Civil aviation passenger oral text standardization agent processing method based on multi-dimensional context perception dynamic adaptation

Abstract

The application discloses a kind of civil aviation passenger oral text standardization intelligent agent processing methods based on multidimensional context perception dynamic adaptation, method includes: obtaining input text is carried out context attention update processing and obtains fragment representation, obtains candidate expression score and screening processing based on candidate expression evaluation system;Select the first K multidimensional context skill unit from multidimensional context skill system as dynamic scheduling and adopt multiple evidence product formula activation score to obtain the activation score of candidate expression to multidimensional context skill unit;Adopt multidimensional comprehensive check gate algorithm to carry out gate evaluation and obtain the gate score of candidate expression, according to gate score, candidate expression is collected and corresponding associated standard expression is generated using constraint generation module.The present application comprehensively evaluates the highest concept probability, probability distribution entropy, overall risk intensity, agent divergence and external evidence support to obtain a determined transformation set for standardized processing, reducing the risk of incorrect decisions.

Description

Technical Field

[0001] This invention relates to the field of intelligent agent collaborative decision-making text standardization processing, and in particular to an intelligent agent processing method for standardizing spoken text of civil aviation passengers based on multi-dimensional context perception and dynamic adaptation. Background Technology

[0002] Research on the normalization of non-standard text has discussed the identification and conversion of non-standard expressions such as numbers, abbreviations, and colloquialisms, providing a basic reference for the standardization of passenger spoken text. Intent recognition and slot filling methods can identify business intent and key information from user input, but their output is usually mainly intent labels and slots, making it difficult to directly generate standard civil aviation expressions. Pre-trained language models can learn contextual semantic representations, providing a foundational capability for candidate expression recognition and semantic matching.

[0003] In scenarios such as civil aviation customer service, airport inquiries, airline app online consultations, self-service inquiry terminals, customer service work order processing, and complaint feedback platforms, the text entered by passengers is often characterized by colloquialisms, regionalisms, incomplete expressions, and context-dependent business meanings. Differences in individual education levels and cognitive abilities also affect the differences in colloquial expression. Retrieval-enhanced generation methods can incorporate external knowledge into the text generation process, helping to alleviate the problem of relying solely on model parameters for knowledge judgment. However, civil aviation passenger colloquial texts contain a large number of expressions related to regions, airports, routes, airline rules, and time events. Simply searching external materials is still insufficient to determine whether a particular expression can be directly replaced with a standard concept. This type of text includes general inquiries as well as different expression types such as feedback, requests, complaints, and suggestions. Although all text types originate from passengers' natural language input, their processing objectives are not entirely the same: Inquiry texts typically focus on business rules or procedures, such as "Can an elderly person with a fever still board the plane?" or "What should I do if my luggage is delayed due to a typhoon?"; Feedback texts usually describe service situations that have already occurred, such as "No one explained the situation after the flight was canceled" or "The system can't find the record after the rebooking"; Request texts usually contain clear processing requests, such as "Requesting accommodation to be arranged as soon as possible" or "Hoping for a refund of the rebooking price difference"; Complaint texts usually convey dissatisfaction and a sense of responsibility, such as "The two airlines are passing the buck, causing me to wait a long time at the airport"; Suggestion texts usually offer suggestions for process improvement, such as "Suggesting that a unified SMS be sent after a flight cancellation to explain the rebooking and accommodation arrangements." Therefore, when standardizing text, the system needs not only to identify the colloquial expressions used by passengers but also to determine the corresponding business actions, processing targets, and risk boundaries based on the text type. These differences in expression are not only differences in wording but also affect the system's judgment of business targets, business actions, regional conditions, route types, and risk boundaries. Standardizing spoken texts of civil aviation passengers cannot rely solely on keyword replacement; it also requires comprehensive judgment based on text type, regional conditions, time events, business scenarios, and risk boundaries.

[0004] The intelligent agent for standardizing spoken text among civil aviation passengers processes passenger input content that has already been formed into text format. This includes text from airline apps for online inquiries, airport website messages, self-service inquiry terminals, human customer service systems, customer service work orders, complaint feedback platforms, consumer service platforms, and transcripts transcribed from airport hotlines. Research on the integration of language models with external environments, knowledge, and tools provides a reference for knowledge retrieval and action execution in complex tasks. Tool-invoking language models further illustrate that the model can select external tools according to task needs and utilize the tool results to complete subsequent processing. However, the standardization of spoken text among civil aviation passengers does not require the invocation of general tools themselves, but rather a controllable Skill scheduling and standard concept discrimination process centered around the boundaries of civil aviation business, regional differences, cognitive differences, and risk conditions.

[0005] Standardization of spoken texts by civil aviation passengers refers to converting colloquial, regional, or non-standardized expressions into standard expressions that conform to the expression habits and business boundaries of civil aviation, while preserving the original intent of passengers. Expressions lacking context, with unclear regional conditions, ambiguous business meanings, or unmet risk conditions are left in a state of pending confirmation. A multi-agent collaborative framework can handle complex tasks through the division of labor and cooperation among different agents, providing a system organization reference for candidate expression identification, context judgment, risk review, and constraint generation in this application. Uncertainty calibration research shows that the credibility of results needs to be calibrated during tool use or external capability invocation to reduce the impact of erroneous invocations or judgments.

