, multiple prediction results are obtained, and then the entropy or standard deviation of the prediction results is calculated to quantify the model's "uncertainty" regarding the prediction. The final confidence score is a weighted combination of the softmax probability and the uncertainty measure.

[0009] Preferably, the hierarchical task network planning steps in step S2. task decomposition include: Based on the intent tag and slot information obtained from the intent parsing in S1, the high-level task is decomposed by the task decomposition module and stored in the domain knowledge base in a formal language. The task decomposition module defines the specific steps for decomposing a high-level task into a set of subtasks or atomic operations, and includes preconditions and effects.

[0010] Preferably, the construction of the directed acyclic graph of the task in step S2. task decomposition includes the following steps: The task is managed and persisted using a graph database. The directed acyclic graph (DAG) includes node attributes and edge attributes. The node attributes of the task DAG include task ID, task type, estimated execution time, input parameters, output parameters, and current state. The edge attributes of the task DAG include dependency type, source task ID, and target task IP. The subtasks or atomic operations output by the hierarchical task network planning are converted into an initial set of nodes, and then directed edges are gradually added through the hierarchical task network planning.

[0011] Preferably, step S2. Propagating and processing various constraints in task decomposition includes the following steps: For sequence constraints, virtual synchronization nodes are added or task priorities are adjusted in the directed acyclic graph of tasks at the hierarchical level to ensure that predecessor tasks complete before successor tasks. For mutual exclusion constraints, mutually exclusive tasks are prevented from executing simultaneously by allocating them to non-overlapping time windows or by using resource locking mechanisms. For resource limit constraints, when constructing the directed acyclic graph of tasks, the task decomposition module performs preliminary resource reservation or capacity checks on tasks with high resource requirements in advance, and optimizes resource allocation through a capacity-based network flow algorithm to ensure that concurrent resource usage at any time does not exceed the system-set threshold.

[0012] Preferably, the working steps of the confidence gating mechanism in step S3, parallel scheduling, include: The system continuously monitors the task execution events and slot position confidence changes in the parallel scheduling module; the preset dynamic threshold is trained and predicted by a machine learning model based on the task's criticality, business scenario, and historical performance data; triggering the confidence gating operation includes sending a confirmation request to the user or initiating a system rollback process. The confirmation request includes providing the user with key information currently understood by the system and possible alternative options, and collecting user feedback to update the slot information and slot position confidence.

[0013] Preferably, the execution layer protocol in step S3. parallel scheduling includes the following steps: The idempotent action is achieved by introducing a unique request identifier in each atomic service interface. On the service side, before each request is processed, a distributed cache system or database is queried to check whether the request identifier has been processed, so as to avoid repeated execution of business logic. The transaction compensation strategy is implemented based on the Saga pattern. Each business process is decomposed into a series of small, independent local transactions. Each local transaction publishes an event upon completion. If a local transaction fails or needs to be rolled back, the local transaction publishes a "compensation event" to trigger the compensation operation of the previously successfully executed local transactions.

[0014] Compared with the prior art, the beneficial effects of the present invention are: 1. The intelligent task orchestration system based on intent recognition of the present invention includes an intent parsing module, a task decomposition module, a parallel scheduling module, and a result aggregation module. The graph parsing module is used to receive user input in natural language form and perform natural language understanding processing, entity and slot extraction, and confidence calculation. The task decomposition module is used to receive intent tags and structured slot information output by the graph parsing module, and perform hierarchical task network planning, task directed acyclic graph generation, and constraint propagation processing based on the received intent tags and structured slot information. The parallel scheduling module is used to receive the task directed acyclic graph and constraint information output by the task decomposition module, and perform critical path priority scheduling, confidence gating mechanism, resource upper limit control, and implementation of execution layer protocol. The result aggregation module is used to receive the execution results of the parallel scheduling module and perform multi-branch result collection, consistency verification, and execution report generation. Through the coordinated operation of the above modules, it has outstanding performance advantages in terms of accuracy when handling complex and ambiguous intentions, error rate under uncertain conditions, and overall task completion efficiency. Through efficient parallel scheduling, it effectively improves the accuracy of intention understanding, significantly improves the success rate of understanding complex and ambiguous intentions, and greatly reduces the error rate when facing uncertain information, thus greatly enhancing the robustness and reliability of the system.

