Aviation text content cleaning and labeling method, system and equipment and medium

Abstract

The invention relates to the technical field of aeronautical text data processing, and discloses an aeronautical text content cleaning and labeling method, system, device and medium wherein the method comprises: noise filtering: identifying and removing noise in an aeronautical text in combination with static cleaning and a general large model; format standardization: converting the aviation text after noise removal into a standardized format text; duplicate removal and error correction: detecting duplicate contents based on a hash algorithm, and correcting spelling errors and grammar errors based on a general large model to obtain an aviation text subjected to duplicate removal and error correction; entity identification: key entities are extracted based on the general large model, and the extracted key entities are labeled; active learning: screening high-value samples in the marked key entities based on an uncertainty query strategy; and dynamic optimization: performing verification and iterative optimization on the marking result in combination with the aviation knowledge base. According to the method, the automation level, the labeling accuracy and the system self-adaptive capability of aviation text processing can be remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of aviation text data processing technology, and in particular to a method, system, device and medium for cleaning and annotating aviation text content. Background Technology

[0002] Aviation text data plays a crucial role in the development of the low-altitude economy, but its processing faces numerous technical challenges. First, aviation text data is highly complex: flight logs, maintenance manuals, airworthiness regulations, safety reports, and other text types are diverse, often existing in multi-source, heterogeneous formats such as PDF, XML, and sensor logs. This data contains a large amount of technical terminology (such as "APU malfunction" and "ECAM warning"), unstructured content, and implicit engineering logic, making it difficult for traditional data processing tools to parse efficiently. For example, existing methods based on rule engines or shallow natural language processing models often fail to accurately extract key information due to a lack of deep understanding of aviation semantics, and may even lead to misjudgments.

[0003] Secondly, the cleaning and annotation of aviation text data is inefficient. Existing technologies rely on manual annotation, which requires domain experts to review each text line, a process that is time-consuming and costly. Furthermore, the accuracy of manual annotation is still difficult to guarantee, and the same event is often labeled differently due to variations in the understanding of different individuals.

[0004] General-purpose large models (such as the GPT series) suffer from significant limitations in their professional adaptability in the aviation field. Because their training data primarily originates from general internet corpora, they lack in-depth integration with authoritative knowledge such as airworthiness regulations and aircraft manuals, leading to a susceptibility to factual errors in specialized annotation tasks. For example, when parsing the critical parameter of "flap angle threshold," the model might generate an incorrect annotation of "30°" based on statistical probability, while actual airworthiness standards (such as CCAR-25 or FAR-121) explicitly limit this value to 25°. Summary of the Invention

[0005] To address the aforementioned issues, this invention proposes a method, system, device, and medium for cleaning and annotating aviation text content. Through a large-scale model architecture enhanced with vertical domain knowledge and a closed-loop mechanism of active learning and manual verification, it can solve problems such as high heterogeneity of aviation text data, complex semantics of professional terms, lack of professional adaptability, and high risk of illusion in existing technologies.

[0006] The technical solution adopted in this invention is as follows: A method for cleaning and annotating aviation text content includes: Noise filtering: Combining static cleaning with a general large model, noise is identified and removed from aviation text; Format standardization: Converting noise-removed aviation text into standardized format text; Deduplication and error correction: Duplicate content in the standardized format text is detected based on a hash algorithm, and spelling and grammatical errors are corrected based on a general large model to obtain the deduplicated and error-corrected aviation text; Entity recognition: Key entities are extracted from the deduplicated and error-corrected aviation text based on a general large model, and the extracted key entities are labeled. Active learning: Filtering high-value samples from labeled key entities based on uncertainty query strategies; Dynamic optimization: The annotation results are verified and iteratively optimized by combining the aviation knowledge base.

[0007] Furthermore, the noise filtering specifically includes: The known noise is matched and removed using regular expressions. The known noise includes HTML tags, special characters, broken newline characters, redundant spaces, tabs, page breaks, and communication check codes. The system calls a general large model to perform context-aware analysis, determines whether word breaks in the text need to be merged, identifies non-standard symbols and performs standardized replacements; at the same time, it selectively retains or transforms chart references according to the needs of downstream tasks, and outputs a denoised clean text sequence.

