A method and system for understanding key intelligence requirements based on big models

Abstract

The present invention discloses a method for understanding the intention of key intelligence requirements based on a large model, including: constructing a large model for understanding the intention of key intelligence requirements; the large model includes an interactive instruction understanding algorithm, an operational plan understanding algorithm, a contextual prompt learning technology, a heterogeneous intelligence association fusion algorithm, and a multimodal intelligence understanding algorithm; establishing automated intelligence analysis and understanding by combining efficient parameter fine-tuning, retrieval enhancement generation, and prompt word technology of the large model; understanding the intention of key intelligence requirements based on the large model; and being used for intelligence-based decision-making, which includes generating key information requirements for associated commanders, mining task decision points, and specifying traction branches and subsequent plans. Also disclosed are a system, an electronic device, and a computer-readable storage medium, which are based on a military large language model, utilize efficient parameter fine-tuning of the large model, implement an iterative intelligence vector database, and a flexible intelligence task prompt word template, thereby significantly improving the intelligent processing capability of intelligence data.

Description

Technical Field

[0001] The present invention relates to the fields of large models, intelligence technology, computer simulation, and modeling and simulation technology, and particularly relates to a method and system for understanding the intention of critical intelligence requirements based on large models. Background Art

[0002] Since 2017 when Transformer emerged, artificial intelligence has gradually entered the era of large-scale pre-trained models. BERT trains the model in an encoder "fill-in-the-blank" manner and can achieve good results in specific tasks through fine-tuning of the model, which has attracted the attention of the industry. GPT trains the model by predicting the next word using a decoder, featuring "generation" ability. On November 30, 2022, OpenAI introduced techniques such as instruction fine-tuning, supervised fine-tuning, and reinforcement learning based on human feedback (RLHF) into the model and finally launched a new conversational general artificial intelligence tool - ChatGPT, which demonstrated a high level of human-computer interaction. Relying on the large amount of knowledge stored in the model, ChatGPT has achieved good application effects in aspects such as human-computer question answering, text summary generation, machine translation, classification, and code generation. Artificial intelligence technology based on large language models is gradually being applied to all aspects of scientific research, production, and life. These advanced technologies enable machines to learn and reason independently, and their application scope ranges from situation awareness to decision support, showing strong productivity and changing the working methods and research paradigms in various fields including military and government organizations.

[0003] Large language models have two core advantages, namely the ability to understand and execute complex instructions and the ability to decompose and plan complex tasks. The ability to understand and execute complex instructions is reflected in that when we give detailed instructions to the large model and clearly express task constraints or specifications, the large model can complete the task as required by the instructions. This ability to be faithful to the instruction requirements is highly related to the contextual generation ability of the large model. Given reasonable prompts, and the more abundant and detailed the prompts are, the higher-quality content the large model can often generate. This contextual generation ability essentially reflects a kind of ability to model the world and is independent of the human way of perceiving the world. The ability to decompose and plan complex tasks is another advantage of the large model. It can decompose complex tasks into multiple steps and reasonably plan the execution order of tasks. This provides an important opportunity for vertical domain applications, enabling the large model to work in coordination with traditional information systems, efficiently coordinating numerous systems such as databases, knowledge bases, office automation systems, and code libraries in traditional IT systems to complete complex decision-making tasks that were difficult for traditional intelligent systems to handle in the past, thereby enhancing the intelligent level of the entire information system.

[0004] The key technologies in the existing technical solutions related to large language models include:

[0005] 1. High - efficiency parameter fine - tuning technology for large language models

[0006] The working principle of a large language model (LLM) is to predict the next word through each node in the model's neural network based on the input text and complete the answer in an autoregressive generation manner. Only a very small part of the neural network nodes that determine the final answer in the entire autoregressive generation process of the model account for the total number of nodes. The neural network contains many dense layers to perform matrix multiplication operations, and the weight matrices in these layers are usually full - rank. After fine - tuning for a specific task, the LLM has a low "intrinsic dimension (eigen - rank)", and can still effectively learn even when randomly projected into a smaller subspace. Therefore, when constructing a military large language model in the scenario of intelligence tasks, this patent also selects LoRA (Low - Rank Adaptation), a technology based on low - rank adaptation, as one of the technical approaches for fine - tuning. This technology can achieve a good effect with only a small number of parameters to be trained when using a large model to adapt to downstream tasks, promoting the model to have higher perception, cognition, and decision - making intelligence in the scenario of multi - source and multi - modal intelligence tasks. LoRA has been widely applied to open - source LLMs (such as LLaMA and BLOOM) to achieve efficient parameter fine - tuning. By freezing the pre - trained model weights and injecting trainable rank - decomposition matrices into each layer of the Transformer architecture, LoRA greatly reduces the number of trainable parameters for downstream tasks and effectively improves the fine - tuning efficiency of the pre - trained model on downstream tasks.

[0007] As Figure 1 shown in the technical architecture of LoRA. The left part inside the rounded rectangle in the figure is the original PLM (Pre - trained Language Model). LoRA adds two structures, A and B, beside it. The parameters of these two structures are initialized as Gaussian distribution and 0 respectively, to implement an operation of dimensionality reduction and then dimensionality increase to simulate the intrinsic rank. Initially, A is initialized with a random Gaussian distribution and B is initialized with a 0 matrix to ensure that this bypass matrix is still a 0 matrix at the beginning of training. During training, the parameters of the PLM are fixed, and only the dimensionality - reduction matrix A and the dimensionality - increase matrix B are trained. The input and output dimensions of the model remain unchanged, and when outputting, BA is superimposed with the parameters of the PLM.

[0008] Suppose we want to fine - tune a pre - trained language model (such as LLaMA) for a downstream task, then we need to update the pre - trained model parameters, which is expressed by the following formula:

[0009] W 0 +ΔW

[0010] W0 is the parameter for pre-training model initialization, and ΔW is the parameter to be updated. If it is full-parameter fine-tuning, its number of parameters is W 0 It can be seen that fine-tuning large language models with all parameters is extremely costly. For LoRA, only ΔW needs to be fine-tuned.

[0011] Specifically, assume the pre-trained matrix is Its update can be expressed as:

[0012]

[0013] where the rank r << min(d, k).

[0014] During the training process of LoRA, W 0 remains fixed, and only A and B are training parameters. During the forward process, both W 0 and ΔW are multiplied by the same input x and finally added together:

[0015] h = W 0 x + ΔWx = W 0 x + BAx

[0016] Full Fine-Tuning can be regarded as a special case of LoRA (when r equals k). During the inference process, LoRA hardly introduces additional inference latency, and only W = W 0 + ΔW needs to be calculated. The combination of LoRA and Transformer is also very simple, just adding a bypass in the calculation of QKV Attention.

[0017] During the above fine-tuning process, only the parameters of A and B are updated, and the pre-trained model parameters remain fixed. When performing subsequent inferences, using the re-parametrization idea, AB and W are merged, so that no additional calculations are introduced during inference. And for different downstream tasks, only AB needs to be re-trained on the basis of the pre-trained model, which can also speed up the training rhythm of the large model.

[0018] In summary, LoRA does not require the cumulative gradient update of the weight matrix to have full rank during the adaptation process. When applying LoRA to all weight matrices and training all biases, setting the rank r of LoRA to the rank of the pre-trained weight matrix can roughly restore the expressiveness of full-scale fine-tuning. That is to say, as the number of trainable parameters increases, training LoRA roughly converges to training the original model; in addition, LoRA has no additional inference latency: when deploying a specific individual soldier Agent in the joint training platform, W = W 0+BA and perform inference normally. When it is necessary to switch to another single-soldier Agent for deployment, W can be restored by subtracting BA 0 , and then add a different B′A′, which is a fast operation that only requires very little memory overhead. Compared with the model fine-tuned on the structural parameters, no additional latency is introduced during the LoRA inference process. This ability to switch tasks at a lower cost during deployment can help this patent create more customized single-soldier entity models, and these models can be switched in real time on machines that store the pre-trained weights in VRAM; finally, LoRA can reduce memory and storage resource consumption. For large Transformers trained with Adam, if r << d model , LoRA reduces the VRAM usage by 2 / 3 because there is no need to store the fixed pre-trained parameter W 0 's optimizer state, and large model training can be performed with fewer GPUs.

