An unmanned aerial vehicle knowledge text database corpus automatic labeling system and method based on a large language model

Abstract

This invention discloses an automatic annotation system and method for a UAV knowledge text database corpus based on a large language model. The system includes: a question sample generation module, used to intelligently segment the input intelligence knowledge text into multiple text blocks, and generate question samples related to the content of each text block based on the text content and a large language model; a tag generation module based on retrieval enhancement generation, used to receive the multiple text blocks and corresponding question samples output by the question sample generation module, retrieve relevant information from the question samples using retrieval enhancement generation technology, and generate corresponding answers as tags to annotate the text blocks; and a module communication and scheduling unit, used to control the data flow between the question sample generation module and the tag generation module. The automatic annotation system based on a large language model used in this invention significantly improves annotation efficiency.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, and in particular to an automatic annotation system and annotation method for UAV knowledge text database corpus based on a large language model. Background Technology

[0002] Currently, large language models excel in natural language understanding and processing. However, in specialized fields, there are often massive and complex amounts of knowledge data, which are difficult for general practitioners to fully grasp. Therefore, it is necessary to leverage large language model question-answering systems to empower professional analysis processes, thereby improving the efficiency of information processing and decision support. However, most current open-source or closed-source large models are not trained using such specialized and domain-specific data, which may lead to inaccurate information or unreliable inferences in related tasks. To address this issue, it is necessary to conduct specialized training and fine-tuning of large models, constructing their knowledge system within the relevant domain, thereby enhancing the reliability and applicability of the output.

[0003] Natural Language Processing (NLP), as an important branch of artificial intelligence, aims to enable computers to understand and generate human natural language. With the rapid development of internet technology, a vast amount of text data has emerged, providing abundant resources for the application of NLP technology. However, for computers to accurately understand and process this text data, effective data annotation is necessary to provide a basis for training and evaluating NLP models.

[0004] Corpus annotation technology has a long history in the field of natural language processing, and its development has gone through several stages, from simple keyword annotation to complex semantic role annotation and sentiment annotation. Early annotation work mainly relied on manual labor, with annotators classifying, annotating, and marking words, sentences, or passages in the text according to predetermined annotation standards. While this traditional annotation method could guarantee a certain level of annotation quality, it faced problems such as low efficiency, high cost, and unstable annotation quality.

[0005] With the rise of machine learning and deep learning technologies, the demand for labeled corpora in natural language processing models has exploded. To meet this demand, researchers have begun exploring automatic annotation techniques, aiming to automate some or all of the annotation work through algorithms. Automatic annotation techniques typically rely on pre-trained language models, which, through pre-training on massive amounts of text, can capture rich linguistic features and patterns, thus playing a crucial role in annotation tasks.

[0006] However, existing automatic annotation technologies still have certain limitations, such as the need to improve annotation accuracy and poor adaptability to specific domain corpora, such as equipment knowledge. Furthermore, achieving rapid annotation of large-scale corpora while ensuring annotation quality remains a major challenge in the field of natural language processing.

[0007] Therefore, there is an urgent need to provide a solution for an automatic annotation system and annotation method for UAV knowledge text database corpora based on large language models. Summary of the Invention

[0008] To address the above issues, this invention provides an automatic annotation system and method for UAV knowledge text database corpora based on a large language model, enabling rapid understanding of natural language corpora. Simultaneously, the system can perform precise analysis and automatically annotate the corpora, significantly reducing the need for manual intervention.

[0009] According to a first aspect of the present invention, an automatic annotation system for UAV knowledge text database corpus based on a large language model is provided, comprising: The question sample generation module is used to intelligently segment the input intelligence knowledge text into multiple text blocks, and generate question samples related to the content of each text block based on the text content and a large language model. The tag generation module based on retrieval enhancement generation is used to receive multiple text blocks and corresponding question samples output by the question sample generation module, retrieve relevant information from the question samples through retrieval enhancement generation technology and generate corresponding answers as tags to achieve annotation of the text blocks; The module communication and scheduling unit is used to control the data flow between the problem sample generation module and the tag generation module.