[0008] Furthermore, the format standardization includes employing a multimodal parsing engine and performing differential processing for different original formats: For PDF documents, integrate OCR and layout analysis technologies to distinguish between main text, tables, and footnotes; For flight data records in XML or JSON format, extract key fields and convert them into semantically coherent natural language descriptions; For time-series data in sensor logs, readable text summaries are generated while preserving structured attributes; Ultimately, all aviation texts will be presented as standardized texts in natural language narrative form.

[0009] Furthermore, the deduplication and error correction specifically include: The approximate duplicate detection technology based on SimHash identifies paragraphs with similar content in a large-scale text set. Based on a general large model, when a spelling error is detected, candidate words are matched based on Levenshtein distance and the correct word is determined by combining the context semantics; when a grammatical error is detected, multiple correction candidates are generated by the general large model and the optimal correction result is selected by confidence score.

[0010] Furthermore, the entity recognition includes sequence labeling using a BERT-BiLSTM-CRF fusion model: The text context embedding is generated by the BERT encoder, the long-distance dependencies of the text are captured by BiLSTM, and the transfer rules of entity labels are modeled by the CRF layer. Key entities are identified and classified through multi-granularity entity annotation, and semantic relationships between key entities are established; the key entities include aircraft model, component name, parameters, fault codes and operation actions.

[0011] Furthermore, the active learning specifically includes: taking samples with a predicted confidence level lower than a preset value during entity recognition as input, and filtering high-value samples based on an uncertainty query strategy, wherein the uncertainty query strategy includes a minimum confidence strategy and an entropy strategy.

[0012] Furthermore, the dynamic optimization specifically includes: The system takes the initial annotation results generated by entity recognition and the updated aviation knowledge source as input, and performs consistency verification on the initial annotation results through the aviation knowledge graph. If a conflict is found between the annotation results and the aviation knowledge source, an alarm is triggered and the review process is initiated. For new knowledge in the aviation knowledge source, key information is automatically extracted and the aviation knowledge graph is updated. At the same time, the entity recognition model is fine-tuned online by combining active learning feedback to achieve dynamic iteration of the annotation logic.

[0013] An aviation text content cleaning and annotation system, comprising: The noise filtering module is configured to combine static cleaning with a general large model to identify and remove noise in aviation text; The format standardization module is configured to convert noise-removed aeronautical text into standardized format text. The deduplication and error correction module is configured to detect duplicate content in the standardized format text based on a hash algorithm, and correct spelling and grammatical errors based on a general large model to obtain the deduplicated and error-corrected aviation text. The entity recognition module is configured to extract key entities from the deduplicated and error-corrected aviation text based on a general large model, and to annotate the extracted key entities. The active learning module is configured to filter high-value samples from labeled key entities based on an uncertainty query strategy; The dynamic optimization module is configured to verify and iteratively optimize the annotation results by combining the aviation knowledge base.

[0014] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the aviation text content cleaning and annotation method.

[0015] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the aviation text content cleaning and annotation method.

[0016] The beneficial effects of this invention are as follows: 1. This invention performs noise filtering, format standardization, and deduplication on multi-source heterogeneous aviation text; utilizes a large model for entity recognition, classification labeling, and semantic consistency verification; combines an active learning strategy to select high-value samples and optimize model parameters; and constructs a distributed system to achieve efficient scheduling and load balancing of the cleaning and labeling tasks. This invention covers data preprocessing, large model inference, active learning, result verification, and user interaction modules, solving core problems in aviation text data cleaning and labeling such as data dispersion, noise interference, low labeling efficiency, and poor system scalability, significantly improving cleaning efficiency and labeling accuracy, and forming an intelligent closed loop for aviation text data processing.

[0017] 2. This invention integrates the deep semantic understanding capabilities of large-scale models with aviation expertise to construct a fully intelligent processing framework encompassing noise filtering, format standardization, deduplication and error correction, entity recognition, active learning, and dynamic optimization. This effectively addresses key technical challenges in traditional methods for processing multi-source heterogeneous aviation text, such as incomplete text cleaning, low annotation accuracy, high reliance on manual intervention, and delayed knowledge updates. This invention significantly improves the automation level, annotation accuracy, and system adaptability of aviation text processing, providing high-quality, traceable, and scalable data support for key applications in the low-altitude economy, such as flight safety monitoring, intelligent operation and maintenance decision-making, and airworthiness compliance review, thus promoting the transformation of aviation big data into highly reliable knowledge assets. Attached Figure Description

[0018] Figure 1 This is a flowchart of an aviation text content cleaning and annotation method according to Embodiment 1 of the present invention.