[0019] 2. Prompt Engineering

[0020] The goal of prompts is to obtain better answers from large language models. Briefly speaking, the process is to repeatedly try and continuously iterate better prompts, discourse structures, and external knowledge for questioning to meet the actual and specific needs of users. Prompt engineering is a series of processes to find more appropriate answers in large language models through more appropriate language organization based on understanding the capabilities and limitations of large language models. Prompts can be a question, a text description, or even a text model with a bunch of parameters. We interact with large models through prompts, and the model will also generate corresponding text, pictures, or even videos and voices according to the information provided by the prompts. Prompts often can include roles, instructions, context information, questions, output requirements, etc. Among them, the instruction part describes the tasks we need the large model to do. Context information can, on the one hand, provide some additional information for the model so that it can better execute your instructions, and on the other hand, it can delimit the information scope for the model. Questions, as the name implies, are the questions that need to be answered by the model. Output requirements can be the formats, numbers, or styles that the model needs to follow when answering, etc.

[0021] MattNight proposed a relatively complete prompt methodology framework CSISPE on ChatGPT-Free-Prompt-List. In the CRISE framework, CR represents Capacity and Role, indicating the role and capabilities that one hopes the large model will play. I represents Insight, which is the background information and context. S is for Statement, describing the task one hopes the model to complete. P is Personality, used to describe the style or manner in which one hopes the large model will answer. E is Experiment, which requires the model to provide multiple answers. The prompt examples based on the CSISP framework are shown in Table 1 below:

[0022] Table 1 Prompt Examples of the CSISP Framework

[0023]

[0024] Specifically, the instructions mainly have the following forms:

[0025] Zero-shot Prompt, which directly uses the large model to perform tasks without providing reference examples. As the most common way of using prompts, zero-shot prompts make full use of the knowledge learned by the large model, being the most flexible and general, and it is also one of the important scenarios for evaluating the capabilities of large models. At the same time, zero-shot prompts rely on pre-trained language models, which may make them subject to limitations and biases in the training data.

[0026] Few-shot Prompt, which realizes the interaction between humans and large models by giving a small amount of prompt text, making the model output more accurate.

[0027] Role Prompt, which plays a "role-playing" game with the large model, making the large model imagine itself as an expert in a certain field to obtain better task effects.

[0028] Chain-of-thought Prompt, commonly used in deduction tasks to solve difficult reasoning problems. Among them, the zero-shot chain-of-thought prompt is to put a sentence "Let's think step by step" at the end of the question. Using the chain-of-thought prompt can make the model's output more coherent and logical. However, we need to note that if one step in the chain of thought is incorrect, the error will also accumulate step by step, resulting in a generated result that does not match the expectation.

[0029] Multimodal Prompt, also known as multi-modal prompt words, enables better interaction with large models by simultaneously inputting multi-modal information such as text and images.

[0030] In summary, the existing technologies have the following defects:

[0031] 1. Traditional intelligence processing systems

[0032] Traditional intelligence processing systems mainly rely on manual operations and use keyword search techniques and information retrieval algorithms. These systems usually search for documents, reports, and other records in a database and extract information that matches specific keywords. Human intelligence analysts are responsible for screening, sorting, and writing intelligence reports.

[0033] The technical defects of this are its low efficiency. Manual operation speed is slow, especially when faced with a large amount of intelligence data, the processing efficiency is low, and it is difficult to quickly respond to battlefield requirements. In addition, keyword search is prone to missing important information that does not contain keywords, resulting in incomplete intelligence. The system has a high dependence on manual operations, requires a large amount of manpower input, increases operation and labor costs, and the system cannot automatically understand and analyze complex intelligence requirements, with low intelligence level, and the quality and efficiency of intelligence analysis are limited by the manual level.

[0034] 2. Rule-based intelligence analysis systems

[0035] Rule-based intelligence analysis systems use predefined rules and logic for intelligence processing. These rules are usually designed by experts and include algorithms such as simple decision trees and logistic regression for automatically classifying and analyzing intelligence data.

[0036] The technical defects of this are its lack of flexibility. The system can only operate according to predefined rules and is difficult to adapt to the rapidly changing battlefield environment and complex and variable intelligence requirements. And it is difficult to maintain. The rules and logic need to be continuously updated and maintained to deal with new intelligence situations and changing environments, which increases the complexity and maintenance cost of the system. In addition, the system performs poorly in dealing with unforeseen new situations and non-standardized data and lacks the ability to respond.

[0037] 3. Machine learning and data mining technologies

[0038] Use machine learning and data mining algorithms to extract valuable intelligence information from large-scale data. These technologies include classification, clustering, association rule mining, etc., and can automatically analyze and predict intelligence trends.

[0039] The technical deficiencies are as follows: it has strong data dependence, the training of the model requires a large amount of high-quality labeled data, the cost of obtaining and labeling data is high, and the time consumption is large; moreover, when machine learning models process unseen data or new scenarios, the effects are often not as good as those in the training dataset, resulting in insufficient generalization ability; in addition, it is difficult to explain the internal analysis process and results for some complex machine learning models (such as deep learning), and commanders may not trust or be unable to make full use of these intelligence analysis results.

[0040] 4. Natural Language Processing (NLP) technology

[0041] Based on natural language processing technology, automated intelligence report generation. These technologies include word vectors, topic models, deep learning, etc., and are used to understand and analyze text data.

[0042] The technical deficiencies are as follows: the understanding ability is limited, traditional NLP models have limited understanding ability for complex contexts and implicit information, and may not be able to comprehensively and accurately analyze and interpret intelligence information; moreover, NLP systems have insufficient processing ability for real-time changing information and dynamic situations, and cannot adjust the analysis focus and direction in real time; in addition, NLP models require a large amount of pre-training and fine-tuning, with high costs, and are less adaptable in specific fields, and need to be continuously adjusted to meet specific intelligence needs. Summary of the Invention

[0043] The purpose of the present invention is to construct a key intelligence requirement intention understanding technology based on large models for task decision points aiming at the defects of the existing technology, and design a key intelligence requirement intention understanding method and system based on large models, including: designing an automated intelligence analysis and understanding solution that combines efficient parameter fine-tuning of large models, retrieval-enhanced generation, and prompt engineering techniques, refining the efficient parameter fine-tuning steps and designing basic prompt templates for different stages of intelligence tasks, so that under the empowerment of multiple intelligent technologies such as large language models, through fine-tuning the existing military large language model in the intelligence task scenario and supplemented by the relevant ideas of prompt engineering, the model can efficiently and high-quality meet the decision-making requirements for key intelligence of decision-makers, analysts, etc., construct a transparent battlefield from the perspective of intelligence, improve the intelligence level of data retrieval, processing, and application, ensure the timeliness and relevance of intelligence, and optimize intelligence task planning through customized prompt templates and fine-tuning datasets, enhance the close combination of intelligence and decision-making, achieve efficient information filtering and sorting, and ultimately improve the intelligence processing efficiency and decision-making quality in military operations.

[0044] The first aspect of the present invention is to provide a key intelligence requirement intention understanding method based on large models, including:

[0045] S1. Build a large model for understanding the intent of critical intelligence requirements. The large model for understanding the intent of critical intelligence requirements includes an interactive instruction understanding algorithm, a combat plan understanding algorithm, a context prompt learning technique, a heterogeneous intelligence association and fusion algorithm, and a multi-modal intelligence understanding algorithm. The large model for understanding the intent of critical intelligence requirements is established based on automated intelligence analysis and understanding using techniques such as efficient parameter fine-tuning of a joint large model, retrieval-augmented generation, and prompt engineering.

[0046] S2. Based on the large model for understanding the intent of critical intelligence requirements, conduct intent understanding of critical intelligence requirements. Among them, the intelligence-based decision-making includes generating key information requirements associated with the commander, mining task decision points, and specifying the traction branch and subsequent plan.

[0047] Preferably, S1 includes:

[0048] S11. Fine-tune the basic military large language model empowered by ChatGLM using initial intelligence data to obtain a fine-tuned model.

[0049] S12. Based on the fine-tuned model, construct a prompt template for a specific domain using prompt engineering techniques to make the fine-tuned model meet the needs of different intelligence tasks.

[0050] Preferably, S11 includes:

[0051] S111. Construct three types of datasets, namely an intelligence instruction fine-tuning dataset, an intelligence vector database, and a preset prompt template library. Among them:

[0052] The intelligence instruction fine-tuning dataset is used for fine-tuning the military large language model to enhance the model's understanding ability and response accuracy for specific types of military tasks. It is designed and generates a high-quality fine-tuning dataset based on specific military intelligence tasks, including task-related questions and answers, as well as elements such as specific scenarios, terms, and protocols.

[0053] The intelligence vector database integrates historical intelligence task data, including past inquiries, reports, analysis results, etc., and incorporates newly generated data by experts into the database to ensure the timeliness and relevance of information.

[0054] The preset prompt template library stores multiple templates designed in advance for different types of intelligence tasks to ensure the accuracy and efficiency of input.