[0010] In the above scheme, the problem sample generation module includes: The intelligent text segmentation unit is used to segment the input text. It first segments the text at paragraph and sentence boundaries, and the segment length is dynamically adjusted within the range of 500 to 4000 tags. The question generation unit is used to guide the large language model to reason about each text block by combining prompt word engineering with mind chain technology, and generate questions that are related to the text block and meet the domain professional requirements. The question verification unit is used to perform relevance detection, semantic deduplication, and difficulty classification on the generated questions.

[0011] In the above scheme, the intelligent text segmentation unit is implemented by combining the recursive character segmentation method of the LangChain framework with a custom text segmenter; In the problem generation unit, the prompt word engineering, combined with the thinking chain technology, breaks down the task into multiple steps through the prompt word chain. The execution steps include the understanding stage, the planning stage, and the generation stage.

[0012] In the above scheme, the tag generation module based on retrieval enhancement includes: The retrieval unit is used to encode the text block into a vector, construct a vector index and add it to the vector database of the retrieval unit, encode the question sample into a query vector, and retrieve the most relevant text block from the vector database; The generation unit is used to integrate the retrieved text blocks with the question samples into a context with prompt words, and input it into a large language model to generate an answer as a label for the question samples.

[0013] In the above scheme, the generation unit also includes a reordering mechanism, which uses a cross encoder to perform a secondary fine sorting of the most relevant text blocks retrieved by the retrieval unit. The answer output by the generation unit includes source document block identifiers and model confidence information, and is stored in a structured format.

[0014] In the above scheme, the module communication and scheduling unit includes a single sequential processing mode and a multi-pipeline parallel processing mode.

[0015] In the above scheme, the system also includes a model management and configuration interface to support plug-and-play loading and parameter configuration of large language models, and provides a visual configuration interface.

[0016] In the above scheme, the system's operating mechanism includes batch processing and real-time processing; The batch processing mechanism uses a distributed task queue for offline batch labeling. The real-time processing mechanism receives processing requests through a RESTful interface.

[0017] In the above scheme, the model management and configuration interface encapsulates the calling differences of different large language models through a unified model adapter, allowing dynamic configuration of generation parameters. The generation parameters include temperature coefficient, Top-p sampling, maximum generation length, whether to enable thought chain, batch size, number of retries and timeout.

[0018] According to a second aspect of the present invention, an automatic annotation method for UAV knowledge text database corpora based on a large language model is provided, applied to the system described in any one of the above solutions, the method comprising the following steps: S1. Problem Sample Generation Steps: The input intelligence knowledge text is intelligently segmented into multiple text blocks, and problem samples related to the content of each text block are generated based on a large language model. S2, Tag Generation Step: Receive the text block and corresponding question sample output from the question sample generation step, retrieve relevant information from the question sample using retrieval enhancement generation technology and generate corresponding answers as tags to label the text block; S3. Process control steps: Control the data flow between the problem sample generation step and the label generation step.

[0019] The beneficial effects of this invention are: 1) High Efficiency: The automatic annotation system based on a large language model used in this invention significantly improves annotation efficiency compared to traditional manual annotation methods. The system can process large amounts of data in a short time, enabling rapid annotation of massive corpora, greatly shortening project cycles, and meeting the needs of real-time and large-scale data processing.

[0020] 2) Cost Savings: This invention reduces manual intervention through automation technology, significantly lowering annotation costs. It eliminates the need to hire a large number of professional annotation personnel, reducing salary and welfare expenses. Simultaneously, it reduces training and supervision costs, making the overall annotation cost more economical.

[0021] 3) Quality Stability: This invention leverages the stability and consistency of large language models to effectively reduce individual differences and errors caused by manual annotation. The annotation results output by the automatic annotation system exhibit high consistency, improving the accuracy and reliability of large language model training.

[0022] 4) High level of intelligence: This invention combines deep learning and artificial intelligence technologies, making the annotation process more intelligent. The system can automatically learn and optimize annotation rules, adapt to different types of corpora and annotation needs, and improve the system's adaptability and generalization ability.

