Intelligence analysis method and device based on large language model hierarchical task decomposition

Abstract

This invention provides an intelligence analysis method and apparatus based on hierarchical task decomposition using a large language model, relating to the field of artificial intelligence technology. The method includes: determining an intelligence analysis task based on the intelligence data to be analyzed; inputting the natural language description corresponding to the intelligence analysis task into a large intelligence analysis model to obtain the current hierarchical workflow output by the large intelligence analysis model; generating a task execution graph using subtasks as graph nodes and encoding preceding dependent tasks as directed edges; and executing the intelligence analysis task based on the task execution graph to obtain the intelligence analysis results. The intelligence analysis method and apparatus based on hierarchical task decomposition using a large language model provided by this invention can accurately and orderly execute intelligence analysis tasks according to a planned task execution graph without relying on a predefined rule base, reducing the cost of intelligence analysis using a large language model and improving the accuracy and real-time performance of intelligence analysis.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to an intelligence analysis method and apparatus based on hierarchical task decomposition of large language models. Background Technology

[0002] By combining the general capabilities of large language models (such as the Tongyi Qianwen Large Model Series, MiniMax Large Model Series, or Zhipu GLM Large Model Series) with the highly specialized, complex, and time-sensitive nature of intelligence analysis, automated and intelligent intelligence analysis can be achieved.

[0003] In existing technologies, the application of large models to intelligence analysis has the following drawbacks: (1) Methods based on predefined rules (such as hierarchical task networks) are difficult to adapt to the real-world needs of changing intelligence topics and flexible adjustment of analysis paradigms, and the construction and maintenance of rule bases are costly; (2) Intelligence analysis methods based on proxy frameworks such as ReAct usually rely on iterative reasoning and external interaction for task decomposition. Due to its serial execution mode, each round of "reasoning-action-observation" loop requires multiple calls to the large model, resulting in a high cumulative delay from task reception to plan generation, which is difficult to meet the high real-time requirements of intelligence analysis; (3) General large models lack in-depth knowledge in fields such as military intelligence and geopolitics, and the accuracy and reliability of output in professional sub-tasks such as sign recognition, situation assessment, and risk prediction are insufficient. Summary of the Invention

[0004] This invention provides an intelligence analysis method and apparatus based on hierarchical task decomposition of a large language model, which solves the technical problems of high cost and untimely and inaccurate output results in the prior art when using large language models for intelligence analysis.

[0005] This invention provides an intelligence analysis method based on hierarchical task decomposition using a large language model, comprising the following steps: Based on the intelligence data to be analyzed, determine the intelligence analysis task; The natural language description corresponding to the intelligence analysis task is input into the large intelligence analysis model to obtain the current hierarchical workflow output by the large intelligence analysis model; wherein, the current hierarchical workflow includes multiple subtasks and multiple pre-dependent task codes; each subtask corresponds to one or more of the pre-dependent task codes; the pre-dependent task codes are determined based on the dependency relationships between the subtasks; the large intelligence analysis model is obtained by training a first large language model based on a pre-constructed training dataset; A task execution graph is generated using the subtasks as graph nodes and the preceding dependent tasks as directed edges. Based on the task execution graph, the intelligence analysis task is executed to obtain intelligence analysis results.

[0006] According to the intelligence analysis method based on hierarchical task decomposition of a large language model provided by the present invention, the sub-tasks further include one or more meta-tasks; before generating the task execution graph, the method further includes a verification step of the current hierarchical workflow; the verification step includes at least one of the following: Based on the aforementioned pre-dependent task encoding, it is determined that there is no closed-loop dependency relationship between the subtasks; Determine that the dependency relationship between the subtask and the metatask conforms to preset rules; An expert model is identified that satisfies the requirements corresponding to the meta-task; the expert model is used to execute the corresponding meta-task; the expert model is a pre-trained convolutional neural network model or a model constructed based on a rule-based algorithm according to different requirements.

[0007] According to the present invention, an intelligence analysis method based on hierarchical task decomposition of a large language model is provided, wherein the intelligence analysis task is executed based on the task execution graph to obtain intelligence analysis results, including: Based on the task execution graph, the expert model is invoked to execute the corresponding meta-task, and the execution result output by the expert model is obtained; The execution results are integrated according to preset rules to generate an intelligence analysis report; The intelligence analysis report is used as the result of the intelligence analysis.

[0008] According to the present invention, an intelligence analysis method based on hierarchical task decomposition of a large language model is provided, wherein the step of integrating the execution results according to preset rules to generate an intelligence analysis report includes: Based on the user's input instructions, the execution result is corrected to obtain a corrected result; Based on the correction results, the intelligence analysis report is generated.