[0055] S112. For a new original question, first perform keyword extraction to identify the core elements and themes of the question. Based on these keywords, select the most appropriate template from a preset prompt template library. Subsequently, fine-tune the basic military large language model using efficient parameter fine-tuning techniques based on an intelligence instruction fine-tuning dataset to form a military large language model adapted to the intelligence field.

[0056] S113. Fine-tune the military large model adapted to the intelligence field based on the parameter-efficient fine-tuning method.

[0057] Preferably, the S113 includes:

[0058] (1) Collect and screen data to obtain intelligence data with different data sources and data modalities, including:

[0059] Collect raw intelligence data: Collect raw data from one or more data sources such as open-source intelligence, human intelligence, signals intelligence, geospatial intelligence, and technical intelligence.

[0060] Screen raw intelligence data: Perform data cleaning, data verification, data classification and marking, data analysis, data fusion, and confidentiality and ethics processing on the raw intelligence data, and collect and screen key information from the multi-source and multi-modal intelligence data as the data basis for the training and decision-making support of the large model.

[0061] (2) Build multiple data models based on the characteristics and requirements of the screened intelligence data and each type of intelligence. Among them, the multiple data models correspond to each type of intelligence category one by one and are used to manage and analyze complex intelligence data. The data models include:

[0062] Entity-Relationship Model (ERM) and text analysis model for open-source intelligence;

[0063] Graph database model and time-driven model for human intelligence;

[0064] Time series model and network traffic analysis model for signals intelligence;

[0065] Geographic Information System (GIS) model and Online Analytical Processing (OLAP) multi-dimensional data model for geospatial intelligence;

[0066] Database model and association rule learning model for technical intelligence;

[0067] (3) Build different categories of datasets based on the constructed data models. Among them, the datasets include text understanding datasets, intelligence-related datasets, decision point mining datasets, and auxiliary task planning datasets.

[0068] The text understanding - related datasets include: an interaction instruction understanding dataset and a multi - modal intelligence understanding dataset;

[0069] The intelligence - related datasets include: an intelligence report dataset, a combat intelligence requirement understanding dataset, and an intelligence task planning and decomposition dataset;

[0070] The decision - point mining datasets include: a decision - point case dataset and a scenario simulation dataset;

[0071] The auxiliary task planning - related datasets include: a command and control communication dataset and a tactics and strategy guide dataset;

[0072] (4) Based on the above - mentioned datasets, construct an intelligence vector dataset and an expected intelligence instruction fine - tuning dataset respectively. The intelligence vector dataset is updated in real - time. Whenever there is the latest real intelligence data, it is stored in it in real - time; the expected intelligence instruction fine - tuning dataset is stored in a data architecture that conforms to the instruction fine - tuning specification after being screened, reviewed, and supplemented by military experts, and is used to fine - tune the basic military large - language model. The expected intelligence instruction fine - tuning dataset is updated periodically according to the fine - tuning requirements;

[0073] (5) Fine - tune the military large - model adapted to the intelligence field based on multi - modal instruction fine - tuning technology and efficient parameter fine - tuning technology for model fine - tuning; where:

[0074] The multi - modal instruction fine - tuning technology is used to generate an intelligence instruction fine - tuning dataset based on prompts and experts by the large - language model. The intelligence instruction fine - tuning dataset contains a seed instruction set and corresponding input and output instance sets to generate a large number of new instruction data and corresponding input - output instances, and then use the generated instruction dataset to fine - tune the large - model for downstream tasks;

[0075] The efficient parameter fine - tuning technology includes setting the model and tokenizer and selecting an appropriate parameter quantization model according to the current fine - tuning data volume.

[0076] Preferably, in S12, based on the prompt engineering technology on the basis of the fine - tuned model, construct a prompt template for a specific domain to make the fine - tuned model meet the needs of different intelligence tasks, including: constructing the prompt template for different intelligence task scenarios; the different intelligence task scenarios include:

[0077] (1) The task scenario of the intelligence data retrieval, processing, and application process;

[0078] (2) The task scenario of the task decision - point mining process;

[0079] (3) The task scenario of the key information requirement generation process for task decision - points; and

[0080] (4) Task scenario for formulating branches and subsequent plans driven by decision points.

[0081] Preferably, the prompt template uses a generalized prompt template as the basis. An effective prompt contains one or more of the following 5 elements: role, instruction, context information, input data, and output indication; a basic prompt template is formed based on historical data, and specific prompt templates are designed based on the basic prompt template for typical different intelligence task scenarios.

[0082] Preferably, for the task scenario of intelligence data retrieval, processing, and application process, the S12 includes:

[0083] S121, determining the mapping relationship between each stage in combat intelligence preparation and the task scenario of intelligence data retrieval, processing, and application process; where the processing of the intelligence data includes intelligence processing and intelligence analysis; for intelligence data retrieval, it corresponds to the intelligence plan formulation and guidance stage and the intelligence collection stage; for intelligence processing, it corresponds to the intelligence processing and utilization stage and the intelligence analysis and production stage; for intelligence analysis, it corresponds to the intelligence distribution and fusion stage and the intelligence evaluation and feedback stage; for intelligence application, it corresponds to the traditional literature service and intelligence product provision stage and the content and knowledge output stage based on understanding the user's intention and related concepts, including literature information provision, professional analysis, processing, and handling of intelligence products.

[0084] S122, designing the prompt template based on the mapping relationship.

[0085] For the task scenario of task decision point mining process, the S12 includes:

[0086] S121’, determining the application scenario of the task decision point mining process task scenario;

[0087] S122’, designing the prompt template based on the application scenario of the task decision point mining process task scenario and the prompt word technology of LLM.

[0088] For the task scenario of generating key information requirements oriented to task decision points, the S12 includes:

[0089] S121”, determining the application scenario of the task scenario of generating key information requirements oriented to task decision points;

[0090] S122”, designing the prompt template based on the application scenario of the task scenario of generating key information requirements oriented to task decision points and natural language processing and deep learning technologies.

[0091] For the task scenario of formulating branches and subsequent plans driven by the decision point, S12 includes:

[0092] S121”’, determine the application scenario of the task scenario of formulating branches and subsequent plans driven by the decision point;

[0093] S122”’, design the prompt template based on the application scenario of the task scenario of formulating branches and subsequent plans driven by the decision point.

[0094] The second aspect of the present invention provides a key intelligence requirement intention understanding system based on a large model, for implementing the method described in the first aspect, including:

[0095] A model construction module, used to construct a key intelligence requirement intention understanding large model; wherein the key intelligence requirement intention understanding large model includes an interactive instruction understanding algorithm, a combat plan understanding algorithm, a context prompt learning technology, a heterogeneous intelligence association and fusion algorithm, and a multi-modal intelligence understanding algorithm; the key intelligence requirement intention understanding large model is established based on automated intelligence analysis and understanding of joint large model efficient parameter fine-tuning, retrieval-augmented generation, and prompt technology;

[0096] A key intelligence requirement intention understanding module, used to perform intention understanding on key intelligence requirements based on the key intelligence requirement intention understanding large model; wherein the intelligence-based decision-making includes generating associated commander key information requirements, mining task decision points, and specifying traction branches and subsequent plans.

[0097] The third aspect of the present invention provides an electronic device, including a processor and a memory, the memory stores multiple instructions, and the processor is used to read the instructions and execute the method described in the first aspect.

[0098] The fourth aspect of the present invention provides a computer-readable storage medium, the computer-readable storage medium stores multiple instructions, and the multiple instructions can be read and executed by the processor to execute the method described in the first aspect.

[0099] The core inventive point of the present invention lies in:

[0100] Construction method of intelligence vector database and intelligence instruction fine-tuning dataset. This method designs a real-time updated intelligence vector database, which collects historical intelligence task data, including inquiries, reports, analysis results, etc., and quickly and effectively retrieves information through advanced data processing technologies (such as vectorization and indexing). In addition, an intelligence instruction fine-tuning dataset generated by experts is designed and constructed for multi-modal instruction fine-tuning of the basic military large language model. Under the support of this solution, the timeliness and relevance of intelligence data are ensured, the understanding and response capabilities of the large model for specific military tasks are improved, and the retrieval, processing, and application processes of intelligence data are optimized. Construct an intelligence vector database and an intelligence instruction fine-tuning dataset to ensure real-time update and quality control of information. In the task inference stage, the model dynamically adjusts the analysis focus and direction by quickly retrieving examples in the intelligence vector database as context resources.

[0101] Large model fine-tuning method based on ChatGLM. Use the intelligence instruction fine-tuning dataset to fine-tune the basic military large language model for specific intelligence tasks, so as to enhance the model's understanding ability and response accuracy for specific types of military tasks. Finally, improve the model's performance in different task scenarios, enhance the real-time data adjustment ability of the generated content of the model, and promote the mining of task decision points and the generation of key intelligence requirements. Perform efficient parameter fine-tuning on the military large language model based on ChatGLM technology to enhance the model's understanding and response capabilities for specific military tasks.