[0006] In summary, passenger input text in civil aviation passenger service scenarios is generally characterized by colloquialism, regionalism, incomplete expression, and context-dependent business meaning, posing significant challenges to text standardization. Existing research focuses on the identification and conversion of numbers, abbreviations, and colloquial expressions, providing a basic reference for the standardization of colloquial text. Intent recognition and slot filling methods can extract business intent and key information from user input, but their output is mainly intent labels and slots, making it difficult to directly generate standard expressions that conform to civil aviation business expression habits. Pre-trained language models can learn deep contextual semantic representations, providing a basic capability for candidate expression identification and semantic matching, but they lack explicit modeling of specific business boundaries and regional rules. Existing technologies struggle to address the problem of dynamically adapting and controllably standardizing various types of civil aviation texts, such as inquiries, complaints, and requests, while preserving the passenger's original intent, by comprehensively utilizing regional awareness, business context, risk constraints, and multi-evidence fusion. Summary of the Invention

[0007] The purpose of this invention is to provide a standardized intelligent agent processing method for civil aviation passenger spoken text based on multi-dimensional context perception and dynamic adaptation. It constructs a full-link processing flow of candidate identification, skill scheduling, standard discrimination and gating collection, which ensures that different types of text such as inquiries, complaints and requests can accurately retain the original intent of passengers after standardization, and output standardized results that conform to the expression habits and internal boundaries of civil aviation business.

[0008] The objective of this invention is achieved through the following technical solution: A method for standardizing spoken text in civil aviation based on multi-dimensional context awareness and dynamic adaptation. This invention's multi-dimensional context awareness refers to the comprehensive identification of regional or area conditions, airport and route conditions, airline rule conditions, business scenarios, text types, colloquialism, completeness of expression, terminology comprehension bias, and risk boundaries in passenger input text. This aims to determine the standard concepts and replaceable ranges corresponding to candidate expressions in the current civil aviation service scenario. The method includes: S1. Construct a multidimensional contextual skill system containing several multidimensional contextual skill units; obtain the input text and divide it into several segments, and use the contextual attention module to perform contextual attention update processing to obtain segment representations. A candidate expression evaluation system was constructed and candidate expression scores were obtained for each fragment. Candidate expressions are selected based on their scores to obtain a candidate expression set Q; S2. Based on the candidate expressions in the candidate expression set Q, select the top K multidimensional contextual skill units from the multidimensional contextual skill system for dynamic scheduling and obtain candidate expressions using multi-evidence product-based activation scoring. Multidimensional contextual skills Activation score ; S3. Construct a multi-agent system comprising several agents, wherein each agent and multi-dimensional contextual skill unit stores corresponding standard concepts, and candidate expressions are obtained using a reliability-calibrated multi-agent fusion method. Probability belonging to candidate criterion concepts, intelligent agents In candidate expressions The reliability of the multi-agent system is assessed by employing a multi-dimensional comprehensive verification gating algorithm to perform gating evaluation on agent disagreements and obtain candidate expressions. Gating score Based on the gate control score Candidate expressions are categorized into a defined transformation set and a pending confirmation set. The constraint generation module is used to associate the content of the defined transformation set with the corresponding standard expressions, and to generate items to be verified from the candidate expressions in the pending confirmation set.

[0009] To better implement this invention, the constraint generation module is used to generate a corresponding standard expression based on the standard concept, skill unit boundary, and verification rules after the candidate expression passes the gating judgment. The method includes: for the candidate expressions in the determined transformation set C, selecting the standard concept with the highest probability after fusion as the target standard concept; calling the allowed replacement boundary, prohibited replacement rule, and executable verification rule in the context-adapted skill unit corresponding to the standard concept to perform legality verification on the candidate expression; if the candidate expression triggers the prohibited replacement rule or risk condition, it is downgraded to the set to be confirmed U; if the verification passes, the candidate expression is converted into a standard expression according to the colloquial expression to standard concept mapping template in the skill unit, and the contextual components that should not be replaced in the original input text are retained; for the candidate expressions in the set to be confirmed U, candidate standard concepts, missing evidence items, risk warnings, and manual verification paths are generated, and the determined standard expression is generated after the supplementary evidence meets the gating conditions.

[0010] Preferably, in method S1, the multidimensional contextual skill unit in the multidimensional contextual skill system The internal associated storage contains data including passenger spoken expressions, civil aviation standard concepts, applicable regions, airports, airlines or routes, business scenarios, allowed replacement boundaries, prohibited replacement or risk conditions, risk levels, skill history credibility and executable verification rules. The corresponding data is encoded to obtain text vectors, and then the data and multi-dimensional contextual skill units are obtained as comprehensive vector representations.

[0011] Preferably, the candidate expression evaluation system includes skill unit-level coarse screening and data item-level fine screening; the skill unit-level coarse screening selects candidate skill units related to the fragment based on the semantic similarity between the fragment representation and the context-adaptive skill unit comprehensive vector representation, and its skill unit-level candidate score expression is as follows: The data item-level fine screening further determines the candidate expression scores based on the semantic similarity between the fragment representation and the data text vector of the candidate skill unit's internal contextual skill data item. The data item-level candidate scores are as follows: in, For fragment representation, The comprehensive vector for the skill unit adapted to the j-th context. For the text vector of the z-th contextual skill data item in the j-th contextualized skill unit, For business trigger strength, For the completeness of the fragment, Enhance alignment items for knowledge. For noise level, to , to These are the weighting coefficients; the candidate expression set Q-screening method is as follows: when At that time, the fragment Add to the candidate expression set Q; otherwise, do not add; where The threshold for the first candidate identification score. The threshold for the second candidate identification score. The threshold for fragment integrity. This is the upper limit threshold for noise.