[0009] According to the intelligence analysis method based on hierarchical task decomposition of a large language model provided by the present invention, the steps for constructing the training dataset include: Obtain seed instructions; the seed instructions include historical task descriptions, corresponding hierarchical historical workflows, and historical reasoning processes. The seed instruction and the first prompt word are input into the second large language model to obtain the first dataset output by the second large language model; the first prompt word is used to instruct the second large language model to perform instruction bootstrapping based on the seed instruction. The first dataset and the second prompt word are input into the second large language model to obtain the second dataset output by the second large language model; the second prompt word is used to instruct the second large language model to perform instruction evolution based on the first dataset. The second dataset is used as the training dataset.

[0010] According to the intelligence analysis method based on hierarchical task decomposition of a large language model provided by the present invention, the second prompt word includes at least one of the following indicative information: Add constraints to the description of the historical tasks; Introduce interfering information into the description of the historical tasks; The reasoning steps in the historical reasoning process are extended.

[0011] This invention also provides an intelligence analysis device based on hierarchical task decomposition of a large language model, comprising the following modules: The determination module is used to determine the intelligence analysis task based on the intelligence data to be analyzed; The task decomposition module is used to input the natural language description corresponding to the intelligence analysis task into the large intelligence analysis model to obtain the current hierarchical workflow output by the large intelligence analysis model; wherein, the current hierarchical workflow includes multiple sub-tasks and multiple pre-dependent task codes; each sub-task corresponds to one or more of the pre-dependent task codes; the pre-dependent task codes are determined based on the dependency relationships between the sub-tasks; the large intelligence analysis model is obtained by training a first large language model based on a pre-constructed training dataset; The generation module is used to generate a task execution graph with the subtasks as graph nodes and the preceding dependent task codes as directed edges. The task execution module is used to execute the intelligence analysis task based on the task execution graph and obtain intelligence analysis results.

[0012] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the intelligence analysis method based on hierarchical task decomposition of a large language model as described above.

[0013] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the intelligence analysis method based on hierarchical task decomposition of a large language model as described above.

[0014] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the intelligence analysis method based on hierarchical task decomposition of a large language model as described above.

[0015] This invention provides an intelligence analysis method and apparatus based on hierarchical task decomposition using a large language model. Based on the intelligence data to be analyzed, an intelligence analysis task is determined. The natural language description corresponding to the intelligence analysis task is input into a large intelligence analysis model to obtain the current hierarchical workflow output by the large intelligence analysis model. The current hierarchical workflow includes multiple sub-tasks and multiple pre-dependent task codes. Each sub-task corresponds to one or more pre-dependent task codes. The pre-dependent task codes are determined based on the dependencies between sub-tasks. The large intelligence analysis model is trained on a pre-constructed training dataset using a first large language model, thereby transforming the intelligence data into a natural language description that the large model can understand, realizing the transformation from natural language description to the current hierarchical workflow. The end-to-end one-time generation of workflows improves task decomposition efficiency, shortens the latency of multi-round inference in traditional large models, and enhances the real-time performance of intelligence analysis. Using subtasks as graph nodes and encoding preceding dependent tasks as directed edges, a task execution graph is generated, transforming the current hierarchical workflow in text form into a task execution graph with clear dependencies. This provides a standardized data structure that can be directly parsed and executed by machines for the parallel and serial scheduling and dependency checks of subsequent tasks. Based on the task execution graph, intelligence analysis tasks are executed to obtain intelligence analysis results. Thus, without the need for a predefined rule base, intelligence analysis tasks can be executed accurately and orderly according to the planned task execution graph, reducing the cost of intelligence analysis using large language models and improving the accuracy of intelligence analysis results. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in this invention 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 some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0017] Figure 1 This is a flowchart illustrating the intelligence analysis method based on hierarchical task decomposition of a large language model provided by the present invention.

[0018] Figure 2 This is a schematic diagram of a directed acyclic graph provided by the present invention.

[0019] Figure 3 This is a schematic diagram of the intelligence analysis device based on hierarchical task decomposition of a large language model provided by the present invention.

[0020] Figure 4 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0022] The following is combined Figures 1 to 4 This invention describes an intelligence analysis method and apparatus based on hierarchical task decomposition using a large language model.

[0023] Figure 1 This is a flowchart illustrating the intelligence analysis method based on hierarchical task decomposition using a large language model provided by the present invention, as shown below. Figure 1 As shown, the method includes the following steps: Step 101: Based on the intelligence data to be analyzed, determine the intelligence analysis task; Specifically, when raw and vague intelligence data is obtained, the corresponding intelligence analysis task is determined according to actual needs.

[0024] For example, in one embodiment, intelligence data could be the future development plan published by organization A, and the intelligence analysis task could be the probability that organization A will take action B at some point in the future.

