Interaction method and device for service consultation

Abstract

The invention provides an interaction method and device for service consultation, and relates to the technical field of artificial intelligence. The method comprises the following steps: performing field-specific text preprocessing on a civil aviation voice instruction text sequence to obtain a Token sequence; encoding the Token sequence through an encoder layer of the joint semantic understanding model; performing intention classification on the word vector representation through an intention classification layer; performing role classification on the word vector representation through a role classification layer; performing slot filling on the word vector representation through a slot filling layer; constructing an annotation task queue; distributing the tasks to corresponding terminals for manual proofreading and labeling based on predefined role permission control logic; and updating parameters of the joint semantic understanding model based on the feedback data. According to the method, the problems of high data annotation cost, difficulty in complex instruction semantic analysis and low non-standard language recognition rate in the civil aviation field are effectively solved, and the accuracy of voice instruction understanding and the self-adaptive capability of the system are remarkably improved.

Description

Technical Field

[0001] This invention relates to the fields of computer, artificial intelligence, and civil aviation traffic management technology, and in particular to an interactive method and device for service consultation. Background Technology

[0002] In civil aviation communications, voice commands between pilots and air traffic controllers are characterized by high structure, dense technical terminology, and rigorous semantics. To build an intelligent voice understanding system for civil aviation, a large number of voice commands need to be annotated, including intent recognition, role determination, and semantic slot filling. Traditional annotation methods rely on manual sentence-by-sentence dictation and annotation, which is inefficient, costly, and inconsistent.

[0003] Existing automatic annotation methods are mostly based on rules or traditional machine learning models, which are difficult to handle the complex semantics, multiple nested intentions, and technical terms in civil aviation instructions. Although pre-trained language models such as BERT (Bidirectional Encoder Representations from Transformers) perform well in general domains, their direct application in vertical domains such as civil aviation still suffers from insufficient domain adaptability and inaccurate identification of specialized slots.

[0004] Therefore, there is an urgent need for an efficient, accurate, and scalable semi-automatic annotation scheme for civil aviation instructions to improve annotation efficiency and quality, and to provide high-quality training data for subsequent speech understanding systems. Summary of the Invention

[0005] To address the technical problems of high data annotation costs, difficulty in parsing complex command semantics, and low recognition rate of non-standard grammar in existing technologies in the civil aviation field, this invention provides an interactive method and apparatus for service consultation. The technical solution is as follows:

[0006] On the one hand, an interactive method for service consultation is provided, which is implemented by an interactive device for service consultation, and the method includes: S1. Receive the sequence of civil aviation voice command texts to be processed.

[0007] S2. Perform domain-specific text preprocessing on the civil aviation voice command text sequence to obtain the Token sequence.

[0008] S3. The token sequence is encoded by the encoder layer of the pre-trained joint semantic understanding model to obtain a context-dependent word vector representation; wherein, the joint semantic understanding model is constructed based on the encoder representation model of the joint bidirectional transformer.

[0009] S4. Classify the word vector representation by intent through the intent classification layer of the joint semantic understanding model to obtain the predicted intent classification result; classify the word vector representation by role through the role classification layer to obtain the predicted speaker role label; fill the word vector representation by slots through the slot filling layer to obtain the predicted entity slot sequence; construct the annotation task queue based on the intent classification result, speaker role label and entity slot sequence.

[0010] S5. Based on predefined role-based access control logic, tasks in the annotation task queue are distributed to the corresponding terminals for manual verification and annotation, and manual annotation feedback data is obtained.

[0011] S6. Update the parameters of the joint semantic understanding model based on manually labeled feedback data.

[0012] Optionally, S2 includes: The civil aviation voice command text sequence is segmented into a token sequence based on a custom dictionary in the civil aviation field. Regular expressions are then used to map the numeric characters in the civil aviation voice command text sequence to standard Arabic numerals, and the standard Arabic numerals are then split into independent characters.

[0013] Optionally, in S4, the word vector representation is classified by the intent classification layer to obtain the predicted intent classification result, including: The sentence-first feature vector in the word vector representation is input into the intent classification layer, and the intent label logical value is output.

[0014] Apply the Sigmoid activation function to the logical values ​​of the intent labels to obtain the probability value for each predefined intent category.

[0015] The probability value is compared with a preset judgment threshold. Multiple intent categories with probability values ​​greater than the preset judgment threshold are selected. The selected multiple intent categories are connected by a predetermined separator to generate a composite intent string as the predicted intent classification result.

[0016] Optionally, in S4, the word vector representation is classified by role through a role classification layer to obtain the predicted speaker role label, including: Input the sentence-first feature vector from the word vector representation into the role classification layer, and output the role label logical value.

[0017] Apply the Softmax activation function to the logical values ​​of the role labels and select the category index with the highest probability as the predicted speaker role label; where the speaker role label is either pilot or controller.

[0018] The speaker's role label is associated with and stored with the intent classification results.

[0019] Optionally, in S4, slot filling is performed on the word vector representation through a slot filling layer to obtain the predicted entity slot sequence, including: The sequence feature vector in the word vector representation is input into the slot filling layer, and the logical value of the slot label corresponding to each token is output.

[0020] The slot tag logic value is used as the emission score input to the conditional random field layer to learn the tag transition matrix.

[0021] Based on the label transition matrix, the Viterbi algorithm is used to calculate the globally optimal slot label sequence as the predicted entity slot sequence.

[0022] Optionally, the predefined role-based access control logic in S5 distributes tasks in the labeled task queue to the corresponding terminals, including: Based on data access requests from different terminals, extract user credentials, task names, and target data IDs from the data access requests.

[0023] The permission verification function is called to verify the legitimacy of the user credentials and identify the role attribute of the user credentials. If the role attribute is administrator, access permissions are granted to all tasks and all data IDs. If the role attribute is an annotator, the task list in the terminal is traversed to verify whether the task name exists in the task list and whether the target data ID falls within the closed range of the terminal. If so, access or modification permissions are granted; otherwise, the request is rejected and an error message is returned.