[0102] Intelligence task planning prompt template driven by task decision points. Through the flexible construction of the prompt word template, customized information filtering and sorting options are provided, and examples and other information are organically combined to form the prompt input of the large language model. Make the model output adapt to the preferences and needs of different commanders, and be able to automatically adjust the information display and analysis methods according to the current task and environment, improving the efficiency of intelligence writing and human-computer interaction. Preset a prompt word template library to realize the customization and systematization of intelligence task planning, ensure the accuracy and efficiency of input, automatically generate high-quality intelligence that meets the current task requirements, provide efficient information filtering and sorting options, and enhance the close combination of intelligence and decision-making.

[0103] Beneficial effects of the method and system of the present invention:

[0104] Based on the military large language model, the present invention uses high-efficiency parameter fine-tuning of the large model, implements an iterative intelligence vector database, and a flexible intelligence task prompt word template, significantly improving the intelligent processing ability of intelligence data. Compared with the traditional system that relies on manual keyword search and writing, the present invention can more quickly identify and extract key information in the following aspects, improving the efficiency and accuracy of intelligence retrieval, integration, and analysis.

[0105] Real-time update and efficient retrieval. By constructing an intelligence vector database, the present invention ensures that intelligence information can be updated in real time to adapt to the rapidly changing military environment. Compared with the static and lagging data processing in traditional intelligence systems, the present invention can dynamically adjust the analysis focus and direction, providing more timely and relevant intelligence support.

[0106] Customized intelligence task planning. The present invention realizes the customization and systematization of intelligence task planning through a preset prompt template library. The templates in the prompt template library are designed in advance according to different types of intelligence tasks, ensuring the accuracy and efficiency of input, enabling the system to generate the optimal intelligence plan according to specific task requirements, and enhancing the close combination of intelligence and decision-making.

[0107] Efficient parameter fine-tuning. Using ChatGLM technology to perform efficient parameter fine-tuning on the military large language model enhances the model's understanding ability and response accuracy for specific military tasks. Compared with traditional models, the present invention can generate intelligence that adapts to different task scenarios more precisely, meeting the personalized needs of commanders.

[0108] Automated intelligence generation. The present invention realizes the automation of intelligence generation by constructing an intelligence instruction fine-tuning data set and an intelligence vector database, combined with efficient parameter fine-tuning technology. Compared with traditional intelligence systems, the present invention can quickly retrieve and analyze historical data, organically combine relevant information to form input, and automatically generate high-quality intelligence that meets the current task requirements, significantly improving the intelligence processing efficiency.

[0109] Flexible adjustment of analysis focus. In the inference stage, the large language model can adjust the analysis focus and direction according to real-time data by quickly retrieving examples in the intelligence vector database as context resources. Compared with traditional systems with a fixed analysis mode, the present invention has stronger adaptability and flexibility, and can provide optimal intelligence analysis and decision support according to changes in tasks and environments. Brief Description of the Drawings

[0110] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in related technologies, the following will briefly introduce the drawings required for use in the description of the specific embodiments or related technologies. Obviously, the following drawings are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0111] Figure 1 It is a schematic diagram of the technical architecture of LoRA provided according to the prior art.

[0112] Figure 2 It is a schematic diagram of the process architecture provided according to the embodiment of the present invention;

[0113] Figure 3 Schematic diagram of the technical architecture provided according to an embodiment of the present invention;

[0114] Figure 4 Schematic diagram of the LoRA weight merging process provided according to an embodiment of the present invention;

[0115] Figure 5 Flowchart of the method for understanding the intent of key intelligence requirements based on a large model provided according to an embodiment of the present invention;

[0116] Figure 6 System architecture diagram of the method for understanding the intent of key intelligence requirements based on a large model provided according to an embodiment of the present invention;

[0117] Figure 7 Structural diagram of an electronic device provided according to an embodiment of the present invention. Detailed implementation manners

[0118] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Apparently, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0119] In the description of the present invention, it should be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc. indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings. It is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and thus should not be construed as a limitation of the present invention. In addition, the terms "first", "second", and "third" are only used for descriptive purposes and cannot be construed as indicating or implying relative importance.

[0120] In the description of the present invention, it should be noted that unless otherwise clearly defined and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be directly connected or indirectly connected through an intermediate medium, and it can be the communication inside two elements. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific situations.

[0121] Such as Figure 2-3As shown, the advantage of the large language model is that it is a "generalist", and many of its potential capabilities require us to be familiar with, explore, and even "awaken" them in our daily work in order to play a greater role alone or in combination with intelligence-specific tools. In the context of a generalist, the purpose of this patent is to fine-tune the basic military large language model empowered by ChatGLM through initial intelligence data. Based on the fine-tuned model, specific domain prompt templates are constructed based on prompt engineering techniques to enable the model to meet the needs of different intelligence tasks. In the initial stage of intelligence preparation, based on the intelligence vector database, optimize the process of intelligence data retrieval, processing, and application; in the mission planning stage, promote the excavation of mission decision points; when the commander reaches the mission decision point, associate the generation of the commander's key intelligence requirements; in the decision-making process, face the mission decision point and drive the formulation of branches and subsequent plans.

[0122] This embodiment aims to solve a number of technical problems existing in the traditional military intelligence processing and application process, including:

[0123] (1) First, the traditional intelligence system relies on manual keyword search and writing, resulting in low intelligence retrieval efficiency, easy omission of key information, and the response speed cannot meet the rapidly changing military needs.

[0124] (2) Second, the existing intelligence analysis lacks intelligence and cannot fully understand and respond to the complex military task requirements, resulting in the disconnection between intelligence and decision-making.

[0125] (3) Third, the timeliness and accuracy of intelligence data processing are insufficient, making it difficult to provide timely and effective support in a dynamic environment.

[0126] (4) Fourth, the intelligence task planning lacks systematization and automation, and cannot generate the key intelligence required by the commander according to the mission decision point, affecting the scientificity and effectiveness of decision-making. Embodiment 1

[0127] As Figure 5 shown, this embodiment provides a method for understanding the intention of key intelligence requirements based on a large model, including:

[0128] S1. Construct a large model for understanding the intent of key intelligence requirements. The large model for understanding the intent of key intelligence requirements includes an interactive instruction understanding algorithm, a combat plan understanding algorithm, a context prompt learning technique, a heterogeneous intelligence association and fusion algorithm, and a multi-modal intelligence understanding algorithm. In this embodiment, S1 includes an automated intelligence analysis and understanding solution that combines high-efficiency parameter fine-tuning of the large model, retrieval-augmented generation, and prompt engineering techniques, refines the high-efficiency parameter fine-tuning steps, and designs basic prompt templates for different stages of intelligence tasks. By constructing a "customized preset prompt template library (updated as needed) + intelligence vector database (updated in real time) + fine-tuned large model", a quick response and in-depth understanding of complex intelligence requirements can be achieved in different task scenarios.

[0129] As a preferred embodiment, S1 includes:

[0130] S11. Fine-tune the basic military large language model empowered by ChatGLM using initial intelligence data to obtain a fine-tuned model.

[0131] S12. Based on the prompt engineering technique, construct prompt templates for specific domains on the basis of the fine-tuned model, so that the fine-tuned model meets the needs of different intelligence tasks.

[0132] As a preferred embodiment, S11 includes:

[0133] S111. Construct three types of datasets, namely, an intelligence instruction fine-tuning dataset, an intelligence vector database, and a preset prompt template library. Among them:

[0134] The intelligence instruction fine-tuning dataset is used for the fine-tuning of the military large language model to enhance the model's understanding ability and response accuracy for specific types of military tasks. It is designed and generates high-quality fine-tuning datasets based on specific military intelligence tasks (such as combat plan generation, intelligence task planning), includes task-related questions and answers, and also includes elements such as specific scenarios, terms, and protocols.

[0135] The intelligence vector database integrates historical intelligence task data, including past inquiries, reports, analysis results, etc., and incorporates newly generated data by experts into the database to ensure the timeliness and relevance of information.

[0136] The preset prompt template library stores multiple templates designed in advance for different types of intelligence tasks to ensure the accuracy and efficiency of input.

[0137] S112. For a new original question, first perform keyword extraction to identify the core elements and themes of the question. Based on these keywords, select the most suitable template from the preset prompt template library. Subsequently, use the efficient parameter fine-tuning technique on the intelligence instruction fine-tuning dataset to fine-tune the basic military large language model to form a military large language model adapted to the intelligence field.