[0023] 5) Easy to expand and upgrade: The system of this invention is built on a large language model and has good scalability. As the model is continuously optimized and upgraded, the system's annotation capabilities will also be improved, enabling it to adapt to the ever-changing technological environment and application needs. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0025] Figure 1 This is a schematic diagram of the overall system of the present invention; Figure 2 This is a schematic diagram of the problem sample generation module of the present invention; Figure 3 This is a schematic diagram of the tag generation module based on search enhancement according to the present invention.

[0026] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0027] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this disclosure as detailed in the appended claims.

[0028] The terms "first," "second," etc., used in this disclosure are for distinguishing similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such use of data can be interchanged where appropriate so that the embodiments of this disclosure described herein can be implemented, for example, in orders other than those illustrated or described herein.

[0029] Furthermore, the terms “comprising” and “having”, and any variations thereof, are intended to cover non-exclusive inclusion, such that a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such process, method, product, or apparatus.

[0030] Multiple, including two or more.

[0031] And / or, it should be understood that, for the purposes of this disclosure, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent three situations: A exists alone, A and B exist simultaneously, and B exists alone.

[0032] like Figure 1 As shown, according to a first aspect of the present invention, an automatic annotation system for UAV knowledge text database corpus based on a large language model is provided, comprising: The question sample generation module is used to intelligently segment the input intelligence knowledge text into multiple text blocks, and generate question samples related to the content of each text block based on the text content and a large language model. The tag generation module based on retrieval enhancement is used to receive multiple text blocks and corresponding question samples output by the question sample generation module, retrieve relevant information from the question samples through retrieval enhancement generation technology and generate corresponding answers as tags to achieve annotation of text blocks; The module communication and scheduling unit is used to control the data flow between the problem sample generation module and the tag generation module.

[0033] like Figure 2 As shown, specifically, the problem sample generation module includes: The intelligent text segmentation unit is used to segment the input text. This invention employs a dynamic segmentation mechanism rather than fixed-length segmentation. It combines the LangChain framework's Recursive CharacterTextSplitter with a custom text segmenter to ensure the segmentation results conform to the document's semantic structure. The semantically complete recursive character segmentation method prioritizes segmentation at paragraph and sentence boundaries. The target length of each block is set to 2000 tokens by default, a value determined based on the following factors: 1) Cover the complete descriptive units of typical equipment technical documents (such as performance parameters and instructions for use of a single type of equipment); 2) Balance retrieval efficiency with contextual integrity, avoiding information fragmentation due to excessively small blocks or impact on retrieval accuracy due to excessively large blocks; 3) Adapt to the context window limitations of mainstream large language models (usually 4K-32K tokens), reserving enough space for prompt words and generated content; This invention can dynamically adjust the block length based on the text type, with a range of 500-4000 tokens. For highly structured equipment parameter tables, smaller blocks of 500-800 tokens can be used; for narrative-driven analyses, larger blocks of 2500-3000 tokens can be used. This dynamic adjustment method is well-suited to the mixed document types characteristic of the UAV field.

[0034] This invention abandons traditional fixed-length block segmentation and innovatively combines the recursive character segmentation method of the LangChain framework with a custom text segmenter. It prioritizes segmentation at natural boundaries such as paragraphs and sentences, ensuring that each text block maintains a complete semantic unit. Since the paragraph semantic integrity of UAV-related knowledge texts is strong, semantic fragmentation is avoided. Furthermore, it can dynamically adjust the block size according to the text type to balance information integrity and processing efficiency.

[0035] The question generation unit employs multi-layered prompt word engineering technology, combining Chain-of-Thought (CoT) and Prompt Chaining to guide the large language model in generating high-quality question samples that are relevant to the text blocks and meet domain-specific requirements.

[0036] Furthermore, during the design of the prompts, the model's role was clearly defined as an "equipment domain expert," enhancing the professionalism of the generated questions. The complex task was also broken down into multiple steps through a chain of prompts. In the first round (understanding phase), the prompts require the model to "deeply understand the following text content and identify key entities (equipment models, technical parameters, uses), relationships (cooperative relationships and adversarial relationships between equipment) and important concepts". In the second round (planning phase), the prompt words guide the model to "plan 5-8 core problem directions based on the above understanding, covering categories such as equipment identification, performance comparison, tactical application, technical principles, and maintenance"; In the third round (generation phase), the prompt word instruction model "generates 2-3 specific, clear, and answerable questions for each question direction. The questions should reflect professionalism and avoid vague expressions."