[0024] Optionally, the predefined role-based access control logic in S5 distributes tasks in the task queue to the corresponding terminals, and also includes: administrator-triggered automatic task distribution.

[0025] Among them, the automatic task distribution triggered by the administrator includes: Read the original data source files of the tasks in the labeling task queue, parse and clean the original data source files, and assign a globally unique ID to each task. The IDs of all tasks are incremented.

[0026] Get the list of active annotators in the current system.

[0027] Calculate the quotient between the total amount of data in the task and the number of annotators, and determine the basic allocation number for each annotator based on the quotient.

[0028] Iterate through the list of annotators, calculate and assign the start ID and end ID to each annotator based on the base allocation quantity, ensuring that the ID ranges are continuous and do not overlap.

[0029] Write the calculated task name and its corresponding ID range into the user configuration document in the database to complete the task binding.

[0030] On the other hand, an interactive device for service consultation is provided, which is applied to an interactive method for service consultation, the device comprising: The data receiving module is used to receive the text sequence of civil aviation voice commands to be processed.

[0031] The data preprocessing module is used to perform domain-specific text preprocessing on civil aviation voice command text sequences to obtain Token sequences.

[0032] The encoding module is used to encode the token sequence through the encoder layer of the pre-trained joint semantic understanding model to obtain context-related word vector representations; wherein, the joint semantic understanding model is constructed based on the encoder representation model of the joint bidirectional transformer.

[0033] The joint reasoning module is used to classify the word vector representation by intent through the intent classification layer of the joint semantic understanding model to obtain the predicted intent classification result; to classify the word vector representation by role through the role classification layer to obtain the predicted speaker role label; to fill the word vector representation by slots through the slot filling layer to obtain the predicted entity slot sequence; and to construct a labeling task queue based on the intent classification result, speaker role label and entity slot sequence.

[0034] The database access control module is used to distribute tasks in the annotation task queue to the corresponding terminals for manual proofreading and annotation based on predefined role access control logic, and obtain manual annotation feedback data.

[0035] The data export module is used to update the parameters of the joint semantic understanding model based on manually labeled feedback data.

[0036] On the other hand, an interactive device for service consultation is provided, the interactive device for service consultation comprising: a processor; a memory storing computer-readable instructions, wherein when the computer-readable instructions are executed by the processor, any one of the interactive methods for service consultation described above is implemented.

[0037] On the other hand, a computer-readable storage medium is provided, wherein at least one instruction is stored therein, the at least one instruction being loaded and executed by a processor to implement any of the above-described interactive methods for service consultation.

[0038] The beneficial effects of the technical solutions provided by the embodiments of the present invention include at least the following: This invention achieves a transformation in civil aviation voice command processing from a "rough manual annotation" model to a "semi-automatic human-machine coupling" model by constructing a complete system encompassing Joint BERT inference, RBAC access control, and automated data flow. The model addresses the challenges of recognizing multiple intents, single intents, and complex entities through Sigmoid, Softmax, and CRF technologies, while the accompanying data management system solves security and efficiency issues in large-scale collaboration through strict segmented access control. This overall solution significantly reduces data production costs while improving recognition accuracy, demonstrating high industrial practical value. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 This is a flowchart of an interactive method for service consultation provided by an embodiment of the present invention; Figure 2 This is a schematic diagram of a model reasoning subsystem, a data management and access control subsystem, and a data interaction and export subsystem provided in an embodiment of the present invention; Figure 3 This is a diagram of the front-end display interface provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the dataset corpus provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the role recognition results provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of the new training model provided in an embodiment of the present invention; Figure 7 This is a schematic diagram of the recognition results of the new training model provided in this embodiment of the invention; Figure 8 This is a block diagram of an interactive device for service consultation provided in an embodiment of the present invention; Figure 9 This is a schematic diagram of the structure of an interactive device for service consultation provided in an embodiment of the present invention. Detailed Implementation

[0041] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0042] In the embodiments of the present invention, words such as "exemplarily" and "for example" are used to give examples, illustrations or explanations. Any embodiment or design solution described as an "example" in the present invention should not be construed as being more preferred or having more advantages than other embodiments or design solutions. Rather, the use of the word "example" is intended to present concepts in a specific manner. In addition, in the embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one of the two can be selected.

[0043] In the embodiments of the present invention, "image" and "picture" can sometimes be used interchangeably. It should be noted that when not emphasizing the difference, the meanings they express are the same. "(of)", "corresponding", and "corresponding" can sometimes be used interchangeably. It should be noted that when not emphasizing the difference, the meanings they express are the same.

[0044] In the embodiments of the present invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When not emphasizing the difference, the meanings they express are the same.

[0045] To make the technical problems, technical solutions and advantages to be solved by the present invention clearer, the following will be described in detail with reference to the accompanying drawings and specific embodiments.

[0046] The embodiments of the present invention provide an interaction method for service consultation. This method can be implemented by an interaction device for service consultation, and the interaction device for service consultation can be a terminal or a server. As Figure 1 shown in the flowchart of the interaction method for service consultation, the processing flow of this method can include the following steps:

[0047] S1. Receive a civil aviation voice command text sequence to be processed.

[0048] S2. Perform domain-specific text preprocessing on the civil aviation voice command text sequence to obtain a Token sequence.

[0049] Specifically, the text cutting logic is configured to: preferentially load a custom dictionary containing civil aviation terms; perform preliminary word segmentation; use a regular pattern ([dongyao two three four five six seven eight nine 0-9]) to perform secondary forced splitting on the Tokens containing numbers in the preliminary word segmentation result, and break up continuous number sequences into single-character Tokens.