[0025] Step 102: Input the natural language description corresponding to the intelligence analysis task into the large intelligence analysis model to obtain the current hierarchical workflow output by the large intelligence analysis model; wherein, the current hierarchical workflow includes multiple sub-tasks and multiple pre-dependent task codes; each sub-task corresponds to one or more of the pre-dependent task codes; the pre-dependent task codes are determined based on the dependency relationships between the sub-tasks; the large intelligence analysis model is obtained by training the first large language model based on a pre-constructed training dataset; Specifically, the intelligence analysis big model is obtained by training the first big language model on a pre-built training dataset. The first big language model can be any big model from series such as the Tongyi Qianwen Big Model Series, the MiniMax Big Model Series, or the Zhipu GLM Big Model Series.

[0026] Furthermore, the steps for constructing the training dataset include: Obtain seed instructions; the seed instructions include historical task descriptions, corresponding hierarchical historical workflows, and historical reasoning processes. The seed instruction and the first prompt word are input into the second large language model to obtain the first dataset output by the second large language model; the first prompt word is used to instruct the second large language model to perform instruction bootstrapping based on the seed instruction. The first dataset and the second prompt word are input into the second large language model to obtain the second dataset output by the second large language model; the second prompt word is used to instruct the second large language model to perform instruction evolution based on the first dataset. The second dataset is used as the training dataset.

[0027] Specifically, first, seed instructions are obtained, which include a historical task description, a corresponding hierarchical workflow, and a historical reasoning process. The historical reasoning process fully records the thought processes, logical judgments, and decision-making basis of human experts when solving relevant problems in the historical task description, serving as "meta-knowledge" for model learning. Then, the seed instructions and the first prompt word are input into a general second-large language model for instruction bootstrapping, outputting the first dataset. Next, the first dataset and a more complex second prompt word are input into the second-large language model for instruction evolution, producing a more challenging second dataset, which serves as the training dataset. The second-large language model can be any of the following series: Tongyi Qianwen Large Model Series, MiniMax Large Model Series, or Zhipu GLM Large Model Series.

[0028] For example, in one embodiment, a seed instruction and a first prompt word (such as "Please generate multiple new instructions with similar semantics but different themes") are input into the large model GLM-4. The large model GLM-4 performs instruction bootstrapping according to the instructions of the first prompt word, that is, it analyzes and learns the task logic and structure implicit in the "seed instruction", and then imitates this logic and structure to automatically generate a large number of new, logically similar task instructions (i.e., the first dataset), but the specific scenarios, objects, and details (i.e., "semantics") are different.

[0029] It is understandable that instruction bootstrapping can easily lead to highly repetitive (homogeneous) data. To address this issue, this embodiment of the invention also introduces diversity control based on vector similarity: all seed instructions in the first dataset are converted into corresponding mathematical vectors (embedded representations), and then the similarity between vectors is calculated. This process removes seed instructions corresponding to vectors with similarity higher than a preset threshold, retaining only samples with sufficient novelty, thereby ensuring the diversity of data in the first dataset.

[0030] Building upon instruction bootstrapping, the large model is further guided by a second prompt word to proactively "increase the difficulty" of the task (i.e., instruction evolution) when generating new seed instructions.

[0031] Furthermore, the second prompt word includes at least one of the following indicative information: Add constraints to the description of the historical tasks; Introduce interfering information into the description of the historical tasks; The reasoning steps in the historical reasoning process are extended.

[0032] Specifically, the second cue word is used to instruct the second large language model to evolve instructions based on the first dataset, and may include at least one of the following: Add constraints to the description of historical tasks: Add specific restrictions to the description of historical tasks, such as specific analytical perspectives, word limits, content formats, etc. Introduce interfering information into the historical task description: Add noisy data that needs to be identified and filtered by the large model to the historical task description, such as irrelevant events and speeches made by relevant people; Extend the reasoning steps in the historical reasoning process: extend single-step tasks into multi-step complex reasoning tasks that require "investigation first, identification second, and deduction third"; Introduce diverse scenarios into the historical reasoning process: keep the logic in the seed instruction unchanged, but modify the reasoning timeline, environmental information, reasoning goal, etc.

[0033] For example, in one embodiment, a first dataset and a second cue word are input into the large model GLM-4. The large model GLM-4 evolves according to the instructions of the second cue word. The second cue word contains explicit instructions: such as evolving "analyze the probability of taking action B" into "analyze the probability of taking action B within 24 hours based solely on open-source intelligence analysis"; expanding the original two-step reasoning of "identifying signs - judging" into a five-step reasoning of "identifying signs - verifying sources - associating with historical events - judging intent - assessing probability". Through these strategies, a tiered training dataset can be systematically constructed to obtain the second dataset output by the large model GLM-4.

[0034] The embodiments of the present invention can systematically generate a series of training datasets with increasing complexity from low to high through instruction bootstrapping and instruction evolution.

[0035] Supervised fine-tuning of the first large language model based on the training dataset enables the large model to gradually progress from learning to solve simple problems to handling complex and challenging tasks, achieving progressive training of capabilities and obtaining a large intelligence analysis model.