[0015] 2. The intelligent task orchestration method based on intent recognition of the present invention includes the following steps: S1. Intent parsing, S2. Task decomposition, S3. Parallel scheduling, and S4. Result aggregation. Through the above steps, the accuracy of intent understanding can be effectively improved, and the multi-dimensional needs of users can be accurately identified; execution efficiency can be effectively improved, the quality of results can be improved, the generated matching scheme meets the user's constraints, and it is traceable, providing a detailed explanation of the decision-making process.

[0016] 3. The intelligent task orchestration method based on intent recognition in this invention improves the accuracy and precision of intent understanding: By employing BERT-like pre-trained language models for deep natural language understanding and advanced technologies such as Span-BERT for entity and slot extraction, the accuracy of understanding complex and multi-dimensional user intents is significantly improved. Compared to traditional solutions based on keyword matching or shallow models, this method can effectively capture the deep semantic structure, multi-level information, and inherent logical relationships in sentences, effectively improving the accuracy of understanding complex and compound intents. Moreover, the ability to process nested entities and implicit slots further enhances the precision of semantic parsing.

[0017] 4. The intelligent task orchestration method based on intent recognition of this invention achieves intelligent and dynamic task orchestration: by introducing hierarchical task network planning and a directed acyclic graph generation mechanism, it can dynamically decompose high-level tasks into a series of executable atomic tasks based on the identified complex intents and their accompanying fine-grained constraints, and construct an optimized task execution flow. Compared with existing rigid fixed flows, it can intelligently adjust the task execution strategy according to specific needs, realize adaptive task decomposition and parallel scheduling, thereby improving task completion efficiency and making full use of the parallel processing capabilities of modern computing resources.

[0018] 5. The intelligent task orchestration method based on intent recognition in this invention significantly enhances the system's fault tolerance under uncertainty: it innovatively introduces a slot position confidence-driven confidence gating mechanism. When processing ambiguous or low-confidence critical information, the system can intelligently pause execution, triggering user confirmation or system rollback processes, preventing misunderstandings from propagating down the task chain. Combined with idempotent actions and transaction compensation strategies in the execution layer protocol, this method effectively reduces the error rate when dealing with task execution failures or uncertain information, greatly improving the system's robustness and reliability.

[0019] 6. The intelligent task orchestration method based on intent recognition in this invention significantly improves the efficiency and stability of parallel processing: Through the critical path priority scheduling algorithm, this invention prioritizes the execution of critical tasks that have the greatest impact on the overall task completion time, effectively shortening the total task execution cycle. Simultaneously, combined with resource limit control, the system can dynamically balance task load and available resources, avoiding resource overload and ensuring efficient concurrent processing capabilities, effectively enhancing the system's parallel processing capacity. The mandatory requirements of idempotency and transaction compensation mechanisms further guarantee the system's stability under various abnormal conditions, maintaining data consistency even in scenarios such as network jitter and service restarts.

[0020] 7. The intelligent task orchestration method based on intent recognition of this invention provides comprehensive traceability and optimizability: the detailed execution report generated by the result aggregation module records the entire link trajectory of the task from intent recognition to final completion, all key decision points, confidence changes, and any anomaly handling processes. This provides a valuable data foundation and transparency for system problem diagnosis, performance bottleneck analysis, and further system optimization, enabling the system to continuously iterate and improve, and constantly enhance its intelligence level and user experience. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart illustrating the intelligent task orchestration system and method based on intent recognition of the present invention. Detailed Implementation

[0023] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain this application, and should not be construed as limiting this application.

[0024] In the description of this application, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, features defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0025] In the description of this application, it should be noted that, unless otherwise expressly specified and limited, the terms "installation" and "connection" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection, an electrical connection, or a connection that allows communication between the components; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication between two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.

[0026] In the description of this application, it should be understood that the terms "upper", "lower", "side", "front", "rear", etc., indicate the orientation or positional relationship based on the installation orientation or positional relationship, and are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this application.

[0027] In the description of this application, it should be noted that the term "and / or" is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can represent three situations: A exists alone, A and B exist simultaneously, and B exists alone.

[0028] It should also be noted that in the embodiments of this application, the same reference numerals are used to represent the same component or part. For the same part in the embodiments of this application, the reference numerals may only be used to mark one part or component as an example. It should be understood that the reference numerals are also applicable to other identical parts or components.