[0012] Preferably, candidate expressions in the candidate expression set Q are used. The routing probability of the context-adaptive skill unit is selected as the benchmark, and the expression for the routing probability is as follows: ;in Indicate candidate expression The probability of being routed to the j-th context-adaptive skill unit; This means that the routing scores of each context-adaptive skill unit are normalized and the top K context-adaptive skill units are selected as dynamic scheduling objects. Indicate candidate expression The query vector is jointly encoded by candidate expressions, context, regional route information, business scenario information, text type information, expression level features, and risk warnings; (d) represents the comprehensive vector representation of the j-th context-adaptive skill unit; (d) represents the vector dimension. This indicates the range of regions, airports, airlines, or routes to which the j-th context-adaptive skill unit applies. Indicate candidate expression Regional route compatibility score with the stated region, airport, airline, or route range; This indicates the business scenario bound to the j-th context-adaptive skill unit. Indicate candidate expression Business scenario matching score between the business scenario and the business scenario; This indicates the text type to which the j-th context-adaptive skill unit applies, including inquiries, feedback, requests, complaints, or suggestions. Indicate candidate expression Text type matching score between the text types; This represents the expression level feature corresponding to the j-th context-adaptive skill unit. The expression level feature includes the degree of colloquialism, expression completeness, terminology comprehension deviation, degree of missing reference, or cognitive expression features. Indicate candidate expression The expression level matching score between the expression level features; This represents the allowed replacement boundary of the j-th context-adaptive skill unit. Indicate candidate expression The degree to which the allowed replacement boundary is met; This represents the rule prohibiting substitution for the j-th context-adaptive skill unit. Indicates the risk level. Indicate candidate expression The intensity of risk triggering a rule prohibiting replacement or a risk level; This represents the credibility of the skill history of the j-th context-adapted skill unit; to These represent the weighting coefficients.

[0013] Preferably, the activation score in the multi-evidence product activation score The expression is as follows: ,in To construct a complete set of evidence types, the complete set of evidence types includes semantic evidence, regional route evidence, business scenario evidence, text type evidence, expression level evidence, replacement boundary evidence, risk evidence, and historical credibility evidence; For multidimensional contextual skills units The corresponding overall data, Present evidence Individual scores; Present evidence The degree of importance in the current candidate expressions; For smoothing terms, For risk mitigation, For the rule of prohibiting substitution Risk level As for the intensity of risk triggering, This represents the risk mitigation coefficient.

[0014] Preferably, the candidate expression This belongs to standard concepts or related explanations. The probability expression is as follows: , Indicates candidate expression after fusion Belongs to the standard concept The probability, The total number of agents. Represents the set of candidate criteria concepts. Represents intelligent agents The given standard concept of probability, For dynamic reliability weights, the numerator The predictive probability of an agent is based on its own reliability. Multiply the exponents together, and the denominator... To traverse all candidate criterion concepts Summing the products together.

[0015] Preferably, the intelligent agent In candidate expressions Reliability The expression is as follows: , Represents intelligent agents In the current business scenario The historical accuracy rate is as follows. Represents intelligent agents In the current business scenario Historical accuracy rate; Represents intelligent agents Under the current risk type The historical error rate, Represents intelligent agents Under the current risk type The historical error rate; Represents intelligent agents Consistency with results verified by other intelligent agents or humans, Represents intelligent agents Consistency with results verified by other intelligent agents or humans; , These are the reliability calibration coefficients.

[0016] Preferably, the candidate expression Gating score The expression is as follows: , The highest standard concept probability, This represents the probability difference between the highest and second-highest concepts. Information entropy represents the probability distribution of standard concepts. Indicates the overall risk intensity. Indicates the degree of divergence among multiple agents. Indicates the degree of support from external evidence. ~ They represent the weighting coefficients, and σ represents... function.

[0017] Preferably, the method for determining the conversion set and the set to be confirmed is as follows: when At that time, candidate expressions Set up as a definite transformation set Otherwise, they are grouped into a pending confirmation set. The threshold for gating scores. As the risk intensity threshold, The threshold for divergence is set; the constraint generation module is used to associate the candidate expressions in the determined transformation set C with the standard expressions; for the candidate expressions in the set to be confirmed U, the items to be verified, candidate standard concepts, missing evidence items and risk warnings are generated, and the standard expressions are associated with the supplementary evidence after the gate conditions are met; if the gate conditions are still not met after the supplementary evidence is provided, the state of pending confirmation is maintained.

[0018] Compared with the prior art, the present invention has the following advantages and beneficial effects: (1) The present invention adopts a reliability calibration multi-agent fusion method. By assigning dynamic reliability weights based on historical accuracy, error rate and consistency calculation to different agents, and weighted fusion of the prediction probabilities of each agent, a high-reliability standard concept classification result is obtained. A multi-dimensional comprehensive verification gating algorithm is designed to comprehensively evaluate the highest concept probability, probability distribution entropy, overall risk intensity, divergence degree between agents and external evidence support. The results are reliably collected into a set of confirmed transformations and a set of pending confirmations. Content with clear facts is directly standardized, while content with insufficient evidence or high risk is retained in a pending confirmation state, which reduces the business risks of wrong decision-making and unauthorized judgment.

[0019] (2) This invention runs through the entire chain of candidate identification, skill scheduling, standard discrimination and gating collection, ensuring that different types of texts such as inquiries, complaints and requests can accurately retain the original intent of passengers after standardization, and output standardized results that conform to the expression habits and internal boundaries of civil aviation business. For ambiguous expressions caused by regional differences, missing context or unmet risk conditions, this invention will not forcibly generate a definite conclusion, but will standardize the state of pending confirmation, so that the final standardized text has significant advantages in business compliance, operability and risk controllability.