[0036] The natural language description corresponding to the intelligence analysis task is input into the intelligence analysis big model to obtain the current hierarchical workflow output by the intelligence analysis big model. The current hierarchical workflow includes multiple sub-tasks and multiple pre-dependent task codes. Each sub-task corresponds to one or more pre-dependent task codes. The pre-dependent task codes are determined based on the dependency relationship between sub-tasks, thus clarifying the task level, dependency relationship and execution resources through the current hierarchical workflow.

[0037] For example, in one embodiment, the natural language description corresponding to the intelligence analysis task is input into the large intelligence analysis model (e.g., a model obtained by supervised fine-tuning using LoRA technology and the aforementioned training dataset based on the Qwen2-7B-Instruct base model). The large intelligence analysis model outputs a structured workflow in JSON format at once, which includes a series of sub-tasks such as t1, multi-source intelligence data collection and preprocessing; t2, anomaly identification and situation assessment; t3, B action trend prediction and risk level; t4, customized intelligence product output; t5, early warning triggering mechanism and response guidelines; t6, intelligence effect evaluation and system optimization, as well as explicit prerequisite dependencies, such as t2 depending on t1.

[0038] This invention overcomes the high latency problem of multiple iterations in the traditional ReAct framework by analyzing and outputting a clear current hierarchical workflow in one go through a large intelligence analysis model (experiments have shown that the latency can be reduced from an average of 3 seconds to 800 milliseconds). A globally optimal execution plan containing explicit dependencies can be generated with a single model call, thereby improving the efficiency and real-time performance of intelligence analysis.

[0039] Step 103: Using the subtasks as graph nodes and the preceding dependent task codes as directed edges, generate a task execution graph; Specifically, the textualized, structured hierarchical workflow obtained in the previous step is transformed into a directed acyclic graph (i.e., a task execution graph) with strict topological logic that can be directly scheduled and executed by a computer program. Each subtask becomes a node in the graph, and dependencies become directed edges between nodes.

[0040] Furthermore, the subtask further includes one or more metatasks; before generating the task execution graph, the method further includes a verification step of the current hierarchical workflow; the verification step includes at least one of the following: Based on the aforementioned pre-dependent task encoding, it is determined that there is no closed-loop dependency relationship between the subtasks; Determine that the dependency relationship between the subtask and the metatask conforms to preset rules; An expert model is identified that satisfies the requirements corresponding to the meta-task; the expert model is used to execute the corresponding meta-task; the expert model is a pre-trained convolutional neural network model or a model constructed based on a rule-based algorithm according to different requirements.

[0041] Specifically, the verification step is used to check the completeness and correctness of the current hierarchical workflow. First, it checks whether all required fields exist, such as the task name, subtask list, and metatask name; the absence of any required field will lead to subsequent execution failure. Second, it checks whether the nesting relationship of the structure is correct, that is, whether the dependency relationship between subtasks and metatasks conforms to preset rules, such as whether the metatask is correctly contained within a subtask. Third, it determines whether an expert model exists that meets the requirements of the metatask, such as checking whether the name of the specified expert model exists in the registered model interfaces. If the name of the specified expert model in the current hierarchical workflow cannot find a corresponding model interface, it means that there is no pre-trained convolutional neural network model or rule-based algorithm model that meets the requirements of the metatask, and the metatask will not be executed. Finally, if an expert model that meets the requirements of the metatask exists, it can further check whether the type and format of the input parameters received by the expert model meet the requirements, such as whether numerical parameters are indeed numbers, and whether time parameters conform to the time format.

[0042] Among these, the expert model can be a pre-trained convolutional neural network model tailored to different needs, or it can be a script specifically designed for data processing tasks. For example, if the meta-task is to generate a structured JSON dataset, the corresponding expert model can be a Python script that encapsulates specific data crawling, cleaning, transformation, and standardization logic. It can receive raw data source information as input and output a JSON dataset with a uniform format and clear structure. If the meta-task is to identify abnormal signs, it is necessary to process satellite imagery or drone aerial photos to detect and locate the aggregation of specific types of ship targets (such as aircraft carriers, destroyers, etc.), aircraft, or ground equipment in a certain area in real time. The corresponding expert model can be a pre-trained ResNet series model (such as ResNet-50, ResNet-101). It can perform target detection on the input target image, directly map image pixels to bounding boxes and class probabilities, and output specific classification labels and confidence scores.

[0043] In addition to the above verification steps, it is also necessary to determine, based on the encoding of the preceding dependency tasks, whether there are no closed-loop dependencies between subtasks, that is, to verify whether the dependencies between subtasks form a directed acyclic graph.

[0044] Dependency verification employs a topological sorting algorithm. First, a directed graph is constructed based on the encoded prerequisite tasks, where nodes represent subtasks. If subtask A depends on subtask B, there is a directed edge from B to A. Then, a topological sort is attempted on this graph. Topological sorting arranges all nodes of the directed acyclic graph into a linear sequence, ensuring that for each directed edge, the starting point precedes the ending point. If the topological sort is successful, the dependencies are valid, there are no circular dependencies, and the execution order can be determined based on the sorting result. If the topological sort fails, it indicates the existence of circular dependencies, such as A depending on B, B depending on C, and C depending on A. In this case, a reasonable execution order cannot be determined, and the current hierarchical workflow needs to be modified or regenerated.