[0029] Furthermore, in this application, the technical features described in an open-ended manner include both closed technical solutions composed of the listed features and open technical solutions that include the listed features.

[0030] To further understand the invention's content, features, and effects, the following embodiments are provided, along with detailed descriptions in conjunction with the accompanying drawings: like Figure 1 As shown, the intelligent task orchestration system and method based on intent recognition includes an intent parsing module, a task decomposition module, a parallel scheduling module, and a result aggregation module. The graph parsing module is used to receive user input in natural language form and perform natural language understanding processing, entity and slot extraction, and confidence calculation. The task decomposition module receives the intent labels and structured slot information output by the graph parsing module, and performs hierarchical task network planning, task directed acyclic graph generation, and constraint propagation processing based on the received intent labels and structured slot information. The parallel scheduling module is used to receive the directed acyclic graph and constraint information of the tasks output by the task decomposition module, and to execute critical path priority scheduling, confidence gating mechanism, resource upper limit control and execution layer protocol implementation; The results aggregation module receives the execution results from the parallel scheduling module and performs multi-branch result collection, consistency verification, and execution report generation.

[0031] In this embodiment, through the coordinated operation of the various modules, it exhibits superior performance advantages in terms of accuracy when processing complex and ambiguous intents, error rate under uncertain conditions, and overall task completion efficiency. Efficient parallel scheduling effectively improves the accuracy of intent understanding, significantly enhancing the success rate of understanding complex and ambiguous intents. When facing uncertain information, the error rate is drastically reduced, greatly enhancing the robustness and reliability of the system. Of course, this is not a limitation; other methods that can enhance the robustness and reliability of the system are also possible.

[0032] Specifically, the intelligent task orchestration method based on intent recognition includes the following steps: S1. Intent Resolution: Receive user input in natural language form, perform natural language understanding processing using a BERT-like pre-trained language model, extract entities and slots using a sequence labeling model, and calculate confidence scores for the identified intent and extracted slot information. In this embodiment, a BERT (meaning bidirectional encoder representation)-like pre-trained language model based on the Transformer architecture is used as the basic encoder. After pre-training on massive corpora and domain-specific fine-tuning, it can effectively capture lexical, syntactic, and semantic information. This system simultaneously performs intent classification and slot filling using a multi-task learning framework. It introduces a Span-BERT (segment-based BERT improvement) structure to identify entities across continuous text spans, supporting the handling of nested and implicit slots. Furthermore, it combines predicted probabilities with Monte Carlo Dropout (a method for calculating probability distributions through over-random sampling; Dropout is a random deactivation method) to calculate the confidence score for each slot and intent. The final output is structured data containing rich semantic information and confidence scores for the identified intents and extracted slots, which can then be used by the task decomposition module. However, it is not limited to this; any other method capable of outputting structured data containing rich semantic information and confidence scores for the identified intents and extracted slots can be used. S2. Task Decomposition: Based on the intent tags and structured slot information output by the intent parsing module, complex tasks are recursively decomposed into atomic tasks using hierarchical task network planning, and a directed acyclic graph (DAG) of tasks is constructed. This clarifies the execution order and dependencies between tasks, while simultaneously propagating and handling various constraints. In this embodiment, hierarchical task network planning recursively decomposes high-level tasks into atomic operation sequences based on user intent tags and slot information, including preconditions and effects to ensure task rationality. The planning results are further organized into a DAG, which includes node attributes and edge attributes. Node attributes represent tasks, and edge attributes represent dependencies. This is managed through a graph database. During the construction of the DAG, the system identifies and propagates various constraints, including sequence constraints, mutual exclusion, and resource constraints. Conflicts are resolved by adding synchronization nodes, reserving resources, or backtracking adjustments. Finally, an executable task graph satisfying all constraints is output. If conflicts still exist, they are dynamically handled by the parallel scheduling module. However, this is not limited to this; any other method capable of outputting an executable task graph satisfying all constraints can be used. S3. Parallel Scheduling: Based on the directed acyclic graph of tasks and constraint information, a critical path priority strategy is adopted to schedule task execution. Simultaneously, a confidence gating mechanism dynamically monitors the slot position confidence. When the confidence falls below a preset threshold, a user confirmation or system rollback process is triggered, and system resource upper limit control is implemented. The execution layer protocol is completed through idempotent actions and transaction compensation strategies. In this embodiment, a critical path priority scheduling algorithm is used. The critical path is dynamically determined by calculating the earliest start time and latest finish time of tasks, prioritizing the scheduling of tasks with the shortest relaxation time to shorten the total execution time. A confidence gating mechanism is introduced to dynamically trigger user confirmation or... The system rollback mechanism avoids executing critical operations based on low-confidence information; it implements resource cap control, monitoring and allocating resources such as CPU, memory, and API quotas through a centralized resource manager to prevent system overload and support elastic scaling; it defines an execution layer protocol that requires tasks to be idempotent and avoids duplicate operations through unique request identifiers, while implementing a transaction compensation mechanism based on the Saga pattern (meaning compensatory transaction pattern) to roll back to a consistent state in case of failure; and it achieves efficient, stable, and adaptive parallel task scheduling through multi-mechanism collaboration. Of course, it is not limited to these; any other mechanism capable of achieving efficient, stable, and adaptive parallel task scheduling is acceptable. S4. Result Aggregation: Collect the results of all parallel execution branches, standardize the data format, verify logical consistency, and generate a detailed task execution report, including the execution trajectory, key decisions, and any exception handling records. In this embodiment, branch result collection is achieved asynchronously through a message queue. Each subtask publishes its standardized results to a designated topic, where the aggregation module caches and transforms them to ensure consistency of heterogeneous data. The consistency verification module, based on a rule engine or ontology reasoning, detects logical or semantic conflicts between results and resolves them according to preset strategies. The execution report generation module stores the entire process data in a time-series database or data warehouse, supporting the generation of customizable, multi-dimensional reports and traceable logs, providing key performance indicator analysis, and supporting API integration with external monitoring and analysis tools. Of course, this is not limited to this; other modules capable of achieving the same function can be used.