[0050] In a feasible implementation, domain-specific text preprocessing is performed, including mapping the characters representing the unique digital readings in civil aviation to standard Arabic numerals using regular expressions and performing word segmentation to generate a Token sequence. Specifically, in the field of civil aviation, the reading and segmentation of numbers are crucial (for example, "one" represents 1 and "seven" represents 7). To ensure that the model can correctly understand, the cut_mode module and the convert_input_file_to_tensor_dataset function in the code implement strict preprocessing logic.

[0051] Furthermore, the domain-specific text preprocessing steps also include special character cleaning and segmentation: Build and load a custom dictionary for the civil aviation field. The dictionary contains waypoint names, airline codes, and specific terms to prevent the tokenizer from incorrectly segmenting proper nouns.

[0052] For special digital expressions in the text (such as "zero", "one", "two", "seven"), use regular expression ([zero one two three four five six seven eight nine 0-9]) to scan the text and force it to be segmented into independent characters to prevent it from merging with adjacent text. Specifically, once the above numbers or their variants are found, the system will force them to be segmented into independent Tokens. For example, the string "altitude one two" will be segmented into ['altitude', 'one', 'two'], rather than ['altitude', 'one two']. This ensures that each digit can be independently labeled and recognized, which is crucial for the extraction of information such as flight numbers and headings.

[0053] Furthermore, before inputting into the model, the method also includes tensor construction and masking processing steps: perform Sub-word Tokenization on the cleaned text using the tokenizer; construct a slot label mask, mark the first Sub-token of each original word as the valid prediction position, and mark the subsequent Sub-tokens and padding positions of the same word as ignored positions; add classification identifiers () and delimiter identifiers () at the beginning and end of the sequence respectively; generate an attention mask, set the position of the real Token to 1 and the padding position to 0; convert the above sequence into the PyTorch tensor format and encapsulate it as a TensorDataset.

[0054] Specifically, since BERT-like models use WordPiece segmentation, a word may be split into multiple sub-words. To maintain input-output alignment, the system constructs a `slot_label_mask`. For the first sub-token of each original word, the mask is set to a valid value (e.g., `pad_token_label_id + 1`); for subsequent sub-tokens, the mask is set to an invalid value (e.g., `pad_token_label_id`). This ensures that when calculating the loss or decoding, the model only outputs a label for the first token with semantic integrity. Special token addition: a token for intent classification is added at the beginning of the sequence, and a token indicating the end of the sentence is added at the end. If the sequence length exceeds `max_seq_len`, truncation is performed.

[0055] The preprocessing steps also include a position index mapping function, which is used to calculate the start and end character positions of each word after word segmentation in the original input string, so as to support the highlighting of entities on the front-end interface.

[0056] S3. The token sequence is encoded by the encoder layer of the pre-trained joint semantic understanding model to obtain a context-dependent word vector representation; wherein, the joint semantic understanding model is constructed based on the encoder representation model of the joint bidirectional transformer.

[0057] In one feasible implementation, the token sequence is encoded using a pre-trained Joint BERT (Joint Bidirectional Encoder Representations from Transformers) model to obtain context-sensitive word vector representations. The Joint BERT model includes a shared BERT encoder layer, an intent classification layer, a role classification layer, and a slot-filling layer. The Joint BERT model supports dynamically selecting BERT, Distil BERT, or ALBERT as the pre-trained backbone network via a configuration file. The core advantage of the JointBERT architecture lies in the shared underlying encoder, enabling intent classification and slot-filling tasks to mutually reinforce each other. For example, recognizing the intent as "command heading" can significantly increase the probability of recognizing subsequent numbers as "heading values" rather than "speed values."

[0058] S4. Classify the word vector representation by intent through the intent classification layer of the joint semantic understanding model to obtain the predicted intent classification result; classify the word vector representation by role through the role classification layer to obtain the predicted speaker role label; fill the word vector representation by slots through the slot filling layer to obtain the predicted entity slot sequence; construct the annotation task queue based on the intent classification result, speaker role label and entity slot sequence.

[0059] In one feasible implementation, based on word vector representation, the intent recognition task, role recognition task, and slot filling task are executed in parallel to obtain the predicted intent classification result, speaker role label, and entity slot sequence. The inference process is encapsulated in the Model (single intent) and Multi_Model (multi-intent) classes, supporting execution on CPU or GPU. Forward propagation: The model receives input_ids, attention_mask, and token_type_ids. In torch.no_grad() mode, the BERT encoder outputs a sequence of context vectors. The intent classification header processes the vector and outputs intent_logits; the slot filling header processes the sequence vector and outputs slot_logits.

[0060] Optionally, the intent recognition task is configured to support multi-label classification, specifically including: Input the sentence-first feature vector output by the BERT encoder into the intent classification layer, and output the intent label logical value.

[0061] Apply the Sigmoid activation function to the logical values ​​of the intent labels to obtain the probability value for each predefined intent category.

[0062] The probability value is compared with a preset judgment threshold. Multiple intent categories with probability values ​​greater than the preset judgment threshold are selected. The selected multiple intent categories are connected by a predetermined separator to generate a composite intent string as the predicted intent classification result.

[0063] In one feasible implementation, all intent categories with probability values ​​greater than a judgment threshold are identified as intent classification results; in response to the existence of multiple categories that meet the conditions, the multiple categories are connected by a predetermined separator to generate a composite intent string.

[0064] The intent classification results cover specific business intents in the civil aviation field, including: requesting changes in flight altitude, instructing adjustments to flight course, confirming runway information, reporting fuel status, requesting landing clearance, instructing go-arounds, and ground taxiing instructions, etc.

[0065] When handling multi-intent recognition, the joint inference module applies the torch.sigmoid function to process the intent logic value; uses numpy.where to filter out intent indices with a probability greater than 0.5; uses the intent label list to map the indexes to text labels, and concatenates them into a single string using the "#" symbol for output.