[0020] (3) This invention uses candidate expressions as a benchmark to calculate routing probabilities by combining multiple factors such as semantic matching, regional / route adaptation, business scenario matching, replacement boundary satisfaction, and risk trigger intensity. It introduces multi-evidence product activation scoring and risk suppression terms to dynamically schedule and activate skills. It can accurately and controllably schedule the top K most relevant multi-dimensional contextual skill units according to the current input context, regional conditions, and risk boundaries, effectively avoiding incorrect standard concept replacement when conditions are not met or there is a high risk, and greatly improving the accuracy and security of processing. This invention constructs a multi-dimensional contextual skill system containing rich business rules, regional applicability, and risk conditions. Combined with contextual attention mechanism and candidate expression evaluation system, it can accurately identify candidate fragments of passengers' colloquial and regional expressions. It integrates multiple factors such as business trigger intensity, fragment completeness, and knowledge enhancement alignment to improve the accuracy and business relevance of candidate expression recognition. Attached Figure Description

[0021] Figure 1 This is a flowchart of the civil aviation text standardization intelligent agent processing method of the present invention; Figure 2 This is a flowchart illustrating the method principle of the civil aviation text standardization intelligent agent processing method in the embodiment; Figure 3 This is a schematic diagram illustrating the principle of the civil aviation text standardization intelligent agent processing method in this embodiment. Detailed Implementation

[0022] The present invention will be further described in detail below with reference to embodiments: Example like Figure 1 , Figure 3 As shown, a method for standardizing spoken text of civil aviation passengers based on multi-dimensional context awareness and dynamic adaptation is presented. The method includes: S1. Construct a multidimensional contextual skill system containing several multidimensional contextual skill units. Multidimensional contextual skill units within the multidimensional contextual skill system. The internal associated storage contains data including passenger spoken expressions, civil aviation standard concepts, applicable regions, airports, airlines or routes, business scenarios, permissible substitution boundaries, prohibited substitutions or risk conditions, risk levels, skill history credibility, and executable verification rules. This is part of the multidimensional contextual skill system's multidimensional contextual skill units. (That is, the overall data of the j-th multi-dimensional contextual skill unit, which can also be denoted as multi-dimensional contextual skill unit j. The multi-dimensional contextual skill unit can naturally cover the differences in perception of regions / areas; differences in airports, routes, and airlines; differences in business scenarios; differences in text types, such as inquiries, complaints, requests, grievances, and suggestions; differences in expression caused by cultural level and cognitive level; degree of colloquialism, completeness of expression, misunderstanding of terminology, dialectal expression, etc. The expression of the multi-dimensional contextual skill unit is also called the multi-dimensional contextual skill unit.) is as follows: ,in This represents a multidimensional contextual standard domain skills library. This represents a collection of spoken expressions used by passengers. This indicates the corresponding civil aviation standard concept. Indicates the scope of application for a region, airport, airline, or route. Indicates the business scenario. This indicates that boundary replacement is allowed. This indicates a prohibition on substitution or a risk condition. Indicates the risk level. Indicates the credibility of skill history. This indicates executable verification rules; the corresponding data is encoded to obtain text vectors, and the data (i.e., a comprehensive vector obtained by weighting multiple types of data, which can be dynamically scheduled according to text semantics, regional rules and business boundaries) and the comprehensive vector representations corresponding to multi-dimensional contextual skill units are obtained respectively.

[0023] Obtain the input text X (the input text can be text data entered by the passenger or text data converted from the passenger's speech using a speech-to-text module) and divide it into several segments. Use a context attention module to perform context attention update processing (this can absorb surrounding or contextual semantics) to obtain segment representations. , ,in Let φ(w) represent the segment vector, and φ(w) represent the context word or phrase vector. For attention mapping parameters, This represents the fragment representation after contextual attention update. Able to judge the meaning of a passage by combining it with the context.

[0024] The preferred first technical solution of this invention is: constructing a candidate expression evaluation system and obtaining the candidate expression scores corresponding to the fragments. The candidate expression evaluation system includes skill unit-level coarse screening and data item-level fine screening. Skill unit-level coarse screening selects candidate skill units related to the fragment based on the semantic similarity between the fragment representation and the context-appropriate comprehensive vector representation of the skill unit. The skill unit-level candidate score expression is as follows: The data item-level fine screening further determines the candidate expression scores based on the semantic similarity between the fragment representation and the data text vector of the candidate skill unit's internal contextual skill data item. The data item-level candidate scores are as follows: in, For fragment representation, The comprehensive vector for the skill unit adapted to the j-th context. For the text vector of the z-th contextual skill data item in the j-th contextualized skill unit, For business trigger strength, For the completeness of the fragment, Enhance alignment items for knowledge. For noise level, to , to These are the weighting coefficients; the candidate expression set Q-selection method is as follows: when At that time, the fragment Add to the candidate expression set Q; otherwise, do not add; where The threshold for the first candidate identification score. The threshold for the second candidate identification score. The threshold for fragment integrity. The upper limit threshold is set to the noise level. Candidate expressions are selected based on their scores to obtain the candidate expression set Q.

[0025] In the first preferred technical solution, candidate expressions in the candidate expression set Q are used. The routing probability of the context-adaptive skill unit is selected as the benchmark. The routing probability expression is as follows: ;in Indicate candidate expression The probability of being routed to the j-th context-adaptive skill unit; This means that the routing scores of each context-adaptive skill unit are normalized and the top K context-adaptive skill units are selected as dynamic scheduling objects. Indicate candidate expression The query vector is jointly encoded by candidate expressions, context, regional route information, business scenario information, text type information, expression level features, and risk warnings. (d) represents the comprehensive vector representation of the j-th context-adaptive skill unit; (d) represents the vector dimension. This indicates the range of regions, airports, airlines, or routes to which the j-th context-adaptive skill unit applies. Indicate candidate expression Regional route fit score between the region, airport, airline, or route range; This indicates the business scenario bound to the j-th context-adaptive skill unit. Indicate candidate expression Business scenario matching score between the business scenario and the actual business scenario; This indicates the text type to which the j-th context-adaptive skill unit applies. Text types include inquiries, feedback, requests, complaints, or suggestions. Indicate candidate expression Text type matching score between text types; This represents the expression level feature corresponding to the j-th context-adaptive skill unit. The expression level features include the degree of colloquialism, expression completeness, terminology comprehension bias, degree of missing reference, or cognitive expression features. Indicate candidate expression Expression hierarchy matching score between expression hierarchy features; This represents the allowed replacement boundary of the j-th context-adaptive skill unit. Indicate candidate expression The degree to which the allowed replacement boundary is met; This represents the rule prohibiting substitution for the j-th context-adaptive skill unit. Indicates the risk level. Indicate candidate expression The intensity of risk triggering a rule prohibiting replacement or a risk level; This represents the credibility of the skill history of the j-th context-adapted skill unit; to These represent the weighting coefficients.