[0045] Figure 2 This is a schematic diagram of a directed acyclic graph provided by the present invention, such as... Figure 2 As shown, the subtasks include data processing, information transformation, product generation, and closed-loop optimization. The metatasks include t1, multi-source intelligence data collection and preprocessing; t2, anomaly identification and situation assessment; t3, B-action trend prediction and risk level; t4, customized intelligence product output; t5, early warning triggering mechanism and response guidelines; and t6, intelligence effectiveness evaluation and system optimization. Taking the metatask as the smallest execution unit, its execution order is: t1-[t2,t3]-[t4,t5]-t6, where t2 and t3, and t4 and t5 can be executed in parallel to improve task execution efficiency.

[0046] This invention significantly improves system stability and output reliability by detecting and intercepting potential logical or formatting errors in large intelligence analysis models in advance, thus avoiding resource waste caused by erroneous workflows flowing into the execution phase.

[0047] By constructing a task execution graph, the granularity and coordination mechanism of each sub-task are clarified. By scheduling appropriate expert models to handle the corresponding meta-tasks, the collaboration between the general large model and the expert model is realized, maximizing the efficiency and professionalism of the overall task execution and improving the accuracy of intelligence analysis.

[0048] Step 104: Based on the task execution graph, execute the intelligence analysis task and obtain the intelligence analysis results.

[0049] Specifically, based on the topological order of the task execution graph, the most professional execution tools or expert models associated with each node (subtask) in the graph are called sequentially or in parallel to execute the corresponding meta-tasks, and the execution results of all meta-tasks are integrated to form the final output and obtain intelligence analysis results.

[0050] This invention provides an intelligence analysis method based on hierarchical task decomposition using a large language model. Based on the intelligence data to be analyzed, the intelligence analysis task is determined. The natural language description corresponding to the intelligence analysis task is input into a large intelligence analysis model to obtain the current hierarchical workflow output by the large intelligence analysis model. The current hierarchical workflow includes multiple sub-tasks and multiple pre-dependent task codes. Each sub-task corresponds to one or more pre-dependent task codes. The pre-dependent task codes are determined based on the dependencies between sub-tasks. The large intelligence analysis model is trained on a pre-constructed training dataset using a first large language model, thereby transforming the intelligence data into a natural language description that the large model can understand, realizing the transformation from natural language description to the current hierarchical workflow. The end-to-end one-time generation improves task decomposition efficiency, shortens the latency of multi-round inference in traditional large models, and enhances the real-time performance of intelligence analysis. Using subtasks as graph nodes and encoding preceding dependent tasks as directed edges, a task execution graph is generated, transforming the current hierarchical workflow in text form into a task execution graph with clear dependencies. This provides a standardized data structure that can be directly parsed and executed by machines for the parallel and serial scheduling and dependency checks of subsequent tasks. Based on the task execution graph, intelligence analysis tasks are executed to obtain intelligence analysis results. Thus, without the need for a predefined rule base, intelligence analysis tasks can be executed accurately and orderly according to the planned task execution graph, reducing the cost of intelligence analysis using large language models and improving the accuracy of intelligence analysis results.

[0051] Furthermore, the step of executing the intelligence analysis task based on the task execution graph to obtain intelligence analysis results includes: Based on the task execution graph, the expert model is invoked to execute the corresponding meta-task, and the execution result output by the expert model is obtained; The execution results are integrated according to preset rules to generate an intelligence analysis report; The intelligence analysis report is used as the result of the intelligence analysis.

[0052] Specifically, based on the task execution graph, a task execution workflow is formed. Subtasks are executed according to dependencies and parallel relationships. Each meta-task invokes a dedicated model or tool during execution, thus achieving multi-task and multi-agent collaboration through a large-model intelligent agent approach. Real-time information such as task execution status, task execution details, and intelligent agent output can be monitored during task execution. Simultaneously, tasks in progress can be stopped, intelligent agent configurations adjusted, and then re-executed. The meta-task is the smallest execution unit; by invoking different expert models to execute different meta-tasks, multiple execution results are obtained. Finally, all execution results are integrated according to preset rules to generate an intelligence analysis report.

[0053] For example, in one embodiment, a dedicated text compilation model (e.g., a trained text generation model or a report template filling engine) is invoked to perform the meta-task, according to the definition in the task execution graph. Assume this model receives risk level data output from the previous node's "information transformation" stage as input and generates a structured warning text. Then, following preset rules (e.g., in the order of "data-analysis-warning"), the outputs of all expert models (data tables, situation maps, warning texts, etc.) are integrated and compiled into a complete "Intelligence Analysis Report" as the intelligence analysis result.