[0066] The intent classification method of this invention employs a multi-intent recognition strategy: Considering that civil aviation instructions often contain compound instructions (such as "turn left to heading 330, ascend to 900 meters"), the Multi_Model class applies a Sigmoid activation function to intent_logits. The system sets a threshold (e.g., 0.5), and all intents with probabilities higher than this threshold are marked as valid. The final output intent string is concatenated with "#" (e.g., TurnHeading#Climb), achieving multi-label classification.

[0067] Optionally, the role recognition task is configured as a single-label classification to identify whether the speaker of an instruction is a pilot or an air traffic controller, specifically including: Input the sentence-first feature vector output by the BERT encoder into the role classification layer, and output the role label logical value.

[0068] Apply the Softmax activation function to the logical values ​​of the role labels and select the category index with the highest probability as the predicted speaker role label; where the speaker role label is either pilot or controller.

[0069] The speaker's role label is associated with and stored with the intent classification results.

[0070] In one feasible implementation, when the joint reasoning module processes single intent recognition, it applies the torch.softmax function to process the role's logical value; directly selects the role index with the highest logical value as the unique predicted role; and uses the role tag list to map the index to a text tag, which is then output as a single string.

[0071] The role recognition of this invention adopts a single intent recognition strategy: understanding the context of civil aviation instructions is crucial, because the same text (such as "altitude 900") is an "instruction" when spoken by an air traffic controller, but a "repetition" when spoken by a pilot. The Model class applies the Softmax activation function to intent_logits and directly selects the category index with the highest probability as the speaker role; the final output role string realizes single-label classification.

[0072] Optionally, the slot filling task further includes a conditional random field decoding step: The sequence feature vector output by the BERT encoder is input into the slot filling layer, which outputs the logical value of the slot label corresponding to each token.

[0073] In response to the model configuration parameters being true, the slot label logic value output by the slot filling layer is used as the emission score input to the CRF (Conditional Random Field) layer.

[0074] By combining the label transition matrix learned from the CRF layer, the Viterbi algorithm is used to calculate the globally optimal slot label sequence.

[0075] The globally optimal slot label sequence is output as the entity slot sequence, and the specific entity content is parsed according to the BIO (Begin-Inside-Outside) annotation specification.

[0076] In one feasible implementation, the entity types in the entity slot sequence include: airline name, flight number, runway number, parking stand number, altitude value, heading value, speed value, and waypoint name, etc.

[0077] When processing slot filling results, the joint inference module uses all_slot_label_mask to filter out special tokens (,) and sub-token filling positions in the input sequence; it only retains the predicted label of the first token corresponding to the original input word, ensuring that the length of the output slot list is consistent with the length of the original word list after word segmentation.

[0078] The slot filling and CRF decoding of this invention: To ensure the validity of slot label sequences (e.g., in BIO labeling, the I-Tag must immediately follow the B-Tag), the system can optionally integrate a conditional random field. If args.use_crf is true, the model uses the transition matrix of the CRF layer to perform Viterbi decoding on slot_logits and outputs the globally optimal path; otherwise, it only takes argmax at each position. This mechanism significantly reduces the generation of illegal label sequences.

[0079] Furthermore, the loss function of the Joint BERT model is constructed as follows: (1) In the formula, For intent recognition tasks, use binary cross-entropy loss (for multi-label classification) or cross-entropy loss (for single-label classification). The CRF loss or cross-entropy loss for slot filling tasks. To adjust the hyperparameter coefficients for the weights of the two tasks.

[0080] S5. Based on predefined role-based access control logic, tasks in the annotation task queue are distributed to the corresponding terminals for manual verification and annotation, and manual annotation feedback data is obtained.

[0081] In one feasible implementation, to address the collaboration and security issues in large-scale annotation, this invention implements a sophisticated RBAC (Role-Based Access Control) system based on MongoDB, with the core logic residing in the DB_Controller class. RBAC distributes samples from the annotation task queue to corresponding terminals for manual verification and annotation, where access control restricts specific users' read and write operations on specific data ID ranges.

[0082] Optionally, the predefined role-based access control logic in S5 distributes samples from the annotation task queue to the corresponding terminals, including: Upon receiving a data access request, extract the user credentials, task name, and target data ID from the request.

[0083] The permission verification function is called first to verify the validity of the user's credentials.

[0084] In response to a user role attribute of "admin (administrator)", grant unconditional access to all tasks and all data IDs.

[0085] In response to a user's role attribute being "labeler", iterate through the task list in that user's configuration.

[0086] Verify that the requested task name exists in the task list.

[0087] Verify whether the data ID of the request falls within the closed interval ([start_id, end_id]) assigned to the user by the task.

[0088] Access or modification permissions are granted only if the task name matches and the ID is within the range; otherwise, the request is rejected and an error message is returned.

[0089] In one feasible implementation, the system defines two roles with distinctly different permissions: Administrator (Admin): God's-eye view. In the check_task authentication function, if the user role is "admin", the system directly returns "access", allowing them to access data for any task and any ID. Administrators have exclusive permissions for user management, global task creation, and data clearing.

[0090] Labeler: Restricted perspective. Their permissions are strictly limited to a specific ID range for a specific task. Each labeler maintains a list of tasks in their database document, in the form of [{'taskname': 'task1', 'ids':},...].

[0091] Furthermore, the system enforces a check_task(name, task, id) check before any data operation: Authentication: First, verify the user's password hash via check_labeler.

[0092] Task matching: Traverse the user's task list and find the target task.

[0093] Range locking: The core innovation. The system checks whether the data ID of the requested operation falls within the assigned closed interval [start, end]. If the ID is out of bounds (e.g., assigned 0-100, and the user attempts to modify 101), the system rejects the operation and returns an error. This ID-based sharding mechanism fundamentally prevents data conflicts during concurrent annotation by multiple users.

