A method for dynamic identification of flight changes

Abstract

The present application relates to a method for dynamically identifying flight changes, and to the field of artificial intelligence technology. The method comprises: obtaining an airline's flight dynamics broadcast. Extracting flight change information from the flight dynamics broadcast through a pre-trained flight information recognition model. Comparing the flight change information with the dynamic data provided by the airline's official website and the system's internal order data to generate a recognition result. The present application can automatically extract flight change information from voice dynamic broadcasts, avoiding manual identification of each item, and greatly improving the efficiency of information extraction.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a method for dynamically identifying flight changes. Background Art

[0002] Currently, flight changes frequently occur in the aviation industry due to factors such as airline aircraft deployment, airport operation scheduling, and bad weather in the departure, destination or route areas. The traditional customer service process requires manual verification of flight information and notification of passengers, which is time-consuming and labor-intensive, and prone to information lags or errors. Summary of the invention

[0003] Based on this, it is necessary to provide a method for dynamically identifying flight changes in response to the above technical issues.

[0004] In a first aspect, a method for dynamically identifying flight changes is provided, the method comprising:

[0005] Get flight status updates from airlines;

[0006] Extracting flight change information from the flight dynamics broadcast using a pre-trained flight information recognition model;

[0007] The flight change information is compared with the dynamic data provided by the official website of the airline and the order data within the system to generate an identification result.

[0008] As an optional implementation, the flight change information includes the flight number, departure time, arrival time, departure place and destination before and after the flight change.

[0009] As an optional implementation, the extracting the flight change information in the flight dynamics broadcast by using a pre-trained flight information recognition model includes:

[0010] The voice data in the flight dynamics broadcast is converted into flight text data through the voice-to-text processing module of the flight information recognition model, and the flight change information in the flight text data is extracted through the flight information recognition model.

[0011] As an optional implementation, the training method of the flight information recognition model includes:

[0012] Performing data preprocessing on the flight text data used for model training, converting the flight text data into input data in a preset format and corresponding expected output data;

[0013] Based on a large-scale pre-trained language model, we add task-specific output layers for dynamic identification of flight changes and inject low-rank adapters into pre-selected key layers;

[0014] Performing model training using the input data and the corresponding expected output data, and simultaneously optimizing the parameters of the low-rank adapter and the parameters of the task-specific output layer in the large-scale pre-trained language model through back propagation;

[0015] The trained large-scale pre-trained language model is tested and verified using the real-time flight data of the airline to obtain the flight information recognition model.

[0016] As an optional implementation manner, the flight text data used for model training is preprocessed to convert the flight text data into input data in a preset format and corresponding expected output data, including:

[0017] Cleaning and standardizing the flight text data to extract key information, including flight number, departure time, arrival time, departure place and destination;

[0018] Embedding special characters to identify the start, end and interval of each key information field in the flight text data;

[0019] Perform word segmentation on flight text data embedded with special tags and map each token to a unique ID;

[0020] Construct the expected output data in a structured format according to preset rules.

[0021] As an optional implementation, the method further includes:

[0022] A random mask strategy is applied to the flight text data, that is, random mask processing is performed on some tokens.

[0023] As an optional implementation, the low-rank adapter adopts LoRA technology.

[0024] As an optional implementation, the real-time flight data of the airline is used to test and verify the trained large-scale pre-trained language model to obtain the flight information recognition model, including:

[0025] Using the real-time flight data as a test sample, inputting it into the model, and obtaining flight change information output by the model;

[0026] Comparing the matching degree between the flight change information output by the model and the real-time flight data, and calculating the recognition accuracy and error rate;

[0027] Based on the evaluation results, the low-rank adapter and task-specific output layer parameters in the model are tuned to optimize the overall performance of the model;

[0028] According to the preset automatic evaluation indicators, the model is verified and fed back.

[0029] As an optional implementation, the preset automatic evaluation indicators include recognition accuracy, response delay and robustness indicators and text summary similarity indicators.

[0030] As an optional implementation, the method further includes:

[0031] When the recognition result is a complete match result, an automatic order scheduling operation is performed;

[0032] When the recognition result is an incomplete match result, a triggering customer service processing operation is executed.

[0033] In a second aspect, a computer device is provided, comprising a memory and a processor, wherein the memory stores a computer program executable on the processor, and when the processor executes the computer program, the method steps described in any one of the first aspects are implemented.