[0037] After generating the prompts, the thought chain mode is activated, adding the guiding text to the prompts: "Please think through and demonstrate your analysis process step by step, including: ① Extraction of core text information → ② Identification of key knowledge points → ③ Determination of question type → ④ Generation of specific questions." The model will output intermediate results containing a complete reasoning chain, facilitating subsequent verification and optimization.

[0038] This invention employs multi-level prompt word engineering technology combined with prompt word chains and thought chain technology to guide a large language model in generating in-depth, professional questions. Complex tasks are broken down into three stages: understanding, planning, and generation. The model's role is explicitly defined as an "equipment domain expert," ensuring the professionalism and comprehensiveness of the questions. Under the thought chain model, the model outputs a complete reasoning process, facilitating quality verification and prompt word optimization.

[0039] The question verification unit is used to perform relevance detection, semantic deduplication, and difficulty classification on the generated questions, as detailed below: 1) Relevance detection: Use embedding vector similarity to ensure that the semantic relevance between the question and the original text is greater than 0.85; 2) Duplicate detection: A semantic deduplication algorithm (based on Sentence-BERT embedding similarity calculation) is used to remove duplicates or highly similar issues; 3) Difficulty level: Automatically label problems with difficulty levels such as "basic cognition / technical analysis / comprehensive application" using small sample learning.

[0040] The problem validation unit removes duplicate problems, improving dataset diversity. Difficulty labels are added to support subsequent differentiated applications.

[0041] In the tag generation module based on retrieval augmented generation, retrieval augmented generation (RAG) is a technique that combines information retrieval and natural language generation. It aims to enhance the generation process through retrieval techniques, thereby improving the quality and relevance of the generated text. Furthermore, as... Figure 3 As shown, the tag generation module based on search enhancement includes: The retrieval unit utilizes RAG technology and information retrieval algorithms to retrieve information relevant to the user's query from a large number of data sources. Specifically, this involves: firstly, encoding the input text blocks using a large vector space; the encoding model employs a domain-adapted BERT variant (such as RoBERTa-wwm-ext or a self-developed domain-pre-trained model) as the text encoder, mapping document blocks and question samples to a high-dimensional vector space. The encoding dimension is set to 768 dimensions, achieving a balance between semantic expressiveness and storage efficiency.

[0042] The preprocessed documents are then indexed using appropriate indexing methods (such as inverted indexes) for fast retrieval. The vector database uses Faiss (Facebook AI Similarity Search) to build an efficient vector index. Specifically, it employs the IndexFlatIP (inner product similarity) or IndexIVFFlat (inverted document index) algorithm, supporting millisecond-level retrieval of millions of document blocks. The add() method is called to add vectors to the index and incremental index updates are performed periodically, supporting the dynamic integration of newly added intelligence data.

[0043] Next, the question sample is encoded into a query vector, and its similarity is calculated with the previous large vector space. Then, the document block that is closest to the query request is selected. Specifically, the Faiss search() method is called to calculate the cosine similarity with the document vector library and return the Top-K (default K=5) most relevant document blocks.

[0044] The generation unit first integrates the relevant documents retrieved by the retrieval unit with the question samples. Specifically, it sorts the retrieved Top-K document blocks by similarity and uses them together with the question samples to construct the context of the prompt words. This invention employs a re-ranking mechanism, using a cross-encoder to perform a secondary fine-grained ranking of the retrieval results, thereby improving relevance.

[0045] Then, the prompt words are combined into a complete context, the configured large language model is called to analyze the context, the structured prompt word template is used, the generated answer is validated in the format and used as the standard label (Ground Truth) of the question sample, and stored in JSON format, containing the fields: question, answer, source_chunks (source document chunk ID), and confidence_score (model confidence).