[0094] Furthermore, the access control logic also includes an automatic task distribution step triggered by the administrator; the manage_task function implements automated workflow allocation. The administrator uploads the raw data file (such as Excel), parses and cleans the data, the system reads the file, calculates the total number of rows N, and assigns a globally unique, incrementing ID to each instruction.

[0095] Get the list of currently active labelers and calculate the average number of labels assigned per person: step = N / / num_labelers.

[0096] Calculate the quotient of the total data volume and the number of annotators to determine the basic allocation number for each annotator.

[0097] Iterate through the list of annotators, calculate and assign the start ID and end ID to each annotator sequentially, ensuring that the ID ranges are continuous and non-overlapping. Automatically calculate and update the IDs range for each annotator: User A gets [0, step], User B gets […]. And so on.

[0098] Write the calculated task name and its corresponding ID range into the user configuration document in the database to complete the task binding.

[0099] The interactive terminal allows annotators to mark data that is difficult to judge as "questionable". "Questionable" data will be automatically excluded from JSON and NLP (Natural Language Processing) training data files in the subsequent data export module and will only be retained in the Excel report for expert review.

[0100] The administrator maintains a global mapping table in the labelerlist to enable visual monitoring of task progress. This allows administrators to view the global task progress by: traversing all defined tasks; for each task, calculating the total amount of data and the amount of labeled data; traversing all labelers assigned to the task, calculating the total amount of data and the amount of labeled data within the ID range for each labeler, and generating a fine-grained progress report.

[0101] When the database access control module executes the administrator's deletion of annotators, it includes security verification logic: it counts the total number of users with the "admin" role in the current system; if the total number is less than or equal to 1, it refuses to execute the deletion operation to prevent the system from becoming unusable due to the lack of an administrator.

[0102] The method also includes: The user authentication unit is used to verify the user's login password based on hash algorithms such as PBKDF2.

[0103] The task management unit is used to respond to administrator commands, parse the uploaded dataset file and mathematically divide it into ID ranges according to the number of annotators, ensuring that the task configurations for persistent storage do not overlap.

[0104] The atomic operation verification unit is used to automatically invoke permission check logic before each database find or update operation to intercept illegal data requests.

[0105] The data2excel function is executed to export the "marked" and "questionable" data to an Excel report. The multi-intent field is flattened into a string separated by " / ", and the "questionable" status is written to the remarks column.

[0106] The data export module is configured to execute the data2json function, export only the "annotated" data as a JSON file, and maintain Unicode encoding to support Chinese characters.

[0107] Execute the serialization export logic, traverse the "annotated" data, and extract word sequences, BIO tag sequences, intent and speaker information respectively to generate the corresponding training set files.

[0108] The labels in the seq.out file generated by the data export module are in BIO format. The generation logic includes: traversing the slot label list of each sample; comparing the current label with the previous label; writing "O" in response to the current label being a non-entity ("O"); writing "I-" prefix followed by the label name in response to the current label being the same as the previous label and not "O"; and writing "B-" prefix followed by the label name in response to the current label being different from the previous label or the previous label being "O".

[0109] The check_task function is used to return a specific error flag string so that the front-end terminal can display the specific reason for permission denial to the user.

[0110] S6. Update the parameters of the joint semantic understanding model based on manually labeled feedback data.

[0111] In one feasible implementation, manually labeled feedback data is received, and the parameters of the JointBERT joint model are updated based on the feedback data.

[0112] Specifically, the data received from manually annotated feedback includes: The front-end interface receives annotation packages for specific ID data submitted by annotators. The annotation packages include a revised list of intents, a speaker identifier, and a sequence of slot labels in BIO format.

[0113] Before performing a database update operation, force the permission verification function to be called again to prevent unauthorized modification.

[0114] Update the intent, sender, and slots fields of the corresponding document in the database, and change the status label from "unlabeled" to "labeled".

[0115] In response to the annotator's correction of the original instruction text, the original text before correction and the correction timestamp are stored in the history field to preserve data traceability.

[0116] Furthermore, the method also includes steps for exporting and formatting model training data. To support the next round of training of the JointBERT model, the system can convert unstructured documents in the database into standard NLP sequence labeling format: Traverse the database and filter out data samples with status labels of "questionable" or "discarded".

[0117] Filter out the sample set whose status label is "labeled".

[0118] Generate the sequence input file (seq.in): Extract the word field of the slot objects in the samples, concatenate them into a string separated by spaces, and write it to the file. Extract the original words from the slots field of the database and reassemble them into sentences separated by spaces.

[0119] Generate sequence label file (seq.out): This is the core logic for generating BIO labels. The system iterates through the BIO markers in the slot object: if the current one is "O", it outputs "O"; if the current marker is the same as the previous one, it adds an "I-" prefix (Inside); if they are different, it adds a "B-" prefix (Begin), generating the corresponding label sequence and writing it to the file. This logic ensures the accurate representation of entity boundaries.

[0120] Generate intent label file (label.txt): Extract the intent field. For multi-intent samples, use the "#" connector to connect the multi-intent labels to generate a composite label (e.g., atis_flight#atis_airfare) and write it to the file.

[0121] Generate speaker file: Export role information independently to assist in training the role recognition model.

[0122] Excel Export: Used for manual review. The system filters out "annotated" and "questionable" data, flattening the complex JSON structure into a table. In particular, it converts the intent list into a " / "-separated string and explicitly writes the "questionable" label into the remarks column, making it easier for experts to quickly locate difficult samples.

[0123] JSON Export: Used for system backup or migration. This function automatically removes all samples marked as "questionable," retaining only high-quality, confidently labeled data to ensure the purity of the training set.

[0124] The Joint Inference module supports iterative training: after each round of labeling tasks is completed, the JointBERT model is fine-tuned using the newly exported seq.in, seq.out, and label.txt files, and the updated model is used for the preprocessing of the next batch of unlabeled data.

[0125] The saveJson function is configured to disable ASCII escaping to ensure that readable Chinese civil aviation instruction text is displayed directly in the exported JSON file.