[0138] The application methods of these three types of datasets in real application scenarios include:

[0139] In the use of a certain task scenario, first retrieve similar historical cases and reports in the intelligence vector database for the original question, select highly relevant historical information as a reference, and use it as part of the input to improve the model's understanding of the current question and the relevance of the answer. In addition, retrieve and obtain the prompt template with the highest adaptability to the current task objective in the preset prompt template library, and effectively concatenate the original question, task examples, and other important prompt words to form the final input. Finally, based on this carefully prepared input, the military large language model will generate a corresponding answer. This answer aims to meet the specific intelligence requirements raised by the original question. Other parts of the specification will elaborate on the model fine-tuning and prompt template design schemes in the intelligence task scenario.

[0140] S113. Fine-tune the military large model adapted to the intelligence field;

[0141] The scale of pre-trained language models (PLMs) represented by ChatGPT is getting larger and larger. Limited by computing power costs and training costs, the cost of full fine-tuning is relatively high, and it is still difficult to optimize the large model in all specific intelligence task scenarios. The parameter-efficient fine-tuning (PEFT) method is proposed to solve these two problems. PEFT can enable the PLM to efficiently adapt to various downstream application tasks without fine-tuning all the parameters of the pre-trained model, fine-tuning a small number of or additional model parameters, and fixing most of the pre-trained parameters, greatly reducing the computing and storage costs. At the same time, the most advanced PEFT technology can also achieve performance comparable to full fine-tuning.

[0142] The solution of this patent embodiment uses efficient parameter fine-tuning methods such as LoRA to conduct preliminary fine-tuning experiments without changing the internal parameters of the model, confirm the effectiveness of the constructed dataset, and then synchronously perform full-parameter instruction fine-tuning to update the internal parameters of the model to form specific capabilities. The detailed construction scheme is described as follows.

[0143] As a preferred implementation, the S113 includes:

[0144] (1) Collect and screen data to obtain intelligence data with different data sources and data modalities;

[0145] Collecting data includes: The quality and quantity of data determine the model's "ability from data to decision". For effective fine-tuning, the data needs to be appropriately structured. When constructing a military large language model for intelligence-related task scenarios, collecting multi-source and multi-modal data is a key step. First, it is necessary to collect intelligence data from different data sources and different data modalities. The embodiments of this patent are designed to collect from the sources shown in Table 2 below.

[0146] Table 2 Main data sources

[0147]

[0148] Screening data includes: performing a series of processes such as cleaning, verifying, classifying, and tagging the intelligence data, collecting and screening key information from multi-source and multi-modal intelligence, and providing a high-quality and strongly relevant data basis for subsequent model training and decision support. Such data processing not only enhances the accuracy and reliability of intelligence but also improves the ability to understand complex combat environments.

[0149] The processing flow is shown in Table 3 below.

[0150] Table 3 Data processing flow

[0151]

[0152] (2) Based on the screened intelligence data integration and the characteristics and requirements of each type of intelligence, construct multiple data models; among them, the multiple data models correspond one-to-one with each type of intelligence category and are used to manage and analyze complex intelligence data;

[0153] For different categories of intelligence, when constructing a data structure model, it is necessary to consider the characteristics and requirements of each type of intelligence. Based on the data sources in step (1), the designed data models are shown in Table 4 below.

[0154] Table 4 Examples of some designed data models

[0155]

[0156] The patent designs the above models to integrate data from different intelligence sources. By customizing data structure models for each type of intelligence category, it can more effectively manage and analyze complex military intelligence data, support more accurate and in-depth intelligence work, and ultimately provide a comprehensive analysis perspective. For example, combining information from OSINT and SIGINT to analyze a specific event or trend. In addition, the data model should be designed to be easily extensible and updatable to adapt to the addition of new intelligence and changes in analysis requirements.

[0157] (3) Based on the constructed data model, different categories of data sets are expected to be constructed, and the detailed classification is as shown in Table 5 below:

[0158] Table 5 Expected Constructed Data Sets

[0159]

[0160] (4) Based on the above 4 categories and a total of 9 data sets, an intelligence vector data set and an expected intelligence instruction fine-tuning data set are respectively constructed. The intelligence vector data set is updated in real time, and whenever there is the latest real intelligence data, it is stored in it in real time; the expected intelligence instruction fine-tuning data set is stored in a data architecture that conforms to the instruction fine-tuning specification after being screened, checked, and supplemented by military experts, and is used for instruction fine-tuning of the basic military large language model. The expected intelligence instruction fine-tuning data set is updated periodically according to the fine-tuning requirements.

[0161] (5) Fine-tune the military large model adapted to the intelligence field based on multi-modal instruction fine-tuning technology and efficient parameter fine-tuning technology;

[0162] In this embodiment:

[0163] Multi-modal instruction fine-tuning technology: The large language generates a large number of new instruction data and corresponding input and output instances based on the prompt template and the intelligence instruction fine-tuning data set generated by experts, which includes a seed instruction set and the corresponding input and output instance sets, and then uses the generated instruction data set to perform instruction fine-tuning on the large model for downstream tasks. Compared with knowledge distillation, there are two significant differences in the construction of this instruction data set: ① Knowledge distillation uses the knowledge of the large model to guide the training of the small model, and this method uses the large model itself to guide the fine-tuning of the large model in downstream tasks; ② Knowledge distillation can only extract the knowledge within the domain learned by the large model, while most of the instruction data and corresponding input and output instances generated by this method are outside the data distribution of the pre-training stage.

[0164] Efficient parameter fine-tuning technology: After the software and hardware environment is set up, set the model and the tokenizer, and select an appropriate parameter quantization model according to the current amount of fine-tuning data. If you want to complete the fine-tuning process in a shorter time and with less memory, give priority to using 4-bit quantization provided by BitsAndBytesConfig. Then load and process the data. The data set is loaded based on the file format specified by the data_path parameter. Load the data set from a JSON or JSON Lines file (specified as the file path by the data_path parameter). Then, use tokenize to preprocess the data. Concatenate the input and output text data and encode them into token IDs, and append the end-of-sequence marker. Then prepare the encoded data as the input of the model.

[0165] Before calling LoRA for LLM fine-tuning, hyperparameters need to be set in advance. In terms of technical implementation, this patent design calls the LoraConfig method provided by the PeFT framework to configure LoRA parameters for subsequent fine-tuning. Table 6 below provides a brief explanation of the key hyperparameters and the selection rules. In subsequent repeated fine-tuning work, the following hyperparameters need to be adjusted by observing the changes in model performance and resource consumption, especially lora_alpha and lora_rank, to find the optimal settings for specific tasks.

[0166] Table 6 Hyperparameter Introduction

[0167]

[0168] During the fine-tuning process, the smaller weight matrices A and B are separate. However, once the training is completed, the weights can be merged into a single new weight matrix. Although LoRA is significantly smaller and the training speed is faster, due to separately loading the base model and the LoRA model, there may be latency issues during subsequent inference. To eliminate the latency, this patent selects the merge_and_unload function to merge the adapter weights with the base model, effectively using the newly merged model as an independent model. Multiple LoRA models can also be merged, enabling the Base Model to handle multiple tasks simultaneously.

[0169] The merging process is as Figure 4 shown, W merged = W + BA replaces the original W. Multiple Lora models can be merged through the merge_and_unlaod method. By merging the CKPTs of different types of tasks, the original model can simultaneously possess the capabilities for multiple downstream tasks without affecting the inference efficiency. Call the save_pretrained method to save the model.

[0170] As a preferred implementation, in step S12, based on the fine-tuned model and prompt engineering techniques, a prompt template for a specific domain is constructed to enable the fine-tuned model to meet the needs of different intelligence tasks, including: constructing the prompt template for different intelligence task scenarios; the different intelligence task scenarios include:

[0171] (1) Task scenarios for intelligence data retrieval, processing, and application processes;

[0172] (2) Task scenarios for mining task decision points;

[0173] (3) Task scenarios for generating key information requirements for task decision points; and

[0174] (4) Task scenarios for formulating branches and subsequent plans driven by decision points.

[0175] As a preferred implementation, the prompt word template uses a generalized prompt word template as a basis. An effective prompt contains one or more of the following five elements: role, instruction, context information, input data and output instruction; a basic prompt word template is constructed based on historical data, and a specific prompt word template is designed based on the basic prompt word template for typical different intelligence task scenarios.

[0176] The prompt word template is the necessary glue to boost the large language model to output better. During the reasoning process, the large language model will retrieve one or more samples from the vector database as demonstrations based on the key information in the original question, and select the prompt word template suitable for the current scenario in the preset template library as the organizational structure of the input information. The selected specific task prompt word template will combine the demonstration data and other information into a natural language prompt in a specific order. Finally, the natural language prompt is input to the fine-tuned large language model to form an accurate and standardized output that meets the commander's needs.