[0019] Figure 2 This is a detailed flowchart of an aviation text content cleaning and annotation method according to Embodiment 1 of the present invention. Detailed Implementation

[0020] To provide a clearer understanding of the technical features, objectives, and effects of the present invention, specific embodiments are now described. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention; that is, the described embodiments are only a part of the embodiments of the invention, not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0021] Example 1 like Figure 1 and Figure 2 As shown, this embodiment provides a method for cleaning and annotating aviation text content, including the following steps: S1. Noise Filtering: By combining static cleaning with large models, noise in aviation text is identified and removed, such as HTML tags, special symbols, redundant spaces, etc.

[0022] S2. Format standardization: Convert multi-source heterogeneous data into a standardized text format.

[0023] S3. Deduplication and Error Correction: Detects duplicate content based on hash algorithms and corrects spelling and grammatical errors using a large model.

[0024] S4. Entity Recognition: Key entities are extracted from aviation text using a large model, and entity annotation is performed using models such as Conditional Random Field or BERT-BiLSTM-CRF.

[0025] S5. Active learning: High-value samples are selected based on uncertainty query strategies (such as minimum confidence and entropy value) to reduce the cost of manual annotation.

[0026] S6. Dynamic optimization: Verify and iteratively optimize the annotation results by combining aviation knowledge bases (such as airworthiness regulations and maintenance manuals).

[0027] It should be noted that the data cleaned in step S1 is raw aeronautical text data, including but not limited to PDF converted text, HTML web page scraping content, XML / JSON structured logs, and raw sensor log streams, which contain noise such as HTML tags, special characters, broken line breaks, redundant spaces, tabs, page breaks, and communication checksums.

[0028] Preferably, the noise filtering in step S1 adopts a dual-channel mechanism of "static rules + large-scale model semantic understanding": firstly, known noise is matched and removed using regular expressions; then, a large-scale model is invoked to perform context-aware analysis, determine whether word breaks need to be merged, identify non-standard symbols and perform standardized replacements, and selectively retain or transform chart references according to the needs of downstream tasks. The output is a denoised clean text sequence that retains the original semantic integrity, providing high-quality input for the processing in step S2.

[0029] It should be noted that the input in step S2 is multi-source heterogeneous text after noise filtering, including PDF text streams (including layout information), XML / JSON data packets, free text logs, etc.

[0030] Preferably, the format standardization module in step S2 employs a multimodal parsing engine to perform differentiated processing for different original formats: for PDF documents, it integrates OCR and layout analysis technologies to distinguish between main text, tables, and footnotes; for flight data records in XML or JSON format, it extracts key fields and converts them into semantically coherent natural language descriptions; for time-series data in sensor logs, it preserves their structured attributes while generating readable text summaries. This process ensures that data from heterogeneous sources such as flight logs, airworthiness documents, and maintenance reports can be unified into standardized text with clear semantics and consistent structure. All data is presented in natural language narrative form, providing a unified input interface for subsequent processing.

[0031] It should be noted that the input in step S3 is a standardized text set, which may contain issues such as duplicate reports, spelling errors, and grammatical errors.

[0032] Preferably, the deduplication stage in step S3 employs an approximate duplicate detection technique based on SimHash to quickly identify paragraphs with similar content in a large-scale dataset. The error correction stage uses a large model. When a spelling error is detected, the model performs candidate word matching based on Levenshtein distance and combines this with contextual semantics to determine the correct word. For grammatical errors such as subject-verb disagreement and tense confusion, the large model generates multiple correction candidates and selects the optimal result through confidence scoring, significantly outperforming the rule-matching methods of traditional grammar checking tools.

[0033] It should be noted that the input in step S4 is the normalized text after deduplication and error correction.

[0034] Preferably, step S4 employs a BERT-BiLSTM-CRF fusion model for sequence labeling. The BERT encoder generates contextual embeddings, BiLSTM captures long-range dependencies, and the CRF layer models label transfer rules. This supports multi-granularity labeling, identifies and classifies entities such as aircraft models, component names, parameters, fault codes, and operational actions, and establishes semantic associations.

[0035] It should be noted that the input in step S5 is the set of samples with low prediction confidence during the entity recognition process.