[0126] This invention utilizes the JointBERT model to classify intents and fill slots for voice commands in the civil aviation field, and optimizes the data annotation and model iteration process through a role-based access control system.

[0127] This invention receives a sequence of civil aviation voice commands to be processed; performs domain-specific text preprocessing, including mapping civil aviation-specific numeric characters to standard Arabic numerals using regular expressions, and performing word segmentation to generate a token sequence; and constructs a joint-based... BERT's joint semantic understanding model encodes token sequences to obtain context-sensitive word vector representations. Based on these word vector representations, it executes intent recognition, role recognition, and slot filling tasks in parallel to obtain predicted intent classification results, speaker role labels, and entity slot sequences. In the data preprocessing stage, it integrates cleaning and digital normalization logic for civil aviation-specific terms. In the model inference stage, it achieves multi-intent recognition by applying the Sigmoid activation function and threshold judgment to intent logic values, and single-intent recognition by applying the Softmax activation function and threshold judgment to role logic values. It also utilizes conditional random field decoding to optimize the slot label sequence. Based on predefined role access control logic, samples in the labeling task queue are distributed to corresponding terminals for manual verification and labeling. Access control restricts specific users' read and write operations on specific data ID ranges, ensuring data security and task isolation. Finally, the model is iteratively fine-tuned based on manually corrected high-quality data (including BIO sequence labels and intent labels).

[0128] The system architecture design of this invention follows the principles of "model-guided, manual correction, and closed-loop iteration." The system mainly comprises three core subsystems: a model inference subsystem, a data management and access control subsystem, and a data interaction and export subsystem. For example... Figure 2 As shown, these subsystems synchronize their states and transfer data through a database, ultimately exhibiting an iterative cycle similar to a "data flywheel." The front-end display interface of this invention is as follows: Figure 3 As shown.

[0129] Experimental procedure: 1. Data Annotation: This joint modeling dataset contains 19,094 annotated corpora, such as... Figure 4 As shown, the results are saved as three files: label, seq.in, and seq.out. The annotation results contain a total of 126 slots and 282 intents, including multiple intents.

[0130] 2. Joint Modeling Pre-training: This training model used 17,399 corpora, with 16,494 corpora in the training set and 905 corpora in the test set. Two models were ultimately trained: one for intent recognition and slot filling, and the other for role recognition. The final results are as follows: Figure 5 As shown.

[0131] 3. Experimental Improvements and Results: After a series of improvements to the model and fine-tuning of its parameters to better suit the task of air traffic control instruction analysis, a total of 5 new models were trained, such as... Figure 6 As shown.

[0132] This training model used 18,287 corpora, of which the training set contained 16,425 corpora, the test set contained 951 corpora, and the validation set contained 911 corpora. The final result is as follows: Figure 7 As shown.

[0133] Based on the final experimental results, this invention selected JointBERT joint modeling and added a conditional random field to the slot filling, thereby realizing intent recognition, role recognition and slot filling.

[0134] This invention utilizes a BERT-based joint modeling model to perform joint intent recognition, role recognition, and slot filling tasks for civil aviation voice commands. It provides a complete technical solution for the automated production of large-scale, high-quality semantically labeled data in the civil aviation field, and is of great value in promoting the development of applications such as intelligent speech recognition, air traffic control automation, and pilot assisted training in civil aviation.

[0135] In this embodiment of the invention, by constructing a complete system including Joint BERT inference, RBAC access control, and automated data flow, the processing of civil aviation voice commands has been transformed from a "rough manual annotation" mode to a "semi-automatic human-machine coupling" mode. The model solves the challenges of recognizing multiple intents, single intents, and complex entities through Sigmoid, Softmax, and CRF technologies, while the accompanying data management system addresses security and efficiency issues in large-scale collaboration through strict segmented access control. This overall solution significantly reduces data production costs while improving recognition accuracy, demonstrating extremely high industrial practical value.

[0136] Figure 8 This is a block diagram illustrating an interactive device for service consultation according to an exemplary embodiment. The device is used in an interactive method for service consultation. (Refer to...) Figure 3 The device includes a data receiving module 310, a data preprocessing module 320, an encoding module 330, a joint inference module 340, a database access control module 350, and a data export module 360. Among them:

[0137] The data receiving module 310 is used to receive the text sequence of civil aviation voice commands to be processed.

[0138] The data preprocessing module 320 is used to perform domain-specific text preprocessing on the civil aviation voice command text sequence to obtain the Token sequence.

[0139] The encoding module 330 is used to encode the token sequence through the encoder layer of the pre-trained joint semantic understanding model to obtain a context-related word vector representation; wherein, the joint semantic understanding model is constructed based on the encoder representation model of the joint bidirectional transformer.

[0140] The joint reasoning module 340 is used to classify the word vector representation by intent through the intent classification layer of the joint semantic understanding model to obtain the predicted intent classification result; to classify the word vector representation by role through the role classification layer to obtain the predicted speaker role label; to fill the word vector representation by slots through the slot filling layer to obtain the predicted entity slot sequence; and to construct a labeling task queue based on the intent classification result, speaker role label and entity slot sequence.

[0141] The database permission control module 350 is used to distribute tasks in the annotation task queue to the corresponding terminals for manual proofreading and annotation based on predefined role permission control logic, and obtain manual annotation feedback data.

[0142] The data export module 360 ​​is used to update the parameters of the joint semantic understanding model based on manually labeled feedback data.

[0143] In one feasible implementation, the data preprocessing module 320 is configured to load a custom dictionary and regular expression rules to perform word segmentation, special character cleaning, and tensor transformation for sub-word alignment on the input civil aviation voice commands.