[0033] Specifically, step S1. Intent parsing, which utilizes a BERT-like pre-trained language model for natural language understanding, includes the following steps: By using a self-attention mechanism to achieve bidirectional contextual encoding of text, its internal architecture consists of multiple stacked Transformer encoders. Each Transformer encoder consists of a multi-head self-attention mechanism and a feedforward neural network. Residual connections and layer normalization techniques are used to improve training stability. The masked language model and next-sentence prediction task are pre-trained on a large text corpus, and fine-tuned on a labeled dataset containing domain-specific terms and expressions when applied to specific domains. Task-specific prediction is achieved by introducing one or more linear classification layers.

[0034] In this embodiment, the internal architecture of the BERT-like pre-trained language model consists of stacked multi-layer Transformer encoders. Each encoder layer includes a multi-head self-attention mechanism and a feedforward neural network. Through residual connections and layer normalization techniques, the vanishing and exploding gradient problems in deep network training are effectively mitigated, significantly improving training stability and model performance. To adapt to specific domain applications, this pre-trained model is fine-tuned on labeled datasets containing domain-specific terms and expressions. During fine-tuning, one or more linear classification layers are typically added on top of the output of the BERT-like pre-trained language model to perform task-specific predictions, including intent classification and slot filling. This deep learning architecture ensures that the system can generate embedding vectors containing rich contextual semantics and syntactic information, thereby achieving a refined understanding of user intent. Of course, it is not limited to this; other methods for achieving a refined understanding of user intent can also be used.

[0035] Specifically, step S1, entity and slot extraction in intent parsing, includes the following steps: A multi-task learning framework is adopted, with task-specific predictions including intent classification and slot filling tasks. The intent classification and slot filling tasks share the parameters of the underlying BERT-like pre-trained language model and perform predictions through their respective independent output layers. The slot filling task specifically adopts the Span-BERT architecture, which can directly predict continuous spans in the text as entities or slot values, effectively handling overlapping and nested entities. For implicit slots, the semantic reasoning module infers and fills slots that are not explicitly expressed but are logically necessary based on explicit information provided by the user and the domain knowledge graph.

[0036] In this embodiment, a multi-task learning framework is employed. Task-specific predictions include intent classification and slot filling tasks. These tasks share parameters from the underlying BERT-like pre-trained language model, enabling knowledge transfer and parameter sharing, thus improving overall performance. The slot filling task specifically adopts the Span-BERT architecture, which can directly predict consecutive spans in text as entities or slot values, effectively handling overlapping and nested entities. The slot structure is a standardized information template used to store specific types of parameter information. To handle nested entities and implicit slots, recursive parsing and contextual reasoning mechanisms are introduced. For implicit slots, the semantic reasoning module infers and fills logically necessary but unexpressed slots based on explicit information provided by the user and the domain knowledge graph, assigning a low initial confidence level to prompt user confirmation. Of course, this is not limited to this; other methods capable of achieving the same function can be used.