[0046] In this invention, the answers are generated based on retrieved evidence, resulting in high accuracy and strong verifiability. Furthermore, the use of JSON structured storage facilitates subsequent knowledge base construction and retrieval applications.

[0047] Furthermore, the modules of this invention communicate using a standard HTTP interface request format, supporting both sequential processing of single data lines and parallel processing of multiple data pipelines.

[0048] Specifically, all interactions between modules follow the RESTful API specification and are transmitted in JSON format.

[0049] Request format example: { "request_id": "uuid-1234", "text_chunk": "The text to be processed...", "chunk_id": "doc001_chunk005", "model_config": { "model_name": "qwen-72b-chat", "temperature": 0.3, "max_tokens": 1500 }, "task_priority": "normal" } Response format example: { "request_id": "uuid-1234", "status": "success", "generated_questions": [ { "question_id": "q001", "question_text": "What is the maximum cruising speed of a certain type of drone?" "difficulty_level": "basic knowledge", "confidence_score": 0.92 } ], "processing_time_ms": 1250 } The two modes of input / output flow are as follows: 1) Sequential processing: A single piece of data follows a serial process of "configuration module → problem generation module → tag generation module", and is traced throughout the entire chain through request_id.

[0050] 2) Parallel pipeline: During batch processing, each module independently listens to the message queue (such as Redis Stream or KafkaTopic), realizing loosely coupled pipeline parallelism, which improves throughput by 3-5 times.

[0051] Standardized interfaces enable loose coupling between modules, facilitating independent upgrades and expansions. Pipeline parallelism increases throughput by 3-5 times, adapting to processing needs of varying scales. Combined with Celery's distributed task queue and dynamic resource scheduling, it achieves efficient resource utilization and prioritizes urgent tasks.

[0052] Furthermore, for large-scale knowledge text databases, this invention supports offline batch annotation using a Celery distributed task queue. Data to be processed is added to the annotation task queue, supporting priority scheduling. The question generation and answer generation modules can form a processing pipeline. After module 1 completes question generation for a batch of text, module 2 can start generating answers to those questions in parallel, while module 1 continues processing the next batch of text. It also features breakpoint resumption and result persistence capabilities, writing the results of each batch of processing to a distributed file system (such as HDFS) or database in real time, supporting resumption after task interruption. The output path can be configured as local disk, network storage, or cloud storage.

[0053] For single or small amounts of newly added intelligence text, this invention employs real-time online annotation and dynamic resource scheduling. The system monitors GPU memory usage and model computing load, dynamically adjusting the number of data entries processed in real time and the processing time per cycle. When the GPU memory idle rate exceeds 30%, the batch size is automatically increased; when the load is too high, a flow control mechanism is triggered.

[0054] For urgent intelligence processing needs, FastAPI provides a RESTful interface to achieve second-level response (average latency <3 seconds); priority preemption allows real-time tasks to preempt low-priority task resources in the batch queue.

[0055] This invention supports the loading, initialization, and dynamic configuration of runtime parameters for pre-trained large models. It provides a visual model selection interface and supports plug-and-play calls to mainstream open-source and commercial large language models. A unified model adapter interface encapsulates the differences in calling different models, ensuring that upper-layer business logic does not need to be modified when switching models. It also supports fine-grained parameter configuration, allowing users to adjust key generation parameters according to task requirements, such as temperature coefficient, top-p sampling, maximum generation length (Max Tokens), whether to enable chain-of-thought, batch size, retry count, and timeout settings. All parameter configurations can be saved as preset schemes, supporting rapid switching for different types of intelligence texts (such as equipment parameters, tactical regulations, intelligence briefings, etc.).

[0056] This invention realizes a fully automated, intelligent, and highly reliable production closed loop from knowledge text to high-quality question-answering training corpus, significantly improving the efficiency and quality of building UAV knowledge text databases.

