CN-121996926-A - Interaction method and device for service consultation

CN121996926ACN 121996926 ACN121996926 ACN 121996926ACN-121996926-A

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 steps of executing 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 a joint semantic understanding model, carrying out intention classification on word vector representation through an intention classification layer, carrying out role classification on the word vector representation through a role classification layer, carrying out slot filling on the word vector representation through a slot filling layer, constructing a labeling task queue, distributing tasks to corresponding terminals based on predefined role authority control logic to carry out manual correction and labeling, and updating parameters of the joint semantic understanding model based on feedback data. The method effectively solves the problems of high data labeling cost, difficult semantic analysis of complex instructions and low non-standard semantic recognition rate in the civil aviation field, and remarkably improves the accuracy of voice instruction understanding and the self-adaptive capacity of the system.

Inventors

WU MAN
WANG BING
LIU LILONG

Assignees

北京科技大学
北京科技大学顺德创新学院

Dates

Publication Date: 20260508
Application Date: 20260114

Claims (10)

1. An interactive method for service consultation, the method comprising: S1, receiving a civil aviation voice instruction text sequence to be processed; s2, performing field-specific text preprocessing on the civil aviation voice instruction text sequence to obtain a Token sequence; S3, encoding the Token sequence through an encoder layer of a pre-trained joint semantic understanding model to obtain a word vector representation related to the context, wherein the joint semantic understanding model is constructed based on an encoder representation model of a joint bidirectional transducer; S4, carrying out intention classification on the word vector representation through an intention classification layer of the joint semantic understanding model to obtain a predicted intention classification result, carrying out role classification on the word vector representation through a role classification layer to obtain a predicted sender role label, carrying out slot filling on the word vector representation through a slot filling layer to obtain a predicted entity slot sequence, and constructing a labeling task queue according to the intention classification result, the sender role label and the entity slot sequence; s5, distributing the tasks in the labeling task queue to corresponding terminals for manual correction and labeling based on the predefined role authority control logic to obtain manual labeling feedback data; And S6, updating parameters of the joint semantic understanding model based on the manual annotation feedback data.
2. The interaction method for service consultation of claim 1, wherein the S2 includes: word segmentation processing is carried out on the civil aviation voice command text sequence according to the civil aviation field custom dictionary to obtain a Token sequence, digital reading characters in the civil aviation voice command text sequence are mapped into standard Arabic numerals by using a regular expression, and the standard Arabic numerals are segmented into independent characters.
3. The interactive method for service consultation according to claim 1, characterized in that the intention classification of the word vector representation by the intention classification layer in S4, obtaining a predicted intention classification result, includes: Inputting sentence head feature vectors in the word vector representation to an intention classification layer, and outputting an intention label logic value; Applying a Sigmoid activation function to the intention tag logic value to obtain a probability value for each predefined intention category; Comparing the probability value with a preset judging threshold value, selecting a plurality of intention categories with probability values larger than the preset judging threshold value, connecting the selected intention categories through a preset separator, and generating a composite intention character string as a predicted intention classification result.
4. The interactive method for service consultation according to claim 1, wherein the step of performing character classification on the character vector representation through the character classification layer in S4 to obtain a predicted speaker character tag includes: inputting sentence characteristic vectors in the word vector representation to a character classification layer, and outputting character label logic values; applying a Softmax activation function to the character tag logic value, and selecting a category index with the highest probability as a predicted sender character tag, wherein the sender character tag is a pilot or a controller; and storing the caller role label and the intention classification result in an associated mode.
5. The interactive method for service consultation according to claim 1, wherein the step of slot filling the word vector representation through the slot filling layer in S4 to obtain a predicted entity slot sequence includes: inputting the sequence feature vector in the word vector representation to a slot filling layer, and outputting a slot label logic value corresponding to each Token; Inputting the logic value of the slot label as a transmitting score to a conditional random field layer, and learning a label transfer matrix; based on the label transfer matrix, calculating a globally optimal slot label sequence by using a Viterbi algorithm to serve as a predicted entity slot sequence.
6. The interaction method for service consultation according to claim 1, wherein the distributing the tasks in the labeling task queue to the corresponding terminals based on the predefined role authority control logic in S5 includes: based on data access requests provided by different terminals, extracting user credentials, task names and target data IDs in the data access requests; And if the role attribute is a label person, traversing a task list in the terminal, verifying whether a task name exists in the task list and verifying whether a target data ID is in a closed zone of the terminal, if so, granting access or modifying permission, otherwise, rejecting the request and returning an error prompt.
7. The interactive method for service consultation according to claim 1, wherein the task in the labeling task queue is distributed to the corresponding terminal based on the predefined role authority control logic in S5, further comprising an administrator triggering automatic task distribution; wherein, the automatic distribution of task is triggered by the administrator, including: reading an original data source file marking tasks in a task queue, analyzing and cleaning the original data source file, and distributing globally unique IDs for each task, wherein the IDs of all the tasks are increased; Acquiring an active annotator list in a current system; calculating the quotient of the total data volume of the task and the number of the annotators, and determining the basic allocation number of each annotator according to the quotient; Traversing the label list, calculating and distributing the starting ID and the ending ID of each label according to the basic distribution quantity, and ensuring that the ID intervals are continuous and non-overlapping; And writing the calculated task name and the corresponding ID range interval into a user configuration document in a database to complete task binding.
8. An interaction means for service advisory for implementing an interaction method for service advisory as claimed in any one of claims 1-7, the means comprising: the data receiving module is used for receiving the civil aviation voice instruction text sequence to be processed; the data preprocessing module is used for executing field-specific text preprocessing on the civil aviation voice instruction text sequence to obtain a Token sequence; The coding module is used for coding the Token sequence through an encoder layer of a pre-trained joint semantic understanding model to obtain a word vector representation related to the context, wherein the joint semantic understanding model is constructed based on an encoder representation model of a joint bidirectional transducer; The joint reasoning module is used for carrying out intention classification on the word vector representation through an intention classification layer of the joint semantic understanding model to obtain a predicted intention classification result; performing role classification on the word vector representation through a role classification layer to obtain a predicted sender role label, performing slot filling on the word vector representation through a slot filling layer to obtain a predicted entity slot sequence, and constructing a labeling task queue according to an intention classification result, the sender role label and the entity slot sequence; The database permission control module is used for distributing the tasks in the marking task queue to the corresponding terminals for manual correction and marking based on the predefined role permission control logic to obtain manual marking feedback data; And the data deriving module is used for updating parameters of the joint semantic understanding model based on the manual annotation feedback data.
9. An interactive apparatus for service consultation, characterized in that the interactive apparatus for service consultation comprises: A processor; A memory having stored thereon computer readable instructions which, when executed by the processor, implement the method of any of claims 1 to 7.
10. A computer readable storage medium having stored therein program code which is callable by a processor to perform the method of any one of claims 1 to 7.