[0144] The Joint Inference Module 340 is configured to load pre-trained Joint BERT model weights and perform intent recognition, role recognition, and slot filling in parallel on the pre-processed instruction tensor. During the initialization phase, the Joint Inference Module 340 automatically detects whether the runtime environment supports CUDA and loads the model into GPU memory to accelerate inference; otherwise, it falls back to CPU execution.

[0145] The database access control module 350 is configured to manage user access and modification permissions for data based on a role-based access control model. Permissions are finely divided based on task name and data ID range, and an administrator interface is provided for automatic task splitting and assignment.

[0146] The device also includes a human-interactive terminal, configured to display samples to be reviewed within the assigned range to the annotators, provide a visual BIO tag editing and intent correction interface, and receive submitted data.

[0147] The data export module 360 ​​is configured to provide report export functions in Excel and JSON formats, and to generate a standard sequence labeling dataset containing seq.in, seq.out, and label.txt for further training needs of the JointBERT model.

[0148] This invention builds a Web service interface based on the Flask framework, providing a RESTful API to support interaction between the front-end terminal and the back-end inference module and database control module.

[0149] The database uses MongoDB, which leverages its document-oriented storage features to embed task configurations and permission scopes directly into user documents, achieving integrated storage of data and permission models.

[0150] In this embodiment of the invention, by constructing a complete system including Joint BERT inference, RBAC access control, and automated data flow, the processing of civil aviation voice commands has been transformed from a "rough manual annotation" mode to a "semi-automatic human-machine coupling" mode. The model solves the challenges of recognizing multiple intents, single intents, and complex entities through Sigmoid, Softmax, and CRF technologies, while the accompanying data management system addresses security and efficiency issues in large-scale collaboration through strict segmented access control. This overall solution significantly reduces data production costs while improving recognition accuracy, demonstrating extremely high industrial practical value.

[0151] Figure 9 This is a schematic diagram of the structure of an interactive device for service consultation provided in an embodiment of the present invention, such as... Figure 9 As shown, the interactive device for service consultation may include the above-mentioned Figure 8 The interactive device shown is for service inquiries. Optionally, the interactive device 410 for service inquiries may include a first processor 2001.

[0152] Optionally, the interactive device 410 for service consultation may also include a memory 2002 and a transceiver 2003.

[0153] The first processor 2001, memory 2002, and transceiver 2003 can be connected via a communication bus.

[0154] The following is combined Figure 9 A detailed description of each component of the interactive device 410 used for service consultation is provided below: The first processor 2001 is the control center of the interactive device 410 for service consultation. It can be a single processor or a collective term for multiple processing elements. For example, the first processor 2001 can be one or more central processing units (CPUs), application-specific integrated circuits (ASICs), or one or more integrated circuits configured to implement embodiments of the present invention, such as one or more digital signal processors (DSPs), or one or more field-programmable gate arrays (FPGAs).

[0155] Optionally, the first processor 2001 can perform various functions of the interactive device 410 for service consultation by running or executing software programs stored in the memory 2002 and calling data stored in the memory 2002.

[0156] In a specific implementation, as one example, the first processor 2001 may include one or more CPUs, for example... Figure 9 CPU0 and CPU1 are shown in the diagram.

[0157] In a specific implementation, as one example, the interactive device 410 for service consultation may also include multiple processors, for example... Figure 9 The first processor 2001 and the second processor 2004 are shown in the diagram. Each of these processors can be a single-core processor or a multi-core processor. Here, a processor can refer to one or more devices, circuits, and / or processing cores used to process data (such as computer program instructions).

[0158] The memory 2002 is used to store the software program that executes the present invention, and is controlled by the first processor 2001 to execute it. The specific implementation method can be referred to the above method embodiment, and will not be repeated here.

[0159] Optionally, the memory 2002 may be a read-only memory (ROM) or other type of static storage device capable of storing static information and instructions, random access memory (RAM) or other type of dynamic storage device capable of storing information and instructions, or electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto. The memory 2002 may be integrated with the first processor 2001 or may exist independently, and may be accessed via the interface circuit of the interactive device 410 for service consultation. Figure 9 (Not shown in the image) is coupled to the first processor 2001, and this embodiment of the invention does not specifically limit this.

[0160] The transceiver 2003 is used to communicate with network devices or with terminal devices.

[0161] Alternatively, transceiver 2003 may include a receiver and a transmitter. Figure 9 (Not shown separately). The receiver is used to implement the receiving function, and the transmitter is used to implement the transmitting function.

[0162] Alternatively, the transceiver 2003 can be integrated with the first processor 2001 or exist independently, and can be connected via the interface circuit of the interactive device 410 for service consultation. Figure 9 (Not shown in the image) is coupled to the first processor 2001, and this embodiment of the invention does not specifically limit this.

[0163] It should be noted that, Figure 9 The structure of the interactive device 410 for service consultation shown in the figure does not constitute a limitation on the router. The actual knowledge structure identification device may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0164] Furthermore, the technical effects of the interactive device 410 for service consultation can be referred to the technical effects of the interactive method for service consultation described in the above method embodiments, and will not be repeated here.

[0165] It should be understood that the first processor 2001 in the embodiments of the present invention may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc.

[0166] It should also be understood that the memory in the embodiments of the present invention can be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of random access memory (RAM) are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate synchronous DRAM (DDR SDRAM), enhanced synchronous DRAM (ESDRAM), synchronous linked DRAM (SLDRAM), and direct rambus RAM (DR RAM).

[0167] The above embodiments can be implemented, in whole or in part, by software, hardware (such as circuits), firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0168] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0169] In this invention, "at least one" means one or more, and "more than one" means two or more. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of a single item or a plurality of items. For example, at least one of a, b, or c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be a single item or multiple items.