[0026] The preferred second technical solution of the present invention is: constructing a candidate expression evaluation system and obtaining the candidate expression scores corresponding to the fragments. The candidate expression evaluation system filters candidate expressions sequentially based on the comprehensive vector representation corresponding to the multidimensional contextual skill unit and the comprehensive vector representation corresponding to the data text vector, and the candidate expression score is calculated accordingly. The expression is as follows: , where σ represents function, This represents the comprehensive vector representation corresponding to the multidimensional contextual skill unit or the comprehensive vector representation corresponding to the data text vector (in this embodiment, it can also be obtained by encoding civil aviation standard terminology, airport service procedures, airline customer service scripts, historical work order corpus and business trigger word set). For context attention modules in fragment representation The corresponding context window, Fragment representation With integrated vector representation semantic similarity, Candidate Expressions Business trigger strength (used to determine whether business triggering factors such as ticketing, check-in, baggage, security check, boarding, transfer, special passenger services, and abnormal services appear in the segment). Indicate candidate expression Fragment completeness (used to determine whether a fragment constitutes a complete and independently comprehensible spoken expression unit at the grammatical and semantic level, rather than just containing scattered words or incomplete phrases with missing references). For knowledge-enhanced alignment items (which can effectively avoid misjudgments caused by relying solely on model semantic similarity; used to determine whether a fragment can form a consistent correspondence with the multidimensional contextual skill base, civil aviation knowledge graph, or retrieved business rules), the multidimensional contextual skill unit of the multidimensional contextual skill system stores multidimensional contextual domain knowledge and candidate expressions. Align with knowledge in the multidimensional context domain; if a match is found, the corresponding knowledge enhancement alignment item will receive bonus points. Indicates the level of noise (used to suppress colloquialisms, unclear references, lack of action, narrow boundaries, or fragments lacking contextual support). ~ These represent the weighting coefficients (which can be obtained by training on historical customer service annotation corpus, and then gradually calibrated through review and feedback). The higher the score, the more closely the fragment matches the candidate expression characteristics for subsequent standard concept mapping. Candidate expressions are selected based on their scores to obtain a candidate expression set Q. In some embodiments, such as... Figures 1-3 As shown, the candidate expression set Q-selection method is as follows: if the fragment represents Corresponding candidate expression score ,and When this happens, the corresponding segment is used as a candidate expression. And add it to the candidate expression set Q, The threshold for candidate identification score. The threshold for fragment integrity. This is the upper limit threshold for noise; if , ,and ,but ;otherwise Candidate expressions Corresponding storage segment representation The comprehensive vector representation of the data text vector and the corresponding data text (including spoken expressions, standard concepts or explanations).

[0027] In the second preferred technical solution, in some embodiments, such as Figures 1-3 As shown, candidate expressions in the candidate expression set Q Multidimensional contextual skill units were selected based on the criteria. Routing probability , ,in To select the top K multidimensional contextual skill units as the function for dynamic scheduling, The query vector representing the candidate expression (can be jointly encoded by the candidate expression, context, regional information, business scenario, and risk warning). For multidimensional contextual skills units The corresponding composite vector, , This refers to the collection of textual materials that constitute this skill unit (including spoken expressions, standard concept descriptions, regional rules, business boundaries, prohibition conditions, and historical customer service examples). This represents the text vector obtained by encoding the data z; Indicates data weights; This indicates normalization processing. Representing candidate representations and integrated vectors Attention matching score between them For vector dimensions, Indicate candidate expression Scope of application of multidimensional contextual skill units The region or route compatibility score. Indicate candidate expression Binding business to multidimensional contextual skill units The score is based on the matching of business scenarios. Indicate candidate expression The set of allowed synonyms in multidimensional contextual skill units The degree to which the boundary replacement is allowed in the middle. Indicate candidate expression The rule of prohibiting substitution in multidimensional context skill units Risk level The intensity of risk triggering. Representing multidimensional contextual skill units The credibility of the skill history; ~ These are the weighting coefficients.

[0028] S2. Based on the candidate expressions in the candidate expression set Q, select the top K multidimensional contextual skill units from the multidimensional contextual skill system for dynamic scheduling and obtain candidate expressions using multi-evidence product-based activation scoring. Multidimensional contextual skills Activation score .

[0029] In some embodiments, such as Figures 1-3 As shown, the activation score in the multi-evidence product activation score The expression is as follows: ,in To construct a complete set of evidence types, the complete set of evidence types includes semantic evidence, regional route evidence, business scenario evidence, text type evidence, expression level evidence, replacement boundary evidence, risk evidence, and historical credibility evidence. For multidimensional contextual skills units Corresponding to the overall data, Present evidence The individual score. Present evidence The importance of the expression in the current candidate expression. This is a smoothing term (used to prevent the overall score from becoming invalid when a certain piece of evidence is zero). This is a risk suppression item (when a candidate expression triggers high-risk conditions such as dangerous goods, refunds and changes, special passengers, compensation, etc., even if the semantic matching is high, the final activation score will be reduced). For the rule of prohibiting substitution Risk level As for the intensity of risk triggering, This represents the risk mitigation coefficient.