[0034] According to a third aspect, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the method steps described in any one of the first aspects are implemented.

[0035] The present application provides a method for dynamically identifying flight changes. The technical solution provided by the embodiments of the present application brings at least the following beneficial effects: obtaining the flight dynamics broadcast of the airline; extracting the flight change information in the flight dynamics broadcast through a pre-trained flight information recognition model; comparing the flight change information with the dynamic data provided by the official website of the airline and the internal order data of the system to generate a recognition result. The present application uses a pre-trained flight information recognition model to automatically extract the flight change information in the voice dynamic broadcast, avoiding manual identification one by one, and greatly improving the efficiency of information extraction. By comparing the extracted information with the official website data of the airline and the internal order data, the accuracy of the identified flight information is ensured. At the same time, by introducing automated evaluation indicators, the model can maintain good robustness under various noises and complex scenarios.

[0036] It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0038] Figure 1 A flowchart of a method for dynamically identifying flight changes provided in an embodiment of the present application;

[0039] Figure 2 A flowchart of a method for training a flight information recognition model provided in an embodiment of the present application;

[0040] Figure 3 A flowchart of a data preprocessing method provided in an embodiment of the present application;

[0041] Figure 4 A flowchart of a method for testing and verifying a flight information identification model provided in an embodiment of the present application;

[0042] Figure 5 A schematic diagram of the structure of a computer device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0043] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0044] The following will describe in detail a method for dynamically identifying flight changes provided by an embodiment of the present application in conjunction with a specific implementation method. Figure 1 A flowchart of a method for dynamically identifying flight changes provided in an embodiment of the present application is shown in FIG. Figure 1 As shown, the specific steps are as follows:

[0045] Step S101, obtaining the flight status report of the airline.

[0046] In implementation, the system can obtain real-time flight dynamics broadcasts from the airline's internal system or official platform through the data acquisition module. Flight dynamics broadcasts can be real-time broadcasts in voice format or real-time text information pushed by the airline. The data acquisition module can obtain the corresponding information through interface calls (API), streaming media subscriptions, or directly from the data repository. For example, the system can subscribe to the airline's real-time flight dynamics data interface. Whenever the flight status changes, the interface will push voice broadcast data containing key information such as flight number, departure / arrival time, and departure and destination. The acquisition module stores the voice data in the local server in preparation for subsequent voice-to-text processing.

[0047] Step S102, extracting flight change information from the flight dynamics broadcast using a pre-trained flight information recognition model.

[0048] In implementation, relying on the pre-trained flight information recognition model, the model can accurately recognize the dynamic change information in the voice broadcast. For example, the voice broadcast collected by the system is: "Due to the adjustment of aviation scheduling, the departure time of the AB1234 flight from A to B is adjusted from 10:00 to 11:30, and the destination is adjusted from B to C." The flight information recognition model extracts the flight change information as follows: flight number before change: AB1234, flight number after change: AB1234, departure time before change: 10:00, departure time after change: 11:30, departure place before change: A, departure place after change: A, destination before change: B, destination after change: C.

[0049] Step S103, comparing the flight change information with the dynamic data provided by the airline's official website and the order data within the system to generate an identification result.

[0050] During implementation, the system can compare and contrast flight change information with real-time dynamic data from the airline's official website and the system's internal order data. By comparing information such as flight number, take-off / arrival time, departure place, destination, etc., the consistency between the data sources is determined to generate recognition results. For example: The system compares the extracted AB1234 flight information with the latest data released on the airline's official website and the records in the internal order database: The flight number can be compared: Confirm that the flight number is "AB1234" in both the official website and the internal data. Compare departure time: The official website data shows that the flight departure time has been updated to 11:30, which is consistent with the extracted information. Compare departure and destination: Both the internal order and official website data indicate that the departure place is "A" and the destination is "C".

[0051] As an optional implementation, the flight change information includes the flight number, departure time, arrival time, departure place and destination before and after the flight change.

[0052] As an optional implementation, in step S102, the method for extracting flight change information in the flight dynamics broadcast by using the pre-trained flight information recognition model is:

[0053] The voice-to-text processing module of the flight information recognition model is used to convert the voice data in the flight dynamic broadcast into flight text data, and the flight change information in the flight text data is extracted through the flight information recognition model.