[0037] Specifically, step S1. Intent parsing involves calculating a confidence score based on the identified intent and extracted slot information, including the following steps: The identified intent and extracted slot information are subjected to posterior probability and prediction uncertainty estimation based on Monte Carlo Dropout. For each slot, the softmax probability value of the slot prediction label is calculated as an initial confidence score. Multiple prediction results are obtained by running the model with Monte Carlo Dropout layer multiple times during the inference stage. Then, the entropy or standard deviation of the prediction results is calculated to quantify the model's "uncertainty" for the prediction. The final confidence score is a weighted combination between the softmax probability and the uncertainty measure.

[0038] In this embodiment, the confidence calculation mechanism combines the posterior probability of the model output with a Monte Carlo Dropout-based estimation of prediction uncertainty. For each slot, the system first calculates the softmax (normalized exponential function) probability value of its predicted label and uses it as the initial confidence score. To quantify the model's "uncertainty" regarding the prediction, multiple prediction results are obtained by running the model with a Monte Carlo Dropout layer multiple times during the inference phase. Then, the entropy or standard deviation of these prediction results is calculated; the higher the entropy or standard deviation, the greater the uncertainty of the model's prediction. The final confidence score is a weighted combination of the softmax probability and the uncertainty measure. This weighting can be optimized using supervised learning methods based on cross-validation and actual performance to achieve a balance between accuracy and robustness. The output of the intent parsing module is a structured data packet, including one or more intent labels and their corresponding confidence scores, as well as a structured set of slot information. Each slot contains its extracted value, original text span, start and end positions, and corresponding confidence score. Of course, this is not limited to this; other mechanisms that achieve a balance between accuracy and robustness can be used.

[0039] Specifically, the steps in step S2, hierarchical task network planning during task decomposition, include: Based on the intent tags and slot information obtained from S1. Intent parsing, the high-level task is decomposed through the task decomposition module and stored in the domain knowledge base in a formal language. The task decomposition module defines the specific steps to decompose a high-level task into a set of subtasks or atomic operations, and includes preconditions and effects.

[0040] In this embodiment, based on the intent tags and slot information obtained from the intent parsing module, a high-level task is decomposed. This decomposition is stored in a domain knowledge base using a formal language (e.g., PDDL or a custom JSON / YAML format). Each formal language defines the specific steps for decomposing a high-level task into a set of subtasks or atomic operations, and includes preconditions and effects. The hierarchical task network planning obtains definitions of tasks, methods, operations, and states by consulting a "Task Planning Domain Definition Language" file containing domain-specific knowledge. This file embeds knowledge about clothing types, material properties, color theory, matching principles, and budget constraints, ensuring the rationality and professionalism of the task decomposition. The hierarchical task network planning employs a depth-first search-based planning algorithm (e.g., a variant of the SHOP2 planner) to search the search space for a plan that satisfies all preconditions and can decompose the initial task into a sequence of atomic operations. Of course, any other algorithm capable of decomposing a high-level task into subtasks can be used.

[0041] Specifically, step S2. The construction of the directed acyclic graph of tasks in task decomposition includes the following steps: The task directed acyclic graph (DAG) is managed and persisted using a graph database. The DAG includes node attributes and edge attributes. The node attributes of the task DAG include task ID, task type, estimated execution time, input parameters, output parameters, and current state. The edge attributes of the task DAG include dependency type, source task ID, and target task IP. The subtasks or atomic operations output by hierarchical task network planning are converted into an initial set of nodes, and then directed edges are gradually added through hierarchical task network planning.

[0042] In this embodiment, the construction of the Directed Acyclic Graph (DAG) is managed and persisted through a graph database (e.g., Neo4j or JanusGraph), which can efficiently represent and query task nodes and their dependent edges. Each node attribute in the DAG represents an atomic task or subtask, and each directed edge attribute represents a dependency relationship, meaning the source node task must be completed before the target node task. Dependencies can be data dependencies, control dependencies, or resource dependencies. Node attributes in the DAG include task ID, task type, estimated execution time, input parameters, output parameters, and current state; edge attributes include dependency type, source task ID, and target task ID. The DAG generation process first converts the atomic task sequence output by the HTN planner into an initial node set, and then, based on the data flow and control flow relationships defined in the HTN method, gradually adds directed edge attributes until all task dependencies are accurately represented. Of course, this is not limited to this; any other method that clearly defines the execution order and dependencies between tasks can be used.