[0030] S3. Construct a multi-agent system containing several agents. Each agent and multi-dimensional contextual skill unit stores corresponding standard concepts. A reliability-calibrated multi-agent fusion method is used to obtain candidate expressions. This belongs to standard concepts or related explanations. probability, agent In candidate expressions The reliability of candidate expressions. In some embodiments, candidate expressions This belongs to standard concepts or related explanations. The probability expression is as follows: , Indicates candidate expression after fusion Belongs to the standard concept The probability, The total number of agents. Represents the set of candidate criteria concepts. Represents intelligent agents The given standard concept of probability, For dynamic reliability weights, the numerator The predictive probability of an agent is based on its own reliability. Multiply the exponents together, and the denominator... To traverse all candidate criterion concepts Summing the products together.

[0031] In some embodiments, the agent In candidate expressions Reliability The expression is as follows: , Represents intelligent agents In the current business scenario The historical accuracy rate is as follows. Represents intelligent agents In the current business scenario The historical accuracy rate. Represents intelligent agents Under the current risk type The historical error rate, Represents intelligent agents Under the current risk type The historical error rate. Represents intelligent agents Consistency with results verified by other intelligent agents or humans, Represents intelligent agents Consistency with results verified by other intelligent agents or humans. , These are the reliability calibration coefficients. Specifically, the agent's historical accuracy in the current business scenario is obtained by smoothing the number of correct judgments and the total number of judgments in the same or similar business scenarios; the agent's historical error rate in the current risk type is obtained by smoothing the number of incorrect judgments and the total number of judgments in the same or similar risk types; and the agent's consistency is obtained by the degree of consistency between the probability distribution of candidate standard concepts output by the agent and the output results of other agents or the results of manual review.

[0032] A multi-dimensional comprehensive verification gating algorithm is used to perform gating evaluation of agent disagreements in a multi-agent system to obtain candidate expressions. Gating score Candidate Expressions Gating score The expression is as follows: , The highest standard concept probability, This represents the probability difference between the highest and second-highest concepts. Information entropy represents the probability distribution of standard concepts. Indicates the overall risk intensity. Indicates the degree of divergence among multiple agents. Indicates the degree of support from external evidence. ~ They represent the weighting coefficients, and σ represents... The function comprises: Overall risk intensity RiR_iRi, calculated based on the prohibition substitution rules triggered by the candidate expression, risk level, business sensitivity type, and real-time rule dependency; multi-agent divergence DiD_iDi, calculated based on the differences in the probability distributions of multiple agents' outputs on the candidate standard concept; and external evidence support EiE_iEi, calculated based on the degree of matching between the candidate expression and contextual skill data items, civil aviation business rules, historical work order examples, or external knowledge retrieval results. This is based on the gating score. Candidate expressions are categorized into a determined transformation set and a pending confirmation set. A constraint generation module is used to associate standard expressions with the content of the determined transformation set and to generate verification items for the candidate expressions in the pending confirmation set. The constraint generation module generates corresponding standard expressions based on standard concepts, skill unit boundaries, and verification rules after a candidate expression passes gating. The method includes: for candidate expressions in the determined transformation set C, selecting the standard concept with the highest probability after fusion as the target standard concept; calling the allowed replacement boundary, prohibited replacement rule, and executable verification rule in the context-adapted skill unit corresponding to the standard concept to perform legality verification on the candidate expression; if the candidate expression triggers a prohibited replacement rule or risk condition, it is downgraded to the pending confirmation set U; if the verification passes, the candidate expression is converted into a standard expression according to the colloquial expression to standard concept mapping template in the skill unit, while retaining the contextual components in the original input text that should not be replaced; for candidate expressions in the pending confirmation set U, candidate standard concepts, missing evidence items, risk warnings, and manual verification paths are generated, and a determined standard expression is generated only after supplementary evidence meets the gating conditions. In some embodiments, the workflow of the constraint generation module is as follows: (1) For the candidate expressions $x_i$ in the determined transformation set $C$, select the standard concept with the highest probability after fusion. As a target standard concept; (2) Call The permissible replacement boundary $A_j$ and the executable verification rule $\Gamma_j$ of the multidimensional context skill unit are used to verify the legality of the replacement. If any rule in the prohibited replacement rule $N_j$ is violated, the candidate expression is downgraded to the set to be confirmed. (3) After the verification is passed, the candidate expression $x_i$ is replaced with the standard concept according to the spoken language → standard mapping template stored in the skill unit. The expression form retains the non-substitutional elements in the context; (4) For candidate expressions in the confirmation set, generate a list of items to be verified, which includes a list of candidate standard concepts, missing evidence items, risk warnings and recommended manual verification paths.

[0033] In some embodiments, the method for determining the aggregation of the transformation set and the set to be confirmed is as follows: when At that time, candidate expressions Set up as a definite transformation set Otherwise, they are grouped into a pending confirmation set. The threshold for gating scores. As the risk intensity threshold, The threshold is set to the degree of divergence. The constraint generation module is used to associate the standard expression with the candidate expressions in the determined transformation set C; for the candidate expressions in the set to be confirmed U, the items to be verified, candidate standard concepts, missing evidence items and risk warnings are generated, and the standard expression is associated with the supplementary evidence after the gate condition is met; if the gate condition is still not met after the supplementary evidence is provided, the state of pending confirmation is maintained.