[0054] In implementation, the speech-to-text processing module (such as ASR) processes the speech data and converts it into flight text data. For example, the converted flight text data is "Due to flight adjustments, the departure time of your flight AB1234 is adjusted from 10:00 to 11:30, the departure point A remains unchanged, and the destination is changed from B to C." The system can clean the text (such as removing noise and standardizing punctuation) and embed preset special character identifiers in the text. For example, add marks before and after "AB1234" for the model to identify the flight number, and add start and end symbols before and after the departure time "10:00" and "11:30" to facilitate the subsequent model to identify the key information of the departure time. The preprocessed text is input into the trained flight information recognition model. The flight information recognition model can recognize: "original flight information": {"flightNo":"AB1234","depTime":"10:00","dep":"Place A","arr":"Place B"}, "new flight information": {"flightNo":"AB1234","depTime":"11:30","dep":"Place A","arr":"Place C"}.

[0055] As an optional implementation, Figure 2 A flowchart of a method for training a flight information recognition model provided in an embodiment of the present application, such as Figure 2 As shown, the specific steps are as follows:

[0056] Step S201 , preprocessing the flight text data used for model training, converting the flight text data into input data in a preset format and corresponding expected output data.

[0057] In implementation, the goal of data preprocessing is to convert the original flight text data (e.g., flight announcements, notification text messages, customer service conversations, etc.) into an input acceptable to the model, and at the same time construct the corresponding expected output data (e.g., structured flight information represented by JSON or a custom format). For example: the flight text data is: "Your flight AB1234 will take off from place A on January 24, 2025, and the destination will be changed to place C." After data preprocessing, the text can be converted into the following format: "Your flight MF8964 will take off from place A on 2025-1-24, and the destination will be changed to place C." The corresponding expected output data can be constructed into a structured format, such as: {"flightNo": "AB1234", "depDate": "2025-1-24", "dep": "place A", "arr": "place C"}. Referring to the above embodiment, the input during model training is a text sequence with special tags, and the target output is the above structured data, which can ensure that the model can learn the corresponding relationship between the tags and the specific flight information.

[0058] As an optional implementation, Figure 3 A flowchart of a data preprocessing method provided in an embodiment of the present application is shown in FIG. Figure 3 As shown, in step S201, the flight text data used for model training is preprocessed, and the specific steps of converting the flight text data into input data in a preset format and corresponding expected output data are as follows:

[0059] Step S301, clean and standardize the flight text data to extract key information, including flight number, departure time, arrival time, departure place and destination.

[0060] In the implementation, noise information is removed from the original flight text data, and the date, time, location and other contents are uniformly formatted to ensure the consistency of subsequent processing. Natural language processing (NLP) technology and regular expressions can be used to automatically identify and extract key information with fixed formats or obvious features in the text, such as flight number (usually a combination of letters and numbers), departure / arrival time (uniform format is "HH:MM" or standard date format), and departure and destination names. For example: the flight text data is: "Dear passengers, your AB1234 flight will take off from A on January 26, 2025, and the destination will be changed to C. Please pay attention to the latest notifications in time." The system can first clean the text: remove irrelevant greetings and redundant words. Standardize the date and time, for example, convert "January 26, 2025" to "2025-1-26" to ensure uniform format. Identify the flight number by matching the "MF\d+" format with regular expressions. Finally, the key information was extracted: flight number: AB1234, departure date: 2025-1-26, departure place: A, destination: C.

[0061] Step S302: embed special characters for identifying the start, end and interval of each key information field into the flight text data.

[0062] In implementation, in order to enable the model to clearly distinguish and locate each key information field, predefined special characters or labels can be added to the text. These special characters or labels can play the role of separation and marking, so that the boundary of each information segment (such as flight number, departure time, etc.) is clear, thereby improving the model's recognition and learning efficiency of key information. Still taking the text extracted in step S301 as an example, the text after cleaning and standardization is:

[0063] "Your flight AB1234 will take off from A on 2025-1-26, and the destination will be changed to C." Insert special tags for each key information field, for example: insert before and after the flight number <flightno> and< / flightno> , insert before and after the departure date <depdate> and< / depdate> , insert before and after the departure place <dep> and< / dep> , insert before and after the destination <arr> and< / arr> .

[0064] Step S303, perform word segmentation processing on the flight text data embedded with special tags, and map each token to a unique ID.