[0036] Preferably, step S5 aims to maximize the utilization efficiency of labeled resources. The most difficult samples to judge are selected by calculating the minimum confidence level predicted by the model and submitted for manual review first.

[0037] It should be noted that the input in step S6 is the preliminary annotation result generated by entity recognition, as well as the updated aerospace knowledge source from an external source.

[0038] Preferably, step S6 uses an aviation knowledge graph for consistency verification, automatically comparing whether the labeled output conforms to engineering standards. Upon detecting a conflict, an alarm is triggered and a review process is initiated. For newly added knowledge, key information is automatically extracted and the knowledge graph is updated. Combined with active learning feedback, the model undergoes online fine-tuning, achieving dynamic iteration of the labeling logic.

[0039] Example 2 This embodiment is based on embodiment 1: This embodiment provides a method for cleaning and annotating aviation text content, including the following steps: Step 1: Noise Filtering: By combining static rules with a large model, noise such as HTML tags, special symbols, and redundant spaces is identified and removed; basic noise patterns are matched using regular expressions, and semantic-level repair is achieved by calling the large model to obtain a clean text sequence. Step 2, Format Standardization: Convert multi-source heterogeneous data into a standardized text format; integrate OCR, layout analysis and natural language generation technologies to achieve a unified mapping of PDF, XML, JSON and other formats to semantically coherent text; Step 3, Deduplication and Error Correction: Duplicate content is detected based on hash algorithms, and spelling and grammatical errors are corrected using a large model; SimHash approximate deduplication and semantic error correction models are combined to improve text consistency; Step 4, Entity Recognition: Key entities are extracted from aviation text using a large model, and entity annotation is performed using the BERT-BiLSTM-CRF model; high-precision fine-grained annotation is achieved by integrating vertical semantics and sequence modeling capabilities. Step 5, Active Learning: Screen high-value samples based on uncertainty query strategies to reduce manual annotation costs; measure model uncertainty by minimum confidence and prediction entropy, and prioritize the submission of difficult samples for review. Step 6, Dynamic Optimization: Verify and iteratively optimize the annotation results by combining the aviation knowledge base; ensure compliance by comparing entities with standard values ​​through the knowledge graph.

[0040] Preferably, Step 1 in this embodiment includes: Step 1.1: Obtain raw, multi-source, heterogeneous aviation text data as input; Raw text data was collected from the airline's operations and maintenance system, including PDF-converted text, HTML webpage logs, XML sensor records, and maintenance reports in free text format. The raw data is as follows: B737-800 Left-hand N2 over-limit (Code: ENG_N2_HIGH), immediate inspection required! Timestamp: 2025-05-20T14:30:00 Step 1.2: Apply regular expressions to remove known structured noise. Use preset regular rules to remove HTML tags, redundant line breaks, and verification marks; re.sub(r'<[^>]+>','', text) removes HTML tags; The intermediate result output is: The N2 of the left engine of B737-800 is overlimit (Code: ENG_N2_HIGH), immediate inspection required! Timestamp: 2025-05-20T14:30:00 Step1.3. Invoke the large model for semantic-level noise repair: Input the intermediate text into the large model to identify and repair potential semantic breaks or non-standard expressions; the model determines that "left engine" is the industry abbreviation of "left engine" and retains it; it identifies "Code: ENG_N2_HIGH" as a standard fault code and retains the content in parentheses; The output of the semantically complete text is: The N2 of the left engine of B737-800 is overlimit (fault code: ENG_N2_HIGH), immediate inspection required! Time: 2025-05-20T14:30:00.

[0041] Preferably, in this embodiment, Step2 includes: Step2.1. Perform OCR and layout analysis on the PDF document; use the OCR engine to extract the text in the PDF maintenance manual image, and combine layout analysis to distinguish the main text, table, and footnote content: An example table is shown in Table 1: Table 1 - Example Table

[0042] Step2.2. Convert structured data (XML / JSON) into natural language descriptions: The XML data is as follows: <event type="engine_alert" param="N2" value="98.5" unit="%" timestamp="2025-05-20T14:30:00" / > Invoke the large model to generate descriptive sentences and unify the expression of time, units, and terms: The N2 speed of the left engine of B737-800 reached 98.5 revolutions per minute at 2025-05-20T14:30:00, exceeding the 97% threshold and triggering the fault code ENG_N2_HIGH.