[0034] Taking a passenger input text as an example: "I'm departing from Airport A to Airport B tomorrow morning. My elderly relative has a slight fever. Can they still board? If there's a typhoon and the flight is delayed, what should I do regarding baggage check-in and rebooking? Will I be notified if the gate changes?", the system identifies candidate expressions including: "Can my elderly relative with a slight fever board?", "Flight delayed due to typhoon", "What should I do regarding baggage check-in and rebooking?", and "Will I be notified if the gate changes?". Based on the text content, the system identifies this scenario as departure from Airport A, arrival at Airport B, domestic flight, pre-trip consultation, and involves multiple business points such as health status, severe weather, flight anomalies, baggage services, ticket rebooking, and gate notification. In this example embodiment, the candidate expression recognition agent first extracts multiple candidate expressions from the passenger input; the multi-dimensional context and business context judgment agent confirms that it belongs to the domestic flight pre-trip consultation scenario based on the departure location, arrival location, airport, and travel time; the skill scheduling agent schedules the health and epidemic prevention rules skill, flight anomaly service skill, baggage check-in skill, ticket rebooking skill, and gate notification skill respectively. For questions like "Can an elderly person with a slight fever still board the plane?", the system doesn't directly generate a definitive answer. Instead, it standardizes it as a consultation on the boarding conditions for passengers with fever, and prompts the user to confirm based on the applicable airline rules, airport requirements, or epidemic prevention requirements. For expressions like "Flight delays due to typhoon weather," "How to handle checked baggage and rebooking," and "Will I be notified of gate changes?", the business context in the text is relatively clear, and the standard concept discrimination agent provides high confidence results, which the gating module then incorporates into the deterministic transformation set C.

[0035]

[0036] Among them, it can be identified as a health-related flight consultation, but it does not directly determine whether boarding is allowed; there are significant risks and uncertainties, which conflict with the low divergence condition of set C; it is recommended to change the gating result of this candidate expression to "entering U, pending supplementary rule evidence", and write in the description "passengers with fever are involved in health risk rules and real-time epidemic prevention policies, and the current evidence is insufficient, so they are classified into the set to be confirmed".

[0037] The standardized text output by the system can be: "Passengers inquire about the boarding conditions and confirmation requirements for elderly passengers with fever on flights from Airport A to Airport B; they also inquire about the handling of baggage and ticket rescheduling in the event of flight delays caused by severe weather such as typhoons, and the notification method after a gate change." In this embodiment, the system standardizes multiple inquiries separately. The issue of boarding with a fever is categorized under health and rule verification inquiries; the system only standardizes its business meaning and does not directly determine whether the passenger can ultimately board. Other expressions with clearer business meanings are entered into a definite transformation set C to generate standardized inquiry text.

[0038] Taking a passenger's input text as an example: "I was originally scheduled to fly from Airport M to Airport N on Airline A, but my flight was temporarily canceled and I was reassigned to another flight on Airline B. Upon arrival at the airport, Airline B claimed there was no record of a rebooking, while Airline A claimed that Airline B was responsible, resulting in me being stranded at the airport and incurring food and accommodation expenses. I request an investigation into the responsible party and compensation." the system identified candidate expressions including: "flight temporarily canceled," "reassigned to another flight on Airline B," "Airline B claimed there was no record of a rebooking," "Airline A claimed that Airline B was responsible," "stranded at the airport and incurred food and accommodation expenses," and "request an investigation into the responsible party and compensation." This text falls under the category of complaint feedback and requests, expressing both the actual flight disruption and the request for accountability and compensation. In this case study, the candidate expression recognition agent first identifies candidate expressions such as flight cancellation, cross-airline rebooking, abnormal rebooking records, disputes over the responsible party, passenger delays, accommodation and food expenses, and compensation processing. The multi-dimensional context and business context judgment agent identifies the text as a complaint scenario following a flight anomaly. The Skill scheduling agent schedules the flight anomaly service Skill, the cross-airline rebooking Skill, the order record verification Skill, the responsible party assistance judgment Skill, and the compensation processing Skill. For factual expressions such as "flight temporarily canceled," "assigned to another flight of Airline B," "Airline B claims no rebooking record," and "delayed at the airport with accommodation and food expenses," the system can standardize them as flight cancellation, cross-airline rebooking connection anomaly, abnormal rebooking record, passenger delays, and explanation of expenses. For "responsible party" and "compensation processing," since further verification is required based on the reason for flight cancellation, carrier relationship, rebooking records, on-site handling records, and airline rules, the gating module includes them in the pending confirmation set U.

[0039]

[0040] The standardized text output by the system can be: "A passenger reported that their original flight, operated by Airline A from Airport M to Airport N, was canceled and they were reassigned to another flight operated by Airline B. However, upon arrival at the airport, Airline B found no rebooking record. Airline A and Airline B disagree on the responsibility for subsequent handling, resulting in the passenger being stranded and incurring food and accommodation expenses. The passenger requests verification of the cross-airline rebooking record, confirmation of the responsible party, and a response regarding related expenses and compensation. The responsible party and compensation standards need further verification based on order records, reasons for flight cancellation, rebooking records, airline rules, and on-site handling." This case example demonstrates the application's approach to handling complaint texts: the system can first standardize the established facts and business events, while retaining the responsible party, compensation amount, compensation method, and other content requiring evidentiary support in the pending confirmation set U, avoiding the direct generation of a definitive liability conclusion when evidence is insufficient.

[0041] Taking the handling of ambiguity related to "passing through customs" using a multi-dimensional contextual skill library as an example, passengers' expressions of "passing through customs" show a clear regional and route dependence. For the text "I'm from airport A to airport B, how long does it take to pass through customs?", the multi-dimensional contextual and business contextual judgment agent identifies it as a domestic flight scenario. The skill scheduling agent activates the security check and pre-boarding process skill and suppresses the border inspection or port clearance skill. The standard concept judgment agent categorizes it into expressions related to "the time required for domestic flight security checks and pre-boarding processes". For the text "I'm from airport C to airport D, how long does it take to pass through customs?", the multi-dimensional contextual and business contextual judgment agent identifies it as a regional flight or port-related scenario. The skill scheduling agent activates the border inspection and port clearance skill, and the standard concept judgment agent categorizes it into expressions related to "the time required for regional flight departures or port clearance processes". The risk review agent performs boundary checks on the above results to avoid uniformly replacing "passing through customs" in different route scenarios with the same standard term.