Description

Interaction method and device for service consultation Technical Field The invention relates to the technical fields of computers, artificial intelligence and civil aviation traffic management, in particular to an interaction method and device for service consultation. Background In civil aviation communication, the voice command between the pilot and the air traffic controller has the characteristics of high structuring, dense technical terms, strict semantics and the like. In order to construct the civil aviation intelligent voice understanding system, a large number of voice instructions need to be marked, including intention recognition, role judgment and semantic slot filling. The traditional labeling mode relies on manual sentence-by-sentence dictation and labeling, and has the advantages of low efficiency, high cost and poor consistency. The existing automatic labeling method is mostly based on rules or traditional machine learning models, and complex semantics, multi-intention nesting and technical terms in civil aviation instructions are difficult to process. Although pre-trained language models such as BERT (Bidirectional Encoder Representations from Transformers, encoder representation of bi-directional transformers) perform well in the general field, direct application in vertical fields such as civil aviation still suffers from insufficient field adaptation, inaccurate professional slot recognition, etc. Therefore, there is a need for an efficient, accurate, and scalable semiautomatic labeling scheme for civil aviation instructions to improve labeling efficiency and quality, and to provide high quality training data for subsequent speech understanding systems. Disclosure of Invention In order to solve the technical problems of high cost of data labeling in the civil aviation field, difficult semantic analysis of complex instructions and low recognition rate of non-standard languages in the prior art, the embodiment of the invention provides an interaction method and device for service consultation. The technical scheme is as follows: in one aspect, there is provided an interaction method for service consultation, the method being implemented by an interaction device for service consultation, the method comprising: s1, receiving a civil aviation voice instruction text sequence to be processed. S2, performing field-specific text preprocessing on the civil aviation voice instruction text sequence to obtain a Token sequence. S3, encoding the Token sequence through an encoder layer of a pre-trained joint semantic understanding model to obtain a word vector representation related to the context, wherein the joint semantic understanding model is constructed based on an encoder representation model of a joint bidirectional transducer. S4, carrying out intention classification on the word vector representation through an intention classification layer of the joint semantic understanding model to obtain a predicted intention classification result, carrying out role classification on the word vector representation through a role classification layer to obtain a predicted sender role label, carrying out slot filling on the word vector representation through a slot filling layer to obtain a predicted entity slot sequence, and constructing a labeling task queue according to the intention classification result, the sender role label and the entity slot sequence. And S5, distributing the tasks in the labeling task queue to the corresponding terminals for manual correction and labeling based on the predefined role authority control logic, and obtaining manual labeling feedback data. And S6, updating parameters of the joint semantic understanding model based on the manual annotation feedback data. Optionally, S2 includes: word segmentation processing is carried out on the civil aviation voice command text sequence according to the civil aviation field custom dictionary to obtain a Token sequence, digital reading characters in the civil aviation voice command text sequence are mapped into standard Arabic numerals by using a regular expression, and the standard Arabic numerals are segmented into independent characters. Optionally, the intention classification layer in S4 performs intention classification on the word vector representation to obtain a predicted intention classification result, including: The sentence head feature vector in the word vector representation is input to the intention classification layer, and the intention label logic value is output. A Sigmoid activation function is applied to the intent tag logic values to obtain probability values for each predefined intent category. Comparing the probability value with a preset judging threshold value, selecting a plurality of intention categories with probability values larger than the preset judging threshold value, connecting the selected intention categories through a preset separator, and generating a composite intention character string as a pr