[0177] Combining the existing prompt word technology, experience and practice, the present invention adopts a generalized prompt word template as the basis. A prompt can contain the following 5 elements, as shown in the following table. An effective prompt does not necessarily have to have all the elements, but is based on the specific analysis of the specific task. The following will design specific prompt word templates for some typical intelligence task scenarios based on the basic prompt word template shown in Table 7 below.

[0178] Table 7 Basic prompt word structure of this patent

[0179]

[0180] As a preferred implementation, for the intelligence data retrieval, processing and application process task scenario, the S12 includes:

[0181] In the field of modern military intelligence, the application of large language models is crucial to improving the efficiency of intelligence retrieval, processing, analysis and application. For military large language models fine-tuned with intelligence data, users can use some basic prompt word templates and structured input to allow the large model to understand user needs and produce answers that meet user expectations.

[0182] S121, determine the mapping relationship between each stage in combat intelligence preparation and the intelligence data retrieval, processing and application process task scenario; wherein the intelligence data processing includes intelligence processing and intelligence analysis;

[0183] In this embodiment:

[0184] 1. Intelligence data retrieval:

[0185] (1) Intelligence Planning and Guidance Phase: In this phase, the large language model, through its general knowledge capabilities and multi-round dialogue optimization, helps intelligence workers refine the novelty search topic, understand the operational environment, and potential threats from the enemy. The model can quickly identify information related to the possible actions of the enemy, improving the accuracy and relevance of intelligence collection. The key is to determine intelligence requirements through efficient information retrieval, laying the foundation for subsequent intelligence collection and analysis.

[0186] (2) Intelligence Collection Phase: In this stage, the large language model helps analyze and filter the large amount of information collected. The model can process multilingual content, translating, summarizing, and classifying important information to improve the usability and comprehensibility of intelligence. In addition, the model can also assist in identifying key time nodes, geographical locations, and types of enemy actions, making the collected intelligence more specific and targeted.

[0187] 2. Intelligence Processing:

[0188] (1) Intelligence Processing and Utilization Phase: In this phase, the analysis ability of the large language model plays a key role. The model can quickly analyze and integrate intelligence data obtained from various sensors and sources. The model helps identify important information gaps, and through comparison and analysis, proposes possible action plans and strategic recommendations. Through intelligent analysis, the model can also assist in identifying potential risks and opportunities, providing support for strategic decision-making.

[0189] (2) Intelligence Analysis and Production Phase: In a joint operational environment, the large language model can be used to generate intelligence products, such as operational environment analysis, enemy capability assessment reports, etc. The model can customize the generation of intelligence products according to user needs and the strategic intentions of commanders, providing targeted intelligence support. In this phase, the model can also assist in completing literature services, providing users with various types of literature information required, including historical data, previous research, etc.

[0190] 3. Intelligence Analysis:

[0191] (1) Intelligence Distribution and Fusion Phase: In this phase, the large language model helps ensure the rapid and effective distribution and fusion of intelligence. The model can assist in integrating intelligence into decision support systems. The model can also help various departments understand and utilize intelligence, ensuring the consistency and accuracy of intelligence.

[0192] (2) Intelligence Evaluation and Feedback Phase: This is a continuous process. In this phase, the large language model can be used to evaluate the effectiveness and accuracy of intelligence. The model can assist in analyzing the effectiveness of the intelligence collection and processing processes to ensure that the needs of commanders and joint forces are met.

[0193] 4. Intelligence Application:

[0194] In terms of intelligence applications, the application of large models is not limited to traditional literature services and the provision of intelligence products, but rather the output of content and knowledge based on understanding user intentions and related concepts. This includes, but is not limited to, the provision of literature information, the professional analysis, processing, and handling of intelligence products.

[0195] The application of large language models in the modern military intelligence field not only improves the efficiency and quality of intelligence processing, but also provides more comprehensive and in-depth support for military decision-making. Through the above methods, large language models can greatly improve the efficiency and quality of intelligence. In the stage of intelligence plan formulation and guidance, the model lays a solid foundation for intelligence collection through precise retrieval and theme refinement. In the intelligence collection stage, the multi-language processing and information classification capabilities of the model make the collected information more specific and targeted. In the stage of intelligence processing and utilization, the analysis and integration capabilities of the model provide strong support for strategic decision-making. In the stage of intelligence analysis and production, the model not only accelerates the circulation and application of information, but also improves the customization and personalization level of intelligence products.

[0196] S122, design the prompt template based on the mapping relationship;

[0197]

[0198] As a preferred implementation, for the task scenario of the task decision point mining process, the S12 includes:

[0199] S121’, determine the application scenario of the task scenario of the task decision point mining process;

[0200] In modern military operations, the identification and analysis of task decision points are crucial. Facing task decision points, large language models can comprehensively understand the impact of battlefield environments such as meteorological factors on combat operations, recommend locations suitable for various types of equipment (including drones, ships, etc.) to perform combat tasks, develop action plans for one's own side, and assist commanders in making decisions.

[0201] Task decision points are divided into expected task decision points and potential task decision points. For expected task decision points, traditional methods use critical event analysis to identify expected decision points. Because regardless of the type of military operation and the operational environment, and no matter how many variables may be encountered during the implementation of the operation, the occurrence of some critical events is inevitable at the initial stage of the operation, and the tasks that must be completed are also relatively certain. The decision points are associated with critical events, and expected decision points are set for each critical event, and a detailed analysis and deduction are carried out on these certain-to-occur critical events and expected decision points, so that when the operation progresses to the decision point, the commander and his staff team can ensure the synchronization and coordination of the operation forces and resources, ensure the achievement of the expected operation effect, and guide the situation onto the development track of the friendly operation. For expected task decision points, traditional methods use wargaming of operation plans to discover potential decision points. Conducting wargaming on operation plans can enable commanders at all levels and their staff teams to better understand and comprehend their own operation plans and the possible threats in the plans, so as to scientifically judge the advantages and disadvantages of each operation plan and lay a foundation for subsequent comparison and approval of operation plans.

[0202] With the progress of technology, large language models (LLMs) provide a novel approach to assist this process, especially in leveraging prompt engineering to extract accurate decision point information. This patent designs a combination of traditional methods and the prompt technology of LLMs to create a more efficient and in-depth decision point extraction process.

[0203] S122’, design the prompt template based on the application scenario of the task scenario in the task decision point extraction process and the prompt technology of the LLM;

[0204] First of all, by leveraging the powerful data processing capabilities of LLMs, we can quickly analyze a large amount of historical military operation and strategic decision-making data. This includes not only public historical records, but also internal reports, communication records, and intelligence data. By designing targeted prompts for the model, we can guide the model to identify and analyze cases that were regarded as critical events in the past, thus helping decision-makers identify possible expected decision points.

[0205] In addition, in terms of mining potential decision points, the prompting techniques of large language models (LLMs) can simulate different battlefield scenarios and action plans. Compared with traditional war gaming, LLMs can generate and analyze various possible scenarios more quickly, thereby revealing potential decision points that may not have been noticed in traditional methods. For example, by setting specific battlefield conditions or enemy behaviors as prompts, LLMs can generate multiple scenarios related to these conditions, helping decision-makers explore and identify key decision points that may arise in different situations. Additionally, leveraging the capabilities of LLMs, we can also analyze the decision-making requirements in specific situations in more depth. For example, by combining Geographic Information System (GIS) data and real-time intelligence, LLMs can provide detailed analyses of specific locations and times, thus helping decision-makers more accurately locate decision points.

[0206] In practical applications, the prompting techniques of LLMs can be combined with traditional war gaming. Traditional war gaming provides judgment based on experience and tactical considerations, while LLMs provide in-depth analysis based on large amounts of data. Through this combination, we can not only improve the efficiency of mining decision points but also enhance its accuracy and comprehensiveness, enabling the model to not only process large amounts of historical data and simulate complex battlefield scenarios but also provide in-depth analysis in real-time situations, thereby helping decision-makers make more accurate and effective decisions in complex and dynamic military environments.

[0207] The prompting template examples designed in this embodiment are as follows:

[0208]

[0209] As a preferred implementation, for the task scenario of the key information requirement generation process for task-oriented decision points, S12 includes:

[0210] S121”, determining the application scenario of the task scenario of the key information requirement generation process for task-oriented decision points;

[0211] In military operations, commanders' decisions are usually faced with urgency in terms of time and space. This requires commanders to make critical decisions at specific time points and locations, and these decisions rely on accurate and timely intelligence support.