[0043] Specifically, step S2. Propagating and processing various constraints in task decomposition includes the following steps: For sequence constraints, virtual synchronization nodes are added to the directed acyclic graph of tasks or task priorities are adjusted to ensure that predecessor tasks complete before successor tasks. For mutual exclusion constraints, mutually exclusive tasks are prevented from executing simultaneously by allocating them to non-overlapping time windows or by using resource locking mechanisms. For resource ceiling constraints, when constructing the hierarchical task network plan, the task decomposition module performs preliminary resource reservation or capacity checks on tasks with high resource requirements in advance, and optimizes resource allocation through a capacity-based network flow algorithm to ensure that the concurrent resource usage at any time does not exceed the system's set threshold.

[0044] In this embodiment, constraint propagation processing includes: for sequential constraints, ensuring that predecessor tasks complete before successor tasks by adding virtual synchronization nodes or adjusting task priorities in the task directed acyclic graph. For mutual exclusion constraints, preventing simultaneous execution of mutually exclusive tasks by allocating them to non-overlapping time windows or utilizing resource locking mechanisms, for example, ensuring atomic access to shared memory in a multi-threaded environment. For resource ceiling constraints, the task decomposition module performs preliminary resource reservation or capacity checks for tasks with high resource requirements when constructing the task directed acyclic graph, and optimizes resource allocation using a capacity-based network flow algorithm to ensure that concurrent resource usage does not exceed a system-set threshold at any given time. When a constraint conflict is detected (e.g., two tasks both require exclusive access to the same scarce resource but are planned to execute in parallel), the task decomposition module initiates a backtracking mechanism, reselects a hierarchical task network planning method, or adjusts the task decomposition path until a valid task directed acyclic graph satisfying all constraints is generated. If the conflict cannot be completely resolved during the decomposition phase, the task decomposition module sends a task directed acyclic graph with conflict warnings to the parallel scheduling module, which then dynamically processes it at runtime. Of course, it is not limited to this; it can be any other tool that can identify, represent, and propagate constraints between tasks during the construction of a directed acyclic graph of tasks.

[0045] Specifically, step S3, the working steps of the confidence gating mechanism in parallel scheduling, include: By continuously monitoring task execution events and slot position confidence changes in the parallel scheduling module; preset dynamic thresholds are trained and predicted by machine learning models based on task criticality, business scenarios, and historical performance data; triggering confidence gating operations includes sending confirmation requests to users or initiating system rollback processes. Confirmation requests include providing users with key information currently understood by the system and possible alternative options, and collecting user feedback to update slot information and slot position confidence.

[0046] In this embodiment, the confidence gating mechanism is implemented in a dedicated decision service, continuously monitoring task execution events and slot position confidence changes in the parallel scheduling module. Preset dynamic thresholds can be adaptively adjusted based on task type, business criticality, and historical data. A reinforcement learning-based strategy can learn at what confidence level to trigger user confirmation or rollback to maximize task success rate and user satisfaction. Triggered gating operations include, but are not limited to, sending a confirmation request to the user or initiating a rollback process. The user confirmation request includes providing the user with key information currently understood by the system and possible alternatives, and collecting user feedback to update slot information and its confidence level. When a rollback process is triggered, the decision service sends a rollback instruction to the parallel scheduling module, instructing it to initiate a transaction compensation mechanism. The rollback process refers to canceling partially executed task operations and restoring the system state to the previous stable point to avoid making incorrect decisions based on low-confidence information. The rollback process works closely with the transaction compensation mechanism in the execution layer protocol to ensure the accuracy and completeness of the state rollback. Of course, this is not limited to this; any mechanism capable of dynamically monitoring the confidence scores of each slot information inherited from the intent parsing module can be used.

[0047] Specifically, step S3. The execution layer protocol in parallel scheduling includes the following steps: Idempotent actions are achieved by introducing a unique request identifier in each atomic service interface. On the service side, before each request is processed, a distributed cache system or database is queried to check whether the request identifier has been processed, so as to avoid duplicate execution of business logic. The transaction compensation strategy is implemented based on the Saga pattern. Each business process is decomposed into a series of small, independent local transactions. Each local transaction publishes an event upon completion. If a local transaction fails or needs to be rolled back, the local transaction publishes a "compensation event" that triggers the compensation operation of the previously successfully executed local transactions.