[0057] According to a second aspect of the present invention, an automatic annotation method for UAV knowledge text database corpus based on a large language model is provided, applied to the system described above. The method includes the following steps: S1. Problem Sample Generation Steps: The input intelligence knowledge text is intelligently segmented into multiple text blocks, and problem samples related to the content of each text block are generated based on a large language model. S2, Tag Generation Step: Receive the text block and corresponding question sample output from the question sample generation step, retrieve relevant information from the question sample using retrieval enhancement generation technology and generate corresponding answers as tags to label the text block; S3. Process control steps: Control the data flow between the problem sample generation step and the label generation step.

[0058] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0059] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0060] Through the above description of the embodiments, those skilled in the art can clearly understand that the above implementation methods can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0061] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. An automatic annotation system for UAV knowledge text database corpus based on a large language model, characterized in that, include: The question sample generation module is used to intelligently segment the input intelligence knowledge text into multiple text blocks, and generate question samples related to the content of each text block based on the text content and a large language model. The tag generation module based on retrieval enhancement generation is used to receive multiple text blocks and corresponding question samples output by the question sample generation module, retrieve relevant information from the question samples through retrieval enhancement generation technology and generate corresponding answers as tags to achieve annotation of the text blocks; The module communication and scheduling unit is used to control the data flow between the problem sample generation module and the tag generation module.

2. The system according to claim 1, characterized in that, The problem sample generation module includes: The intelligent text segmentation unit is used to segment the input text. It first segments the text at paragraph and sentence boundaries, and the segment length is dynamically adjusted within the range of 500 to 4000 tags. The question generation unit is used to guide the large language model to reason about each text block by combining prompt word engineering with mind chain technology, and generate questions that are related to the text block and meet the domain professional requirements. The question verification unit is used to perform relevance detection, semantic deduplication, and difficulty classification on the generated questions.

3. The system according to claim 2, characterized in that, The intelligent text segmentation unit is implemented by combining the recursive character segmentation method of the LangChain framework with a custom text segmenter. In the problem generation unit, the prompt word engineering, combined with the thinking chain technology, breaks down the task into multiple steps through the prompt word chain. The execution steps include the understanding stage, the planning stage, and the generation stage.

4. The system according to claim 1, characterized in that, The tag generation module based on search enhancement includes: The retrieval unit is used to encode the text block into a vector, construct a vector index and add it to the vector database of the retrieval unit, encode the question sample into a query vector, and retrieve the most relevant text block from the vector database; The generation unit is used to integrate the retrieved text blocks with the question samples into a context with prompt words, and input it into a large language model to generate an answer as a label for the question samples.

5. The system according to claim 4, characterized in that, The generation unit also includes a reordering mechanism, which uses a cross encoder to perform a secondary fine sorting of the most relevant text blocks retrieved by the retrieval unit. The answer output by the generation unit includes source document block identifiers and model confidence information, and is stored in a structured format.

6. The system according to claim 1, characterized in that, The module communication and scheduling unit includes a single sequential processing mode and a multi-pipeline parallel processing mode.

7. The system according to claim 1, characterized in that, The system also includes a model management and configuration interface to support plug-and-play loading and parameter configuration of large language models, and provides a visual configuration interface.

8. The system according to claim 1, characterized in that, The system's operating mechanism includes batch processing and real-time processing; The batch processing mechanism uses a distributed task queue for offline batch labeling. The real-time processing mechanism receives processing requests through a RESTful interface.

9. The system according to claim 1, characterized in that, The model management and configuration interface encapsulates the calling differences of different large language models through a unified model adapter, allowing dynamic configuration of generation parameters, including temperature coefficient, Top-p sampling, maximum generation length, whether to enable thought chain, batch size, number of retries and timeout.

10. An automatic annotation method for UAV knowledge text database corpus based on a large language model, characterized in that, Applied to the system of any one of claims 1-9, the method comprises the following steps: S1. Problem Sample Generation Steps: The input intelligence knowledge text is intelligently segmented into multiple text blocks, and problem samples related to the content of each text block are generated based on a large language model. S2, Tag Generation Step: Receive the text block and corresponding question sample output from the question sample generation step, retrieve relevant information from the question sample using retrieval enhancement generation technology and generate corresponding answers as tags to label the text block; S3. Process control steps: Control the data flow between the problem sample generation step and the label generation step.

Images