[0054] Furthermore, the step of integrating the execution results according to preset rules to generate an intelligence analysis report includes: Based on the user's input instructions, the execution result is corrected to obtain a corrected result; Based on the correction results, the intelligence analysis report is generated.

[0055] Specifically, building upon automated intelligence analysis processes, human intervention is introduced at key decision points to ensure reliable and controllable task execution. For example, in solution design tasks, multiple solutions can be generated. By setting key decision points, multiple solutions can be output for the user to choose from, and a report is generated based on the selected solution. Furthermore, by adding key decision point markers to the corresponding key decision node tasks in the training data, the large model can learn to predict decision nodes requiring human intervention through training. For instance, before generating the report, a "review outline" decision node can be automatically inserted, allowing for adjustments to the execution result based on user input at key decision points. These adjusted results are then integrated to generate the intelligence analysis report.

[0056] For example, in one embodiment, when the workflow reaches the "early warning triggering mechanism and response guidelines" (such as...), Figure 2 When a critical decision point (such as a task breakpoint) is reached, the system automatically pauses the execution of subsequent meta-tasks and generates two different response plans, A and B, to present to the user. The system then submits an application to the user until one plan is approved. Based on the user's approval, the execution of subsequent meta-tasks resumes. For example, if a user (such as an analyst) evaluates the plan through the human-computer interface, selects plan A, and adds a correction instruction: "Add 'Take additional action C' as a priority option to plan A," the system receives this user input instruction, uses it as context, and continues executing the subsequent "Intelligence Effectiveness Evaluation and System Optimization" task. The final intelligence analysis report will reflect the user's selection and correction.

[0057] This invention, by embedding human expertise and decision-making capabilities into automated processes during the integration of execution results, effectively prevents the accumulation of possible errors in long-chain tasks, enhances the controllability, reliability, and flexibility of the system when handling complex and high-risk tasks, and improves the accuracy of intelligence analysis results.

[0058] Based on any of the above embodiments, support for multi-turn dialogue capabilities can be extended, enabling users to gradually complete complex intelligence analysis tasks through continuous human-computer interaction. For example, maintaining dialogue history includes all historical questions raised by the user in the current session, workflows generated by the large model for these questions, results obtained from executing the workflows, and user feedback on the results, such as satisfaction ratings and modification suggestions. By storing dialogue history data in a database or file system, relevant historical information can be loaded each time a user asks a question.

[0059] To enable large-scale models to learn from real-world usage and continuously improve performance, online learning mechanisms can be established. For example, feedback can be collected from user behavior, including explicit and implicit feedback. Explicit feedback includes proactive evaluations of workflows and results, such as ratings, comments, and suggestions for improvement; this directly reflects user satisfaction and feedback. Implicit feedback includes user behavioral data, such as whether intelligence analysis tasks were successfully completed, whether users resubmitted their questions, and how users utilized the intelligence analysis results; although not explicitly stated, these behaviors reflect the performance of the large-scale model. Human-computer interaction interface design can guide users to provide feedback, such as asking for satisfaction upon receiving results, providing rating options and comment boxes, thus lowering the barrier to feedback.

[0060] The intelligence analysis device based on hierarchical task decomposition of a large language model provided by the present invention will be described below. The intelligence analysis device based on hierarchical task decomposition of a large language model described below can be referred to in correspondence with the intelligence analysis method based on hierarchical task decomposition of a large language model described above.

[0061] Figure 3 This is a schematic diagram of the intelligence analysis device based on hierarchical task decomposition of a large language model provided by the present invention, as shown below. Figure 3 As shown. This invention provides an intelligence analysis device based on hierarchical task decomposition using a large language model, comprising a determination module 301, a task decomposition module 302, a generation module 303, and a task execution module 304, wherein: The determination module 301 is used to determine the intelligence analysis task based on the intelligence data to be analyzed; the task decomposition module 302 is used to input the natural language description corresponding to the intelligence analysis task into the large intelligence analysis model to obtain the current hierarchical workflow output by the large intelligence analysis model; wherein, the current hierarchical workflow includes multiple sub-tasks and multiple pre-dependent task codes; each sub-task corresponds to one or more pre-dependent task codes; the pre-dependent task codes are determined based on the dependency relationships between the sub-tasks; the large intelligence analysis model is obtained by training a first large language model based on a pre-constructed training dataset; the generation module 303 is used to generate a task execution graph with the sub-tasks as graph nodes and the pre-dependent task codes as directed edges; the task execution module 304 is used to execute the intelligence analysis task based on the task execution graph to obtain the intelligence analysis result.