[0065] In the implementation, the flight text data with special tags is segmented, that is, the long text is split into several basic units (tokens). Each token can be mapped to a unique ID, which ensures that the text is converted into a digital sequence before entering the model. The digital representation after word segmentation provides a standard input format for the model and retains the embedded special tag information, which facilitates the model to learn the correspondence between tags and specific information. For example: "Your flight AB1234 will take off from A on 2025-1-26, and the destination will be changed to C." The above text can be split into the following token sequence: Token1: "Your", Token2: "", Token3: "AB1234", Token4: "", Token5: "Flight", Token6: "Will", Token7: "", Token8: "2025-1-26", Token9: "", Token10: "From", Token11: "", Token12: "A", Token13: "", Token14: "Take off," Token15: "Destination", Token16: "Changed to", Token17: "", Token18: "C", Token19: "". Each token is mapped to a unique ID through the vocabulary, for example, "your" corresponds to ID101, "" corresponds to ID202, and so on. The generated sequence of numbers is used as the input of the model, which is conducive to the identification and matching of special tags and key information during the training process.

[0066] Step S304: constructing expected output data in a structured format according to preset rules.

[0067] In the implementation, each key information extracted from the original text is organized in the form of structured data to form a one-to-one correspondence with the predicted output of the model. JSON or a custom format is usually used to construct structured data, and each field (such as flight number, departure time, departure place, destination, etc.) is clearly identified to provide a clear supervision signal for model training. For example: the system can construct the following structured output data according to preset rules: {"flightNo": "AB1234", "depDate": "2025-1-24", "dep": "A", "arr": "C"}. Structured data is the output that the model expects to obtain during the training process, which helps the model learn how to extract and organize key information from tokenized text.

[0068] As an optional implementation, the method further includes:

[0069] A random mask strategy is applied to the flight text data, that is, some tokens are randomly masked.

[0070] In implementation, some tokens in the input text data can be randomly replaced with special mask tags (such as "[MASK]") to simulate the situation where information is missing in the input data. This strategy can force the model to pay more attention to contextual information during training, thereby improving the model's ability to capture key information and robustness, preventing overfitting, and enhancing the model's fault tolerance to abnormal inputs in practical applications. For example: when constructing training samples, the masked token sequence can be used as the model input, while retaining the original token sequence as a supervisory signal, requiring the model to predict the original token at the mask position. By comparing the difference between the predicted token and the original token, the loss is calculated and the parameters are back-propagated to optimize. The random mask strategy can still recover the correct key information from the context when encountering incomplete or noisy input, thereby improving the overall recognition accuracy and robustness of the model.

[0071] Step S202, based on the large-scale pre-trained language model, add a task-specific output layer for dynamic identification of flight changes, and inject a low-rank adapter into the pre-selected key layer.

[0072] In implementation, the model structure is customized based on a large-scale pre-trained model (such as LLaMA) to adapt to the task of flight information extraction. Specifically, task-specific output layers (for example, sequence annotation or classification heads) can be added to the last few layers or other key layers of the original model structure to specifically predict fields such as flight number and departure time. At the same time, low-rank adapters (such as LoRA modules) are injected to improve training efficiency and reduce computing resource requirements by tuning only a small number of parameters without significantly modifying the weights of the pre-trained model. For example, developers can add LoRA modules to the 11th and 12th layers of the LLaMA model. These modules have low parameter dimensions, only perform low-rank decomposition on part of the weight matrix, and add the output of the adapter to the original activation value during the forward propagation of the model. At the same time, a task-specific linear layer is added to the output of the model to map the hidden state of the Transformer to a token sequence containing special tags or a prediction result of structured data. In this way, when the pre-processed flight text data is input, the model can not only use the original language understanding ability of LLaMA, but also capture and output information related to flight changes through the newly added layer.

[0073] Step S203, using the input data and the corresponding expected output data to perform model training, and simultaneously optimizing the parameters of the low-rank adapter and the parameters of the task-specific output layer in the large-scale pre-trained language model through back propagation.

[0074] In practice, during the training process, the error between the input data and the expected output data is fed back to the model through back propagation, specifically updating the parameters of the low-rank adapter and the task-specific output layer, while keeping the main parameters of the pre-trained model basically unchanged, so as to achieve the purpose of efficient adaptation to specific tasks. For example: in one training iteration, after the model inputs the pre-processed text, the output sequence passes through the task-specific output layer to obtain the predicted result, which is compared with the expected output data and the loss is calculated. Subsequently, the back propagation algorithm is used to update the parameters of the LoRA module and the output layer. After many iterations, the model gradually learns how to accurately extract information such as flight number, time and location from the text.