[0043] Preferably, in this embodiment, Step3 includes: Step3.1. Use the SimHash algorithm to deduplicate the text set: Calculate the SimHash values of multiple reports. If it is less than 3, it is considered a duplicate; merge three reports with similar content and only retain the version with the most complete semantics.

[0044] Step 3.2: Check and correct spelling errors: Input text: “B737-800 left-hand N2 overline”; The model identifies “overline” as a spelling error “overlimit”; The corrected version is: "B737-800 left-hand N2 exceeds limit".

[0045] Step 3.3: Correct syntax errors. Input: "N2 speed high and engine not start"; The large model generates the corrected sentence: "N2 speed is too high and the engine cannot start."

[0046] Step 3.4: Output cleaned text that conforms to language standards and is free of duplicates: The left engine of B737-800 reached 98.5 rpm at 2025-05-20T14:30:00, exceeding the 97% threshold and triggering the fault code ENG_N2_HIGH.

[0047] Preferably, Step 4 in this embodiment includes: Step 4.1: Load the pre-trained aviation domain BERT-BiLSTM-CRF model: supports five types of entities: aircraft type, component, parameter, fault type, and fault code.

[0048] Step 4.2: Perform sequence labeling on the cleaned text: The model outputs the IOB tag sequence.

[0049] Step 4.3: Parse the model output and extract structured entities: Output the annotation results in JSON format.

[0050] Step 4.4: Send the annotation results into the active learning module for confidence evaluation. Since the confidence level is 0.83 < 0.85, proceed to the S5 process.

[0051] Preferably, Step 5 in this embodiment includes: Step 5.1 Calculate the minimum confidence score for each sample: Take the maximum value from the probability distribution output by the model as the confidence score.

[0052] Step 5.2: Calculate the prediction entropy to assess the overall uncertainty. Entropy If the entropy is greater than 0.65, it is considered to have high uncertainty.

[0053] Step 5.3: Filter samples with confidence scores below the threshold or entropy values ​​above the threshold; in this embodiment, a confidence score of 0.83 < 0.85 triggers the review.

[0054] Step 5.4: Push high-value samples to the manual review platform to obtain expert annotation feedback; engineers confirm the annotation is correct and add it to the database.

[0055] Preferably, Step 6 in this embodiment includes: Step 6.1: Load the aviation knowledge base (including CCAR-25, aircraft type manuals, etc.): Query the standard description and threshold corresponding to "ENG_N2_HIGH".

[0056] Step 6.2, Perform consistency verification between annotation results and knowledge base: verify whether the threshold of "N2 speed" is 97% (compliant); verify whether "B737-800" supports the fault code (compliant).

[0057] Step 6.3: Initiate review and supplementation for conflicting or missing items: If "N2=105%" is found but the standard is 97%, an alarm will be triggered.

[0058] Step 6.4: Add the approved samples to the training set to trigger incremental learning of the model: update the model parameters and improve the accuracy of subsequent predictions.

[0059] Example 3 This embodiment provides an aviation text content cleaning and annotation system, including: The noise filtering module is configured to combine static cleaning with a general large model to identify and remove noise in aviation text; The format standardization module is configured to convert noise-removed aeronautical text into standardized format text. The deduplication and error correction module is configured to detect duplicate content in the standardized format text based on a hash algorithm, and correct spelling and grammatical errors based on a general large model to obtain the deduplicated and error-corrected aviation text. The entity recognition module is configured to extract key entities from the deduplicated and error-corrected aviation text based on a general large model, and to annotate the extracted key entities. The active learning module is configured to filter high-value samples from labeled key entities based on an uncertainty query strategy; The dynamic optimization module is configured to verify and iteratively optimize the annotation results by combining the aviation knowledge base.

[0060] Example 4 This embodiment is based on embodiment 1: This embodiment provides a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the aviation text content cleaning and annotation method of Embodiment 1. The computer program can be in the form of source code, object code, executable file, or some intermediate form.

[0061] Example 5 This embodiment is based on embodiment 1: This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aerospace text content cleaning and annotation method of Embodiment 1. The computer program can be in the form of source code, object code, executable file, or some intermediate form. The storage medium includes any entity or device capable of carrying computer program code, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc. It should be noted that the content contained in the storage medium can be appropriately added to or subtracted according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, the storage medium does not include electrical carrier signals and telecommunication signals.