[0062] This invention provides an intelligence analysis device based on hierarchical task decomposition using a large language model. Based on the intelligence data to be analyzed, the device determines the intelligence analysis task; it inputs the natural language description corresponding to the intelligence analysis task into a large intelligence analysis model to obtain the current hierarchical workflow output by the model. The current hierarchical workflow includes multiple sub-tasks and multiple pre-dependent task codes; each sub-task corresponds to one or more pre-dependent task codes; the pre-dependent task codes are determined based on the dependencies between sub-tasks; the large intelligence analysis model is trained on a pre-constructed training dataset using a first large language model, thereby transforming the intelligence data into a natural language description that the large model can understand, realizing the transformation from natural language description to the current hierarchical workflow. The end-to-end one-time generation improves task decomposition efficiency, shortens the latency of multi-round inference in traditional large models, and enhances the real-time performance of intelligence analysis. Using subtasks as graph nodes and encoding preceding dependent tasks as directed edges, a task execution graph is generated, transforming the current hierarchical workflow in text form into a task execution graph with clear dependencies. This provides a standardized data structure that can be directly parsed and executed by machines for the parallel and serial scheduling and dependency checks of subsequent tasks. Based on the task execution graph, intelligence analysis tasks are executed to obtain intelligence analysis results. Thus, without the need for a predefined rule base, intelligence analysis tasks can be executed accurately and orderly according to the planned task execution graph, reducing the cost of intelligence analysis using large language models and improving the accuracy of intelligence analysis results.

[0063] Figure 4 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 4As shown, the electronic device may include: a processor 410, a communications interface 420, a memory 430, and a communication bus 440, wherein the processor 410, the communications interface 420, and the memory 430 communicate with each other via the communication bus 440. The processor 410 can call logical instructions in the memory 430 to execute an intelligence analysis method based on hierarchical task decomposition of a large language model. This method includes: Based on the intelligence data to be analyzed, determine the intelligence analysis task; The natural language description corresponding to the intelligence analysis task is input into the large intelligence analysis model to obtain the current hierarchical workflow output by the large intelligence analysis model; wherein, the current hierarchical workflow includes multiple subtasks and multiple pre-dependent task codes; each subtask corresponds to one or more of the pre-dependent task codes; the pre-dependent task codes are determined based on the dependency relationships between the subtasks; the large intelligence analysis model is obtained by training a first large language model based on a pre-constructed training dataset; A task execution graph is generated using the subtasks as graph nodes and the preceding dependent tasks as directed edges. Based on the task execution graph, the intelligence analysis task is executed to obtain intelligence analysis results.

[0064] Furthermore, the logical instructions in the aforementioned memory 430 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0065] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the intelligence analysis method based on hierarchical task decomposition of a large language model provided by the above methods, the method including: Based on the intelligence data to be analyzed, determine the intelligence analysis task; The natural language description corresponding to the intelligence analysis task is input into the large intelligence analysis model to obtain the current hierarchical workflow output by the large intelligence analysis model; wherein, the current hierarchical workflow includes multiple subtasks and multiple pre-dependent task codes; each subtask corresponds to one or more of the pre-dependent task codes; the pre-dependent task codes are determined based on the dependency relationships between the subtasks; the large intelligence analysis model is obtained by training a first large language model based on a pre-constructed training dataset; A task execution graph is generated using the subtasks as graph nodes and the preceding dependent tasks as directed edges. Based on the task execution graph, the intelligence analysis task is executed to obtain intelligence analysis results.

[0066] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the intelligence analysis method based on hierarchical task decomposition of a large language model provided by the above methods, the method comprising: Based on the intelligence data to be analyzed, determine the intelligence analysis task; The natural language description corresponding to the intelligence analysis task is input into the large intelligence analysis model to obtain the current hierarchical workflow output by the large intelligence analysis model; wherein, the current hierarchical workflow includes multiple subtasks and multiple pre-dependent task codes; each subtask corresponds to one or more of the pre-dependent task codes; the pre-dependent task codes are determined based on the dependency relationships between the subtasks; the large intelligence analysis model is obtained by training a first large language model based on a pre-constructed training dataset; A task execution graph is generated using the subtasks as graph nodes and the preceding dependent tasks as directed edges. Based on the task execution graph, the intelligence analysis task is executed to obtain intelligence analysis results.

[0067] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0068] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0069] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element. Furthermore, it should be noted that the scope of the methods and apparatuses in the embodiments of this application is not limited to performing functions in the order shown or discussed, but may also include performing functions substantially simultaneously or in the reverse order, depending on the functions involved. For example, the described methods may be performed in a different order than described, and various steps may be added, omitted, or combined. Additionally, features described with reference to certain examples may be combined in other examples.

[0070] In this application's embodiments, "determine B based on A" means that factor A must be considered when determining B. It is not limited to "B can be determined based solely on A," but should also include: "determine B based on A and C," "determine B based on A, C, and E," "determine C based on A, and further determine B based on C," etc. Additionally, it can include using A as a condition for determining B, for example, "when A meets the first condition, determine B using the first method"; another example, "when A meets the second condition, determine B," etc.; another example, "when A meets the third condition, determine B based on the first parameter," etc. Of course, it can also be a condition where A is a factor in determining B, for example, "when A meets the first condition, determine C using the first method, and further determine B based on C," etc.