[0075] As an optional implementation, the low-rank adapter may adopt LoRA technology.

[0076] Step S204, using the real-time flight data of the airline to test and verify the trained large-scale pre-trained language model to obtain a flight information recognition model.

[0077] In implementation, after the model training is completed, it can be verified through the actual real-time flight data of the airline to evaluate the recognition performance of the model in real scenarios. The testing process includes inputting the latest real-time voice or text flight dynamic data into the model, comparing the output results with the data on the airline's official website or internal order system, and calculating the recognition accuracy, error rate and other indicators. The verification results are used to confirm whether the model has achieved the expected recognition effect and whether it has the robustness and stability for actual production applications. For example, the real-time broadcast data of the day's flights is converted into text through the ASR module and input into the trained model, and the model outputs the predicted structured flight information. Then, the system automatically compares the predicted results with the flight information published on the airline's official website and the information recorded in the internal system. Assume that during the comparison process, it is found that the flight number matching rate is 99%, the departure time matching rate is 97%, and the departure and destination matching rates are both 98%. At this point, the calculated overall recognition accuracy exceeds the preset standard (for example, 95%), thereby confirming that the model meets the actual application requirements, and the model is used as a flight information recognition module in the subsequent order scheduling and customer service processing process.

[0078] As an optional implementation, Figure 4 A flowchart of a method for testing and verifying a flight information identification model provided in an embodiment of the present application is shown in FIG. Figure 4 As shown, in step S204, the real-time flight data of the airline is used to test and verify the trained large-scale pre-trained language model, and the specific steps of obtaining the flight information recognition model are as follows:

[0079] Step S401, using real-time flight data as a test sample, inputting it into the model, and obtaining flight change information output by the model.

[0080] In implementation, the real-time flight data of airlines can be used as test samples to simulate the working conditions of the model in real scenarios. Considering the quality and consistency of the data, the data in production is screened, several representative topics are selected, and the process of model evaluation is standardized. The data can be real-time text information or text converted by the ASR module, which is input into the trained large-scale pre-trained language model. The model processes the input data based on the features and structured tags learned in the previous training, and outputs the corresponding flight change information. For example, the system receives a piece of flight dynamic broadcast data in real time: "Flight CD4567 was originally scheduled to take off at 10:00, but it is now adjusted to 10:30, and the departure point is changed from D to E." After conversion by the ASR module, it is formed into text and then input into the flight information recognition model. The flight information recognition model outputs structured results based on internal training rules and labeling information, such as: {"flightNo":"CD4567","origDepTime":"10:00","newDepTime":"10:30","origDep":"D","newDep":"E"}. This output is the flight change information extracted by the model.

[0081] Step S402, comparing the flight change information output by the model with the matching degree of the real-time flight data, and calculating the recognition accuracy and error rate.

[0082] In implementation, the structured flight information output by the model is compared with the real-time flight data provided by the airline's official website or other authoritative channels. By comparing the matching of key information fields (such as flight number, departure time, departure place and destination), the recognition accuracy and error rate of the model under actual data are calculated. A high degree of matching indicates that the model has strong extraction capabilities, and vice versa, it reflects that the model has deficiencies in extracting certain information. For example: if all key fields are matched, the recognition accuracy is 100%; if a field deviates, the error rate is calculated. For example, if the number of correctly matched fields is 3 / 4, the accuracy is 75% and the error rate is 25%. These statistical indicators can be used for subsequent model tuning.

[0083] Step S403: Based on the evaluation results, the low-rank adapter and task-specific output layer parameters in the model are tuned to optimize the overall performance of the model.

[0084] In the implementation, according to the recognition accuracy and error rate obtained in step S402, the low-rank adapter (LoRA module) and the task-specific output layer in the model specifically used for flight information extraction are fine-tuned using back propagation and parameter adjustment techniques. Through continuous iterative training, these parameters are adjusted to reduce errors and improve overall recognition accuracy and robustness. This process belongs to the later fine-tuning of the model to ensure that the model can reach the preset performance indicators in the actual production environment. For example: the test results show that in some test samples, there are errors in the recognition of the take-off time or departure point (for example, the error rate is 10%). The system will adjust the learning rate, update step size or parameter weight of the low-rank adapter or the task-specific output layer. After one or more rounds of tuning, real-time data is used for testing again. If the accuracy rate after adjustment is increased to more than 95%, it proves that the model parameter tuning is effective. For example, after the original parameter adjustment, the recognition accuracy rate on the validation set is increased from 85% to 96%, which proves that the tuning measures have successfully optimized the overall performance of the model.