[0048] In this embodiment, idempotency is achieved by introducing a unique request identifier in each atomic service interface. On the service side, before each request is processed, a distributed cache system or database is queried to check whether the request identifier has already been processed. If it has been processed, the result of the previous processing is returned directly without repeating the business logic, thus avoiding the side effects of repeated operations. The generation of the request identifier uses globally unique identifiers or snowflake algorithms to ensure its uniqueness worldwide. The transaction compensation strategy is a fault-tolerant mechanism designed to address potential execution failures in complex task flows. When an atomic task in the task execution chain fails, the system automatically triggers a predefined compensation operation to roll back previously successfully executed related operations, restoring the system state to its state before the transaction began or a consistent intermediate state. The compensation operation is typically the inverse operation of the original operation, and the transaction compensation strategy is implemented based on the Saga pattern. Each business process is decomposed into a series of small, independent local transactions, each of which publishes an event upon completion. If a local transaction fails, or a rollback is detected, a "compensation event" is published, triggering the compensation operation of previously successfully executed local transactions. The compensation operation is designed to be atomic and also idempotent. Of course, this is not a limitation; other specifications and behaviors that define task execution and ensure system stability and data consistency can also be used.

[0049] The technical features of the above embodiments can be combined in any way. 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.

[0050] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. An intelligent task orchestration system based on intent recognition, characterized in that: It includes an intent parsing module, a task decomposition module, a parallel scheduling module, and a result aggregation module. The graph parsing module is used to receive user input in natural language form and perform natural language understanding processing, entity and slot extraction, and confidence calculation. The task decomposition module is used to receive the intent tags and structured slot information output by the graph parsing module, and perform hierarchical task network planning, task directed acyclic graph generation, and constraint propagation processing based on the received intent tags and structured slot information. The parallel scheduling module is used to receive the directed acyclic graph of tasks and the constraint information output by the task decomposition module, and to execute critical path priority scheduling, confidence gating mechanism, resource upper limit control and execution layer protocol implementation. The result aggregation module is used to receive the execution results of the parallel scheduling module and perform multi-branch result collection, consistency verification, and execution report generation.

2. A task orchestration method based on the system of claim 1, characterized in that: Includes the following steps: S1. Intent parsing: Receives user input in natural language form, performs natural language understanding processing using a BERT-like pre-trained language model, extracts entities and slots using a sequence labeling model, and calculates confidence scores for the identified intent and extracted slot information. S2. Task decomposition: Based on the intent tags and structured slot information output by the intent parsing module, the complex task is recursively decomposed into atomic tasks using hierarchical task network planning, and a directed acyclic graph of the task is constructed to clarify the execution order and dependencies between tasks, while propagating and processing various constraints. S3. Parallel Scheduling: Based on the directed acyclic graph of the tasks and the constraint information, the critical path priority strategy is adopted to schedule task execution. At the same time, the confidence of slot position is dynamically monitored through a confidence gating mechanism. When the confidence is lower than a preset threshold, a user confirmation or system rollback process is triggered, and the system resource upper limit is controlled. The execution layer protocol is completed through idempotent actions and transaction compensation strategies. S4. Result Aggregation: Collect the results of all parallel execution branches, standardize the data format and verify logical consistency, and generate a detailed task execution report, including execution trajectory, key decisions and any exception handling records.

3. The intelligent task orchestration method based on intent recognition according to claim 2, characterized in that: Step S1, intent parsing, which utilizes a BERT-like pre-trained language model for natural language understanding, includes the following steps: This method achieves bidirectional contextual encoding of text by using a self-attention mechanism. Its internal architecture consists of multiple stacked Transformer encoders, each composed of a multi-head self-attention mechanism and a feedforward neural network. Residual connections and layer normalization techniques are used to improve training stability. The masked language model and next-sentence prediction task are pre-trained on a large text corpus, and fine-tuned on a labeled dataset containing domain-specific terms and expressions when applied to specific domains. Task-specific prediction is achieved by introducing one or more linear classification layers.