[0071] It should also be noted that the terms "target," "first," and "second" in this invention are used to distinguish similar objects, not to describe a specific order or sequence. It should be understood that such terms can be used interchangeably where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first" and "second" are generally of the same class, without limiting the number of objects; for example, the first object can be one or more.

[0072] In this invention, the term "multiple" refers to two or more kinds, and other quantifiers are similar.

[0073] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An intelligence analysis method based on hierarchical task decomposition using a large language model, characterized in that, include: Based on the intelligence data to be analyzed, determine the intelligence analysis task; The natural language description corresponding to the intelligence analysis task is input into the large intelligence analysis model to obtain the current hierarchical workflow output by the large intelligence analysis model; wherein, the current hierarchical workflow includes multiple subtasks and multiple pre-dependent task codes; each subtask corresponds to one or more of the pre-dependent task codes; the pre-dependent task codes are determined based on the dependency relationships between the subtasks; the large intelligence analysis model is obtained by training a first large language model based on a pre-constructed training dataset; A task execution graph is generated using the subtasks as graph nodes and the preceding dependent tasks as directed edges. Based on the task execution graph, the intelligence analysis task is executed to obtain intelligence analysis results.

2. The intelligence analysis method based on hierarchical task decomposition of a large language model according to claim 1, characterized in that, The subtask further includes one or more metatasks; prior to generating the task execution graph, the method further includes a verification step of the current hierarchical workflow; the verification step includes at least one of the following: Based on the aforementioned pre-dependent task encoding, it is determined that there is no closed-loop dependency relationship between the subtasks; Determine that the dependency relationship between the subtask and the metatask conforms to preset rules; An expert model is identified that satisfies the requirements corresponding to the meta-task; the expert model is used to execute the corresponding meta-task; the expert model is a pre-trained convolutional neural network model or a model constructed based on a rule-based algorithm according to different requirements.

3. The intelligence analysis method based on hierarchical task decomposition of a large language model according to claim 2, characterized in that, The process of executing the intelligence analysis task based on the task execution graph to obtain intelligence analysis results includes: Based on the task execution graph, the expert model is invoked to execute the corresponding meta-task, and the execution result output by the expert model is obtained; The execution results are integrated according to preset rules to generate an intelligence analysis report; The intelligence analysis report is used as the result of the intelligence analysis.

4. The intelligence analysis method based on hierarchical task decomposition of a large language model according to claim 3, characterized in that, The step of integrating the execution results according to preset rules to generate an intelligence analysis report includes: Based on the user's input instructions, the execution result is corrected to obtain a corrected result; Based on the correction results, the intelligence analysis report is generated.

5. The intelligence analysis method based on hierarchical task decomposition of a large language model according to claim 1, characterized in that, The steps for constructing the training dataset include: Obtain seed instructions; the seed instructions include historical task descriptions, corresponding hierarchical historical workflows, and historical reasoning processes. The seed instruction and the first prompt word are input into the second large language model to obtain the first dataset output by the second large language model; the first prompt word is used to instruct the second large language model to perform instruction bootstrapping based on the seed instruction. The first dataset and the second prompt word are input into the second large language model to obtain the second dataset output by the second large language model; the second prompt word is used to instruct the second large language model to perform instruction evolution based on the first dataset. The second dataset is used as the training dataset.

6. The intelligence analysis method based on hierarchical task decomposition of a large language model according to claim 5, characterized in that, The second prompt word includes at least one of the following indications: Add constraints to the description of the historical tasks; Introduce interfering information into the description of the historical tasks; The reasoning steps in the historical reasoning process are extended.

7. An intelligence analysis device based on hierarchical task decomposition of a large language model, characterized in that, include: The determination module is used to determine the intelligence analysis task based on the intelligence data to be analyzed; The task decomposition module is used to input the natural language description corresponding to the intelligence analysis task into the large intelligence analysis model to obtain the current hierarchical workflow output by the large intelligence analysis model; wherein, the current hierarchical workflow includes multiple sub-tasks and multiple pre-dependent task codes; each sub-task corresponds to one or more of the pre-dependent task codes; the pre-dependent task codes are determined based on the dependency relationships between the sub-tasks; the large intelligence analysis model is obtained by training a first large language model based on a pre-constructed training dataset; The generation module is used to generate a task execution graph with the subtasks as graph nodes and the preceding dependent task codes as directed edges. The task execution module is used to execute the intelligence analysis task based on the task execution graph and obtain intelligence analysis results.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the intelligence analysis method based on hierarchical task decomposition of a large language model as described in any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the intelligence analysis method based on hierarchical task decomposition of a large language model as described in any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the intelligence analysis method based on hierarchical task decomposition of a large language model as described in any one of claims 1 to 6.

Images