[0212] S122”, designing the prompting template based on the application scenario of the task scenario of the key information requirement generation process for task-oriented decision points and natural language processing and deep learning techniques;

[0213] In this embodiment, the application of large language model prompting engineering in the process of generating key information requirements for task-oriented decision points includes:

[0214] The information requirements of commanders are mainly divided into two categories: Priority Intelligence Requirements and Information Requirements of Own Troops. Priority Intelligence Requirements mainly focus on the enemy or potential threats, involving the formulation of intelligence collection plans, the clarification of intelligence forces and task priorities. Information Requirements of Own Troops involve the status, capabilities and operation plans of own troops. The practice of the US military shows that these information requirements are crucial for ensuring the effectiveness of battlefield decisions. The intelligence department plays a dual role in combat: one is to provide intelligence support for combat planning; the other is to plan intelligence reconnaissance operations according to combat requirements. In meticulous operation planning, the intelligence department prepares for possible emergencies through a complete intelligence planning and support line. In crisis operation planning, due to time constraints, intelligence support and planning content need to be tailored to fit the combat plan.

[0215] In this context, the empowerment of large language model prompt engineering becomes crucial. This technology can help intelligence analysts quickly generate key information requirements related to specific decision points. Through natural language processing and deep learning technologies, the model can: Based on historical intelligence data and past operations, the model can predict future information requirements, thus guiding the priority of intelligence collection; identify potential patterns and trends in enemy actions, providing strategic insights for commanders; as the battlefield situation changes, the model can update information requirements in real time to ensure that commanders obtain the latest intelligence; evaluate the potential risks and benefits of different decision-making options to assist commanders in making more scientific decisions.

[0216] The example of the prompt template designed in this embodiment is as follows:

[0217]

[0218] As a preferred implementation method, for the task scenario of formulating branches and subsequent plans led by the decision point, the S12 includes:

[0219] S121”’, determine the application scenario of the task scenario of formulating branches and subsequent plans led by the decision point;

[0220] In the modern military combat environment, decision-making and intelligence processing are two important and complementary links. Especially in a complex and ever-changing battlefield environment, the formulation of strategic decisions often needs to be based on a large amount of accurate and timely intelligence information.

[0221] S122”’, design the prompt template based on the application scenario of the task scenario of formulating branches and subsequent plans led by the decision point.

[0222] In this embodiment, the deduction and revision of the opponent's action plan, as well as the intelligence requirement model for strategic decision-making, provide two important cases to help the present invention understand how to use large language model prompt engineering to empower and formulate branches and subsequent plans driven by decision points under high-pressure and high-uncertainty conditions.

[0223] First, consider the process of deducing and revising the opponent's action plan. In actual operation, after the first deduction of the action plan, some deficiencies and potential decision points will be exposed. At this time, the staff team needs to adjust and optimize the plan to improve its pertinence and response flexibility. In this process, the large language model can play a key role. By analyzing historical data, battlefield environment, and possible hostile actions, the model can help decision-makers identify potential weaknesses and opportunities. For example, if the model predicts that the enemy may change the attack direction or a new threat appears, the staff team can formulate branch plans and subsequent plans accordingly.

[0224] In practical applications, the formulation of branches and subsequent plans driven by decision points generally includes the following steps:

[0225] · Intelligence collection and analysis: Use the large language model to analyze the collected data and identify key decision points, such as the potential actions of the enemy or the strategic weaknesses of one's own side.

[0226] · Plan deduction and revision: Based on the analysis results provided by the model, conduct a preliminary deduction of the plan and identify the parts that need to be revised.

[0227] · Branch plan formulation: For the identified key decision points, formulate branch plans to cope with possible battlefield changes or emergencies.

[0228] · Subsequent plan formulation: Consider the possibilities of different battle outcomes, such as victory, defeat, or stalemate, and formulate corresponding subsequent action plans for each situation.

[0229] · Real-time adjustment and optimization: During the actual operation process, continuously revise and optimize the branches and subsequent plans according to the real-time changes in the battlefield situation.

[0230] · Decision execution and feedback: Execute the decision, and based on the feedback of the action results, continuously optimize the model and decision-making process.

[0231] In this way, large language models not only provide support in the intelligence analysis stage, but also play a crucial role throughout the entire decision-making process. By incorporating some thought chain ideas through structured prompt templates, the model predicts and simulates different battlefield situations step by step, helping decision-makers better understand and predict possible changes, and thus make more flexible and adaptable decisions. This decision-making process not only increases the success rate of operations, but also enables rapid strategic adjustments in response to unexpected situations to cope with the ever-changing battlefield environment.

[0232] An example of the designed prompt template in this embodiment is as follows:

[0233]

[0234] S2. Based on the key intelligence requirement intention, the large model is used to understand the intention of the key intelligence requirement; wherein the intelligence-based decision-making includes generating associated commander key information requirements, mining task decision points, and specifying the traction branch and subsequent plans. Embodiment 2

[0235] As Figure 6 shown, this embodiment provides a system for understanding the intention of key intelligence requirements based on a large model, which is used to implement the method described in the first aspect, including:

[0236] A model construction module for constructing a large model for understanding the intention of key intelligence requirements; wherein the large model for understanding the intention of key intelligence requirements includes an interactive instruction understanding algorithm, a combat plan understanding algorithm, a context prompt learning technique, a heterogeneous intelligence association and fusion algorithm, and a multi-modal intelligence understanding algorithm;

[0237] A key intelligence requirement intention understanding module for understanding the intention of key intelligence requirements based on the large model for understanding the intention of key intelligence requirements; wherein the intelligence-based decision-making includes generating associated commander key information requirements, mining task decision points, and specifying the traction branch and subsequent plans.

[0238] The core inventive point of this embodiment lies in:

[0239] Method for constructing an intelligence vector database and an intelligence command fine-tuning dataset. This method designs a real-time updated intelligence vector database that collects historical intelligence mission data, including inquiries, reports, analysis results, etc., and retrieves information quickly and efficiently through advanced data processing technologies (such as vectorization and indexing). In addition, an expert-generated intelligence command fine-tuning dataset is designed and constructed for multi-modal command fine-tuning of basic military large-scale language models. With the support of this solution, the timeliness and relevance of intelligence data are ensured, the understanding and response capabilities of large models for specific military tasks are improved, and the retrieval, processing and application processes of intelligence data are optimized. Construct an intelligence vector database and an intelligence command fine-tuning dataset to ensure real-time information updates and quality control. In the task reasoning stage, the model dynamically adjusts the focus and direction of analysis by quickly retrieving samples in the intelligence vector database as context resources.

[0240] Large model fine-tuning method based on chatglm. Use the intelligence instruction fine-tuning dataset to fine-tune the basic military large language model for specific intelligence tasks to enhance the model's understanding and response accuracy for specific types of military tasks. Ultimately, improve the model's performance in different mission scenarios, enhance the model's ability to adjust real-time data of generated content, and promote mission decision point mining and key intelligence demand generation. Based on ChatGLM technology, perform efficient parameter fine-tuning on the military large language model to enhance the model's understanding and response capabilities for specific military tasks.

[0241] Task decision point driven intelligence mission planning prompt template. Through the flexible construction of prompt word templates, customized information filtering and sorting options are provided, and examples and other information are organically combined to form prompt inputs for large language models. The model output is adapted to the preferences and needs of different commanders, and the information display and analysis methods can be automatically adjusted according to the current mission and environment, improving the efficiency of intelligence writing and human-computer interaction. The preset prompt word template library realizes the customization and systematization of intelligence mission planning, ensures the accuracy and efficiency of input, automatically generates high-quality intelligence that meets the current mission requirements, and provides efficient information filtering and sorting options to enhance the close integration of intelligence and decision-making.

[0242] The present invention also provides a memory storing a plurality of instructions, the instructions being used to implement the method of the first embodiment.

[0243] If Figure 7 As shown, the present invention also provides an electronic device, including a processor 301 and a memory 302 connected to the processor 301, wherein the memory 302 stores a plurality of instructions, which can be loaded and executed by the processor, so that the processor can execute the method as in the first embodiment.