4. The intelligent task orchestration method based on intent recognition according to claim 3, characterized in that: Step S1. Entity and slot extraction in intent parsing includes the following steps: A multi-task learning framework is adopted, in which task-specific prediction includes an intent classification task and a slot filling task. The intent classification task and the slot filling task share the parameters of the underlying BERT-like pre-trained language model and perform prediction through their respective independent output layers. The slot filling task specifically adopts the Span-BERT architecture, which can directly predict continuous spans in the text as entities or slot values, effectively handling overlapping and nested entities. For implicit slots, the semantic reasoning module infers and fills slots that are not explicitly expressed but are logically necessary based on explicit information provided by the user and the domain knowledge graph.

5. The intelligent task orchestration method based on intent recognition according to claim 4, characterized in that: Step S1. Intent parsing involves calculating the confidence score for the identified intent and the extracted slot information, which includes the following steps: The identified intent and extracted slot information are subjected to posterior probability and Monte Carlo Dropout-based prediction uncertainty estimation. For each slot, the softmax probability value of the slot prediction label is calculated as an initial confidence level. By running the model with the Monte Carlo Dropout layer multiple times during the inference phase, multiple prediction results are obtained. Then, the entropy or standard deviation of the prediction results is calculated to quantify the model's "uncertainty" regarding the prediction. The final confidence score is a weighted combination of the softmax probability and the uncertainty measure.

6. The intelligent task orchestration method based on intent recognition according to any one of claims 2-5, characterized in that: The steps of hierarchical task network planning in step S2, task decomposition, include: Based on the intent tag and slot information obtained from the intent parsing in S1, the high-level task is decomposed by the task decomposition module and stored in the domain knowledge base in a formal language. The task decomposition module defines the specific steps for decomposing a high-level task into a set of subtasks or atomic operations, and includes preconditions and effects.

7. The intelligent task orchestration method based on intent recognition according to claim 6, characterized in that: The construction of the directed acyclic graph of the task in step S2, task decomposition, includes the following steps: The task is managed and persisted using a graph database. The directed acyclic graph (DAG) includes node attributes and edge attributes. The node attributes of the task DAG include task ID, task type, estimated execution time, input parameters, output parameters, and current state. The edge attributes of the task DAG include dependency type, source task ID, and target task IP. The subtasks or atomic operations output by the hierarchical task network planning are converted into an initial set of nodes, and then directed edges are gradually added through the hierarchical task network planning.

8. The intelligent task orchestration system and method based on intent recognition according to claim 7, characterized in that: Step S2, propagating and processing various constraints in task decomposition, includes the following steps: For sequence constraints, virtual synchronization nodes are added to the directed acyclic graph of tasks or task priorities are adjusted to ensure that predecessor tasks complete before successor tasks. For mutual exclusion constraints, mutually exclusive tasks are prevented from executing simultaneously by allocating them to non-overlapping time windows or by using resource locking mechanisms. For resource limit constraints, when constructing the directed acyclic graph of tasks, the task decomposition module performs preliminary resource reservation or capacity checks on tasks with high resource requirements and optimizes resource allocation through a capacity-based network flow algorithm to ensure that concurrent resource usage at any time does not exceed the system-set threshold.

9. The intelligent task orchestration method based on intent recognition according to claim 7 or 8, characterized in that: The working steps of the confidence gating mechanism in step S3, parallel scheduling, include: The system continuously monitors the task execution events and slot position confidence changes in the parallel scheduling module; the preset dynamic threshold is trained and predicted by a machine learning model based on the task's criticality, business scenario, and historical performance data; triggering the confidence gating operation includes sending a confirmation request to the user or initiating a system rollback process. The confirmation request includes providing the user with key information currently understood by the system and possible alternative options, and collecting user feedback to update the slot information and slot position confidence.

10. The intelligent task orchestration method based on intent recognition according to claim 9, characterized in that: The execution layer protocol in step S3, parallel scheduling, includes the following steps: The idempotent action is achieved by introducing a unique request identifier in each atomic service interface. On the service side, before each request is processed, a distributed cache system or database is queried to check whether the request identifier has been processed, so as to avoid repeated execution of business logic. The transaction compensation strategy is implemented based on the Saga pattern. Each business process is decomposed into a series of small, independent local transactions. Each local transaction publishes an event upon completion. If a local transaction fails or needs to be rolled back, the local transaction publishes a "compensation event" to trigger the compensation operation of the previously successfully executed local transactions.

Images