[0170] It should be understood that, in various embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0171] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0172] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices, apparatuses, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0173] In the several embodiments provided by this invention, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0174] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0175] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0176] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0177] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An interactive method for service consultation, characterized in that, The method includes: S1. Receive the sequence of civil aviation voice command texts to be processed; S2. Perform domain-specific text preprocessing on the civil aviation voice command text sequence to obtain the Token sequence; S3. Encode the token sequence through the encoder layer of the pre-trained joint semantic understanding model to obtain context-related word vector representations; wherein, the joint semantic understanding model is constructed based on the encoder representation model of the joint bidirectional transformer; S4. Classify the word vector representation by intent through the intent classification layer of the joint semantic understanding model to obtain the predicted intent classification result; classify the word vector representation by role through the role classification layer to obtain the predicted speaker role label; fill the word vector representation by slots through the slot filling layer to obtain the predicted entity slot sequence; construct the annotation task queue based on the intent classification result, speaker role label and entity slot sequence. S5. Based on predefined role-based access control logic, tasks in the annotation task queue are distributed to the corresponding terminals for manual proofreading and annotation, and manual annotation feedback data is obtained. S6. Update the parameters of the joint semantic understanding model based on manually labeled feedback data.

2. The interactive method for service consultation according to claim 1, characterized in that, The S2 includes: The civil aviation voice command text sequence is segmented into a token sequence based on a custom dictionary in the civil aviation field. Regular expressions are then used to map the numeric characters in the civil aviation voice command text sequence to standard Arabic numerals, and the standard Arabic numerals are then split into independent characters.

3. The interactive method for service consultation according to claim 1, characterized in that, The intention classification process in S4, which classifies the word vector representation through an intention classification layer to obtain the predicted intention classification result, includes: Input the sentence-first feature vector from the word vector representation into the intent classification layer, and output the intent label logical value; Apply the Sigmoid activation function to the logical values ​​of the intent labels to obtain the probability value for each predefined intent category; The probability value is compared with a preset judgment threshold. Multiple intent categories with probability values ​​greater than the preset judgment threshold are selected. The selected multiple intent categories are connected by a predetermined separator to generate a composite intent string as the predicted intent classification result.

4. The interactive method for service consultation according to claim 1, characterized in that, The step S4 involves classifying the word vector representation by role through a role classification layer to obtain the predicted speaker role label, including: Input the sentence-initial feature vector from the word vector representation into the role classification layer, and output the role label logical value; Apply the Softmax activation function to the logical values ​​of the role labels and select the category index with the highest probability as the predicted speaker role label; where the speaker role label is either pilot or controller. The speaker's role label is associated with and stored with the intent classification results.

5. The interactive method for service consultation according to claim 1, characterized in that, The process in S4, which involves filling slots in the word vector representation using a slot-filling layer to obtain the predicted entity slot sequence, includes: Input the sequence feature vector in the word vector representation into the slot filling layer, and output the slot label logical value corresponding to each token; The slot tag logic value is used as the emission score input to the conditional random field layer to learn the tag transition matrix; Based on the label transition matrix, the Viterbi algorithm is used to calculate the globally optimal slot label sequence as the predicted entity slot sequence.

6. The interactive method for service consultation according to claim 1, characterized in that, The predefined role-based access control logic in S5 distributes tasks in the labeled task queue to the corresponding terminals, including: Based on data access requests submitted by different terminals, extract user credentials, task names, and target data IDs from the data access requests. The permission verification function is called to verify the legitimacy of the user credentials and identify the role attribute of the user credentials. If the role attribute is administrator, access permissions are granted to all tasks and all data IDs. If the role attribute is an annotator, the task list in the terminal is traversed to verify whether the task name exists in the task list and whether the target data ID falls within the closed interval of the terminal. If so, access or modification permissions are granted; otherwise, the request is rejected and an error message is returned.

7. The interactive method for service consultation according to claim 1, characterized in that, The S5-based predefined role-based access control logic distributes tasks in the task queue to the corresponding terminals, and also includes: the administrator triggers automatic task distribution. Among them, the automatic task distribution triggered by the administrator includes: Read the original data source files of the tasks in the labeling task queue, parse and clean the original data source files, and assign a globally unique ID to each task. The IDs of all tasks are incremented. Get the list of active annotators in the current system; Calculate the quotient between the total amount of data in the task and the number of annotators, and determine the basic allocation number for each annotator based on the quotient; Iterate through the list of annotators, calculate and assign the start ID and end ID to each annotator based on the basic allocation quantity, and ensure that the ID ranges are continuous and do not overlap; Write the calculated task name and its corresponding ID range into the user configuration document in the database to complete the task binding.

8. An interactive device for service consultation, said interactive device for service consultation being used to implement the interactive method for service consultation as described in any one of claims 1-7, characterized in that, The device includes: The data receiving module is used to receive the text sequence of civil aviation voice commands to be processed; The data preprocessing module is used to perform domain-specific text preprocessing on civil aviation voice command text sequences to obtain Token sequences; The encoding module is used to encode the token sequence through the encoder layer of the pre-trained joint semantic understanding model to obtain context-related word vector representations; wherein, the joint semantic understanding model is constructed based on the encoder representation model of the joint bidirectional transformer; The joint reasoning module is used to classify the word vector representation by intent through the intent classification layer of the joint semantic understanding model to obtain the predicted intent classification result; to classify the word vector representation by role through the role classification layer to obtain the predicted speaker role label; to fill the word vector representation by slots through the slot filling layer to obtain the predicted entity slot sequence; and to construct a labeling task queue based on the intent classification result, speaker role label, and entity slot sequence. The database access control module is used to distribute tasks in the annotation task queue to the corresponding terminals for manual proofreading and annotation based on predefined role access control logic, and obtain manual annotation feedback data. The data export module is used to update the parameters of the joint semantic understanding model based on manually labeled feedback data.

9. An interactive device for service consultation, characterized in that, The interactive device for service consultation includes: processor; A memory storing computer-readable instructions that, when executed by the processor, implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium contains program code that can be invoked by a processor to execute the method as described in any one of claims 1 to 7.

Images