[0085] Step S404: provide verification feedback to the model according to preset automated evaluation indicators.

[0086] In implementation, a series of automated evaluation indicators, such as recognition accuracy, response delay, robustness indicators, and text summary similarity (such as ROUGE indicators), can be used to comprehensively evaluate the overall performance of the model. The model output is compared with the standard results through preset indicators to form a quantitative feedback report. This feedback can not only intuitively show the performance of the model in different test dimensions, but also guide the next step of model optimization or parameter adjustment. For example, the system sets the evaluation indicators to require that the recognition accuracy is not less than 95%, the response delay is within 500 milliseconds, and the ROUGE score is above 90%. Through the automatic evaluation program, the system evaluates a batch of real-time test samples, with a recognition accuracy of 96.5%, an average response delay of 450 milliseconds, and a ROUGE score of 91%. Based on these results, feedback can be given to the training module to confirm that the current model has met the expected performance requirements. If a certain indicator does not meet the standard, the system can prompt that further tuning is required.

[0087] As an optional implementation, the preset automatic evaluation indicators include recognition accuracy, response delay and robustness indicators, and text summary similarity indicators.

[0088] As an optional implementation, the method further includes:

[0089] When the recognition result is a complete match, an automatic order scheduling operation is performed.

[0090] When the recognition result is not a complete match, the trigger customer service processing operation is executed.

[0091] In implementation, when the system compares the flight change information extracted with the airline's official website and internal order data, if all key information (such as flight number, departure time, arrival time, departure place, and destination) is completely consistent, it means that the information extraction is accurate and the data source is authoritative. At this time, the order scheduling operation can be automatically completed according to the preset business rules without manual intervention. On the contrary, if the match is not complete, there may be problems such as information recognition errors or data update delays. To ensure the safety and accuracy of scheduling, the system can automatically trigger customer service processing operations, which will be further verified and processed manually. For example: After the system receives the flight dynamic broadcast, it compares it with the airline's official website and internal order data, and all fields are consistent. According to the preset rules, the order scheduling instructions can be automatically generated, the order status can be updated to "flight adjustment confirmation", and the passengers can be notified through SMS, APP messages or emails. In this process, no manual intervention by customer service is required, and a fully automatic process is realized. If the real-time data obtained from the airline's official website shows that the departure time is still displayed as the time before the change or other fields are inconsistent, it can be judged as an incomplete match based on the comparison results, and the abnormal information can be automatically recorded and the customer service processing process can be triggered. After receiving the system notification, the customer service staff can manually check the relevant data, confirm and modify the order information, or contact the passenger to further coordinate the handling plan to ensure that subsequent scheduling will not be confused due to data errors.

[0092] An embodiment of the present application provides a method for dynamically identifying flight changes, the method comprising: obtaining an airline's flight dynamics broadcast. Extracting flight change information from the flight dynamics broadcast using a pre-trained flight information recognition model. Comparing the flight change information with the dynamic data provided by the airline's official website and the system's internal order data to generate a recognition result. The embodiment of the present application utilizes a pre-trained flight information recognition model to automatically extract flight change information from voice dynamic broadcasts, avoiding manual identification of each item, and greatly improving the efficiency of information extraction. By comparing the extracted information with the airline's official website data and internal order data, the accuracy of the identified flight information is ensured. At the same time, by introducing automated evaluation indicators, the model can maintain good robustness under various noises and complex scenarios.

[0093] It should be understood that although Figures 1 to 4 The steps in the flowchart are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, Figures 1 to 4At least part of the steps may include multiple steps or multiple stages. These steps or stages are not necessarily performed at the same time, but can be performed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed in turn or alternately with other steps or at least part of the steps or stages in other steps.

[0094] It can be understood that the same / similar parts between the various embodiments of the above method in this specification can refer to each other, and each embodiment focuses on the differences from other embodiments. For related points, please refer to the description of other method embodiments.