[0062] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

[0063] It should be noted that, for the sake of simplicity, the foregoing method embodiments are described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

Claims

1. An aviation text content cleaning and labeling method, characterized in that, The method comprises the following steps: Noise filtering: combine static cleaning with general large models to identify and remove noise in aviation texts; Format standardization: convert the noise-removed aviation texts into standardized format texts; De-duplication and error correction: detect duplicate content in the standardized format texts based on a hash algorithm, and correct spelling and grammar errors based on a general large model to obtain de-duplicated and error-corrected aviation texts; Entity recognition: extract key entities from the de-duplicated and error-corrected aviation texts based on a general large model, and label the extracted key entities; Active learning: select high-value samples from the labeled key entities based on an uncertainty query strategy; Dynamic optimization: verify and iteratively optimize the labeling results in combination with an aviation knowledge base.

2. The method of claim 1, wherein, The noise filtering specifically includes: Match and remove known noise through regular expressions, including HTML tags, special symbols, line break breaks, redundant spaces, tabs, page breaks, and communication check codes; Call a general large model for context-aware analysis to determine whether to merge word breaks, identify non-standard symbols and perform standardized replacement; at the same time, selectively retain or convert chart references according to the requirements of downstream tasks, and output pure text sequences after noise removal.

3. The method of claim 1, wherein, The format standardization includes using a multi-modal parsing engine and performing differential processing for different original formats: For PDF documents, integrate OCR and layout analysis techniques to distinguish between text, tables, and footnote content; For flight data records in XML or JSON format, extract key fields and convert them into semantically coherent natural language descriptions; For time series data in sensor logs, retain structured attributes while generating readable text summaries; Finally, all aviation texts are presented as standardized texts in natural language narrative form.

4. The method of claim 1, wherein, The de-duplication and error correction specifically includes: Approximate duplicate detection technology based on SimHash to identify similar content in large-scale texts; Based on the general large model, when a spelling error is detected, perform candidate word matching based on Levenshtein distance and determine the correct word based on context semantics; when a grammar error is detected, the general large model generates multiple correction candidates and selects the optimal correction result through confidence scoring.

5. The method of claim 1, wherein, The entity recognition includes using a BERT-BiLSTM-CRF fusion model for sequence labeling: Generate context embeddings of the text through the BERT encoder, capture long-distance dependencies of the text through BiLSTM, and model the transition rules of entity labels through the CRF layer; Identify and classify key entities through multi-granularity entity labeling, and establish semantic associations between key entities; the key entities include aircraft models, component names, parameters, fault codes, and operation actions.

6. The method of claim 1, wherein, The active learning specifically includes: taking samples with a prediction confidence lower than a preset value in the entity recognition process as input, and selecting high-value samples based on an uncertainty query strategy, which includes a minimum confidence strategy and an entropy strategy.

7. The method of claim 1, wherein, The dynamic optimization specifically includes: The preliminary labeling result generated by entity recognition and the externally updated aviation knowledge source are input, and the consistency of the preliminary labeling result is checked by the aviation knowledge graph; if the labeling result conflicts with the aviation knowledge source, an alarm is triggered and a review process is started; For the newly added knowledge in the aviation knowledge source, the key information is automatically extracted and the aviation knowledge graph is updated, and the entity recognition model is fine-tuned online combined with active learning feedback, realizing the dynamic iteration of the labeling logic.

8. An aerial text content cleaning annotation system, comprising: It comprises: A noise filtering module configured to identify and remove noise in aviation text in combination with static cleaning and general large models; A format standardization module configured to convert the aviation text after removing noise into a standardized format text; A deduplication and error correction module configured to detect duplicate content in the standardized format text based on a hash algorithm, and correct spelling and grammar errors based on a general large model to obtain aviation text after deduplication and error correction; An entity recognition module configured to extract key entities from the aviation text after deduplication and error correction based on a general large model, and label the extracted key entities; An active learning module configured to screen high-value samples from the labeled key entities based on an uncertainty query strategy; A dynamic optimization module configured to verify and iteratively optimize the labeling result in combination with an aviation knowledge base. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-8 when the computer program is executed by the processor. The processor executes the computer program to realize the aviation text content cleaning and labeling method of any one of claims 1-7.

10. A computer readable storage medium storing a computer program, characterized in that, The computer program is executed by the processor to realize the aviation text content cleaning and labeling method of any one of claims 1-7.

Images