[0244] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for understanding key intelligence requirements based on a large model, characterized in that: include: S1, building a big model for understanding key intelligence requirements and intentions; The key intelligence demand intention understanding big model includes interactive command understanding algorithm, combat plan understanding algorithm, contextual prompt learning technology, heterogeneous intelligence association fusion algorithm and multimodal intelligence understanding algorithm; the key intelligence demand intention understanding big model is based on the automated intelligence analysis and understanding of the joint big model efficient parameter fine-tuning, retrieval enhancement generation and prompt word technology; S2, understanding the intent of key intelligence requirements based on the key intelligence requirements intention understanding big model; Intelligence-based decision-making includes generating key information requirements for associated commanders, exploring mission decision points, and specifying traction branches and follow-up plans; S1 includes: S11, fine-tuning the basic military large language model enabled by chatglm through initial intelligence data to obtain a fine-tuned model; S12, based on the fine-tuning model, based on the prompt word engineering technology, construct prompt word templates for specific fields, so that the fine-tuning model can meet the needs of different intelligence tasks; S11 includes: S111, construct three types of data sets, namely, intelligence instruction fine-tuning data set, intelligence vector database and preset prompt word template library; S112, for a new original question, first perform keyword extraction to identify the core elements and themes of the question, and based on these keywords, select the most appropriate template from the preset prompt word template library; then, based on the intelligence command fine-tuning dataset, use efficient parameter fine-tuning technology to fine-tune the basic military large language model to form a military large language model adapted to the intelligence field; S113, fine-tuning a large military model adapted to the intelligence domain based on an efficient parameter fine-tuning method; S113 includes: (1) Collect and filter data to obtain intelligence data with different data sources and data modalities; (2) Building multiple data models based on the integration of the filtered intelligence data and the characteristics and requirements of each intelligence category; wherein multiple data models correspond to each intelligence category one by one and are used to manage and analyze complex intelligence data; (3) Based on the constructed data model, different types of data sets are expected to be constructed; the data sets include text understanding data sets, intelligence-related data sets, decision point mining data sets, and auxiliary task planning data sets; (4) Based on the text understanding dataset, intelligence-related dataset, decision point mining dataset, and auxiliary task planning dataset, an intelligence vector database and an expected intelligence instruction fine-tuning dataset are constructed respectively. The intelligence vector database is updated in real time, and whenever there is the latest real intelligence data, it is stored in real time. The expected intelligence instruction fine-tuning dataset is screened, checked, and supplemented by military experts and stored in a data architecture that meets the instruction fine-tuning specifications. It is used to fine-tune the basic military large language model. The expected intelligence instruction fine-tuning dataset is updated periodically according to the fine-tuning needs. (5) Fine-tuning a large military model adapted to the intelligence field based on multimodal instruction fine-tuning technology and efficient parameter fine-tuning technology; the multimodal instruction fine-tuning technology is used to generate an intelligence instruction fine-tuning dataset based on prompt templates and experts through a large language, and then use the generated expected intelligence instruction fine-tuning dataset to fine-tune the instructions of the large military model for downstream tasks; the efficient parameter fine-tuning technology includes setting a model and a word segmenter and selecting a suitable parameter quantization model according to the current amount of fine-tuning data.

2. A method for understanding key intelligence requirements based on a large model according to claim 1, characterized in that: The intelligence instruction fine-tuning dataset is used for fine-tuning the military large language model to enhance the model's ability to understand and respond to specific types of military tasks. It is a high-quality fine-tuning dataset designed and generated based on specific military intelligence tasks, and contains task-related questions and answers, as well as specific scenarios, terms, and protocol elements. The intelligence vector database integrates historical intelligence mission data, including past inquiries, reports, and analysis results, and integrates new data generated by experts into the database to ensure the timeliness and relevance of information; The preset prompt word template library stores a plurality of templates designed in advance for different types of intelligence tasks to ensure the accuracy and efficiency of input.

3. A method for understanding key intelligence requirements based on a large model according to claim 2, characterized in that: The intelligence instruction fine-tuning dataset further includes a seed instruction set and a corresponding input and output instance set to generate a large amount of new instruction data and corresponding input and output instances, and then use the generated instruction dataset to fine-tune the instructions of the large model for downstream tasks.

4. A method for understanding key intelligence requirements based on a large model according to claim 3, characterized in that: The data collection and screening to obtain intelligence data with different data sources and data modalities includes: Collecting raw intelligence data: Collecting raw data from one or more data sources including open source intelligence, human intelligence, signals intelligence, geospatial intelligence, and technical intelligence; Screening of raw intelligence data: Performing data cleaning, data verification, data classification and labeling, data analysis, data fusion, and confidentiality and ethical processing on raw intelligence data, collecting and screening key information from multi-source, multi-modal intelligence data as the data basis for large-scale model training and decision support.

5. A method for understanding key intelligence requirements based on a large model according to claim 4, characterized in that: The data model includes: Entity-relationship model (ERM) and text analysis model for open source intelligence; Graph database models and time-driven models for human intelligence; Time series models and network traffic analysis models for signals intelligence; Geographic Information System (GIS) model and multidimensional data model (OLAP) for geospatial intelligence; Database models and association rule learning models for technical intelligence.

6. A method for understanding key intelligence requirements based on a large model according to claim 5, characterized in that: The text understanding data sets include: interactive instruction understanding data sets and multimodal intelligence understanding data sets; intelligence-related data sets include: intelligence report data sets, operational intelligence requirements understanding data sets and intelligence task planning and decomposition data sets; decision point mining data sets include: decision point case data sets and scenario simulation data sets; auxiliary task planning data sets include: command and control communication data sets and tactics and strategy guidance data sets.

7. A method for understanding key intelligence requirements based on a large model according to claim 6, characterized in that: S12, based on the fine-tuning model, based on the prompt word engineering technology, construct prompt word templates for specific fields, so that the fine-tuning model can meet the needs of different intelligence tasks, including: constructing prompt word templates for different intelligence task scenarios; different intelligence task scenarios include: (1) Mission scenarios for intelligence data retrieval, processing, and application; (2) Task decision point mining process task scenario; (3) mission scenarios for the key information requirements generation process at mission decision points; and (4) The branches driven by the decision point and the task scenarios for formulating subsequent plans.

8. A method for understanding key intelligence requirements based on a large model according to claim 7, characterized in that: The prompt word template uses a generalized prompt word template as its basis. An effective prompt contains one or more of the following five elements: role, instruction, context information, input data and output indication. The basic prompt word template is constructed based on historical data, and specific prompt word templates are designed based on the basic prompt word template for typical different intelligence task scenarios.

9. A method for understanding key intelligence requirements based on a large model according to claim 8, characterized in that: For intelligence data retrieval, processing and application process task scenarios, S12 includes: S121, determine the mapping relationship between each stage in combat intelligence preparation and the task scenario of intelligence data retrieval, processing and application process; the processing of intelligence data includes intelligence processing and intelligence analysis; for intelligence data retrieval, it corresponds to the intelligence plan formulation and guidance stage and the intelligence collection stage; for intelligence processing, it corresponds to the intelligence processing and utilization stage and the intelligence analysis and production stage; for intelligence analysis, it corresponds to the intelligence distribution and fusion stage and the intelligence evaluation and feedback stage; for intelligence application, it corresponds to the traditional document service and intelligence product provision stage and the content and knowledge output stage based on understanding user intentions and related concepts, including document information provision, professional analysis, processing and handling of intelligence products; S122, designing a prompt word template based on the mapping relationship; For the task decision point mining process task scenario, S12 includes: S121', determining the application scenario of the task scenario of the task decision point mining process; S122', designing a prompt word template based on the application scenario of the task scenario in the task decision point mining process and the prompt word technology of LLM; For the mission scenario of key information requirement generation process for mission decision points, S12 includes: S121”, determine the application scenarios of the task scenarios in the process of generating key information requirements for task decision points; S122”, based on the application scenarios of the task scenarios of the key information demand generation process for task decision points and the design of prompt word templates using natural language processing and deep learning technologies; For the branch driven by the decision point and the subsequent plan formulation task scenario, S12 includes: S121”’, determine the application scenarios of the task scenarios for branches pulled by the decision points and subsequent planning; S122”’, design the prompt word template based on the application scenario of the branch driven by the decision point and the formulation task scenario of the subsequent plan.

10. A key intelligence requirement intention understanding system based on a large model, used to implement any method of claims 1-9, comprising: Model building module, used to build a large model for understanding key intelligence requirements and intentions; The key intelligence demand intention understanding model includes interactive instruction understanding algorithm, combat plan understanding algorithm, contextual prompt learning technology, heterogeneous intelligence association fusion algorithm and multimodal intelligence understanding algorithm; The key intelligence requirement intention understanding module is used to understand the intention of key intelligence requirements based on the key intelligence requirement intention understanding large model; it is used for intelligence-based decision-making, which includes associating commanders' key information requirements generation, mining mission decision points, and specifying traction branches and subsequent plans.

11. An electronic device, characterized in that: The method comprises a processor and a memory, wherein the memory stores a plurality of instructions, and the processor is used for reading the instructions and executing any one of the methods according to claims 1-9.

12. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a plurality of instructions, and the plurality of instructions can be read by a processor to execute the method according to any one of claims 1 to 9.

Images