[0095] In one embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program that can be executed on the processor. When the processor executes the computer program, the method steps of dynamically identifying flight changes are implemented. Figure 5 A schematic diagram of the structure of a computer device provided in an embodiment of the present application is shown in FIG. Figure 5 As shown, the computer device may include a processor 501, a system bus 502, a non-volatile storage medium 503, an internal memory 504, a network interface 505, a display screen 506 and an input device 507. Among them, the non-volatile storage medium 503 stores an operating system 5031 and a computer program 5032. The processor 501 is used to execute the computer program 5032 to implement the above-mentioned method steps for dynamic identification of flight changes. The system bus 502 is used to connect the processor 501, the non-volatile storage medium 503, the internal memory 504, the network interface 505, the display screen 506 and the input device 507 to ensure efficient communication between the components. The internal memory 504 is used to temporarily store the running programs and data to help the processor 501 quickly access the required information, thereby improving the overall system performance. The network interface 505 (such as a network card) enables the computer device to be connected to a local area network or the Internet to achieve data transmission and remote communication. The display screen 506 is used to present the information of dynamic identification of flight changes processed by the computer device to the user in the form of graphics or text. The input device 507 (such as a keyboard, a mouse, a touch screen, etc.) is used to allow the user to input flight data and commands into the computer device to achieve interactive operations with the computer device.

[0096] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0097] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0098] It should also be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for display, data for analysis, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.

[0099] Each embodiment in this specification is described in a related manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0100] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0101] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.

Claims

1. A method for dynamically identifying flight changes, characterized in that: The method comprises: Get flight status updates from airlines; Extracting flight change information from the flight dynamics broadcast using a pre-trained flight information recognition model; Compare the flight change information with the dynamic data provided by the airline's official website and the order data within the system to generate an identification result; The training method of the flight information recognition model comprises: Performing data preprocessing on the flight text data used for model training, converting the flight text data into input data in a preset format and corresponding expected output data; Based on a large-scale pre-trained language model, we add task-specific output layers for dynamic identification of flight changes and inject low-rank adapters into pre-selected key layers; Performing model training using the input data and the corresponding expected output data, and simultaneously optimizing the parameters of the low-rank adapter and the parameters of the task-specific output layer in the large-scale pre-trained language model through back propagation; The trained large-scale pre-trained language model is tested and verified using the real-time flight data of the airline to obtain the flight information recognition model.

2. The method according to claim 1, characterized in that The flight change information includes the flight number, departure time, arrival time, departure place and destination before and after the flight change.

3. The method according to claim 1, characterized in that: The extracting of flight change information from the flight dynamics broadcast using a pre-trained flight information recognition model includes: The voice data in the flight dynamics broadcast is converted into flight text data through the voice-to-text processing module of the flight information recognition model, and the flight change information in the flight text data is extracted through the flight information recognition model.

4. The method according to claim 1, characterized in that: The data preprocessing of the flight text data used for model training, converting the flight text data into input data in a preset format and corresponding expected output data, includes: Cleaning and standardizing the flight text data to extract key information, including flight number, departure time, arrival time, departure place and destination; Embedding special characters to identify the start, end and interval of each key information field in the flight text data; Perform word segmentation on flight text data embedded with special tags and map each token to a unique ID; Construct the expected output data in a structured format according to preset rules.

5. The method according to claim 4, characterized in that The method further comprises: A random mask strategy is applied to the flight text data, that is, random mask processing is performed on some tokens.

6. The method according to claim 1, characterized in that The low-rank adapter adopts LoRA technology.

7. The method according to claim 1, characterized in that The method of testing and verifying the trained large-scale pre-trained language model using the real-time flight data of the airline to obtain the flight information recognition model includes: Using the real-time flight data as a test sample, inputting it into the model, and obtaining flight change information output by the model; Comparing the matching degree between the flight change information output by the model and the real-time flight data, and calculating the recognition accuracy and error rate; Based on the evaluation results, the low-rank adapter and task-specific output layer parameters in the model are tuned to optimize the overall performance of the model; According to the preset automatic evaluation indicators, the model is verified and fed back.

8. The method according to claim 7, characterized in that The preset automatic evaluation indicators include recognition accuracy, response delay and robustness indicators and text summary similarity indicators.

9. The method according to claim 1, characterized in that: The method further comprises: When the recognition result is a complete match result, an automatic order scheduling operation is performed; When the recognition result is an incomplete match result, a triggering customer service processing operation is executed.

Images