[0042]

[0043] This case example illustrates that this application does not simply replace "pass" with a fixed standard term, but rather dynamically determines its standard meaning based on a multi-dimensional contextual skill library, route type, and business process.

[0044] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for standardizing spoken text of civil aviation passengers using intelligent agents based on multi-dimensional context awareness and dynamic adaptation, characterized in that: the method include: S1. Construct a multidimensional contextual skill system containing several multidimensional contextual skill units; obtain the input text and divide it into several segments, and use the contextual attention module to perform contextual attention update processing to obtain segment representations. A candidate expression evaluation system was constructed and candidate expression scores were obtained for each fragment. Candidate expressions are selected based on their scores to obtain a candidate expression set Q; S2. Based on the candidate expressions in the candidate expression set Q, select the top K multidimensional contextual skill units from the multidimensional contextual skill system for dynamic scheduling and obtain candidate expressions using multi-evidence product-based activation scoring. Multidimensional contextual skills Activation score ; S3. Construct a multi-agent system comprising several agents, wherein each agent and multi-dimensional contextual skill unit stores corresponding standard concepts, and candidate expressions are obtained using a reliability-calibrated multi-agent fusion method. This belongs to standard concepts or related explanations. probability, agent In candidate expressions The reliability of the multi-agent system is assessed by employing a multi-dimensional comprehensive verification gating algorithm to perform gating evaluation on agent disagreements and obtain candidate expressions. Gating score Based on the gate control score Candidate expressions are categorized into a defined transformation set and a pending confirmation set. The constraint generation module is used to associate the content of the defined transformation set with the corresponding standard expressions, and to generate items to be verified from the candidate expressions in the pending confirmation set.

2. The method for standardizing spoken text of civil aviation passengers based on multi-dimensional context awareness and dynamic adaptation as described in claim 1, characterized in that: The constraint generation module is used to generate a corresponding standard expression based on the standard concept, skill unit boundary, and verification rules after the candidate expression passes the gating judgment. The method includes: for the candidate expressions in the determined transformation set C, selecting the standard concept with the highest probability after fusion as the target standard concept; calling the allowed replacement boundary, prohibited replacement rule, and executable verification rule in the context-adapted skill unit corresponding to the standard concept to perform legality verification on the candidate expression; if the candidate expression triggers the prohibited replacement rule or risk condition, it is downgraded to the set to be confirmed U; if the verification passes, the candidate expression is converted into a standard expression according to the colloquial expression to standard concept mapping template in the skill unit, and the contextual components that should not be replaced in the original input text are retained; for the candidate expressions in the set to be confirmed U, candidate standard concepts, missing evidence items, risk warnings, and manual verification paths are generated, and the determined standard expression is generated after the supplementary evidence meets the gating conditions.

3. The method for standardizing spoken text of civil aviation passengers based on multi-dimensional context awareness and dynamic adaptation as described in claim 1, characterized in that: In method S1, the multidimensional contextual skill unit in the multidimensional contextual skill system The internal associated storage contains data including passenger spoken expressions, civil aviation standard concepts, applicable regions, airports, airlines or routes, business scenarios, allowed replacement boundaries, prohibited replacement or risk conditions, risk levels, skill history credibility and executable verification rules. The corresponding data is encoded to obtain text vectors, and then the data and multi-dimensional contextual skill units are obtained as comprehensive vector representations.

4. The method for standardizing spoken text of civil aviation passengers based on multi-dimensional context awareness and dynamic adaptation as described in claim 1, characterized in that: The candidate expression evaluation system includes skill unit-level coarse screening and data item-level fine screening. Skill unit-level coarse screening selects candidate skill units related to the fragment based on the semantic similarity between the fragment representation and the context-adaptive skill unit's comprehensive vector representation. The skill unit-level candidate score expression is as follows: The data item-level fine screening further determines the candidate expression scores based on the semantic similarity between the fragment representation and the data text vector of the candidate skill unit's internal contextual skill data item. The data item-level candidate scores are as follows: in, For fragment representation, The comprehensive vector for the skill unit adapted to the j-th context. For the text vector of the z-th contextual skill data item in the j-th contextualized skill unit, For business trigger strength, For the completeness of the fragment, Enhance alignment items for knowledge. For noise level, to , to These are the weighting coefficients; the candidate expression set Q-screening method is as follows: when At that time, the fragment Add to the candidate expression set Q; otherwise, do not add; where The threshold for the first candidate identification score. The threshold for the second candidate identification score. The threshold for fragment integrity. This is the upper limit threshold for noise.

5. The method for standardizing civil aviation passenger spoken text using intelligent agents based on multi-dimensional context awareness and dynamic adaptation as described in claim 1, characterized in that: Candidate expressions in the candidate expression set Q The routing probability of the context-adaptive skill unit is selected as the benchmark, and the expression for the routing probability is as follows: ;in Indicate candidate expression The probability of being routed to the j-th context-adaptive skill unit; This means that the routing scores of each context-adaptive skill unit are normalized and the top K context-adaptive skill units are selected as dynamic scheduling objects. Indicate candidate expression The query vector is jointly encoded by candidate expressions, context, regional route information, business scenario information, text type information, expression level features, and risk warnings; (d) represents the comprehensive vector representation of the j-th context-adaptive skill unit; (d) represents the vector dimension. This indicates the range of regions, airports, airlines, or routes to which the j-th context-adaptive skill unit applies. Indicate candidate expression Regional route compatibility score with the stated region, airport, airline, or route range; This indicates the business scenario bound to the j-th context-adaptive skill unit. Indicate candidate expression Business scenario matching score between the business scenario and the business scenario; This indicates the text type to which the j-th context-adaptive skill unit applies, including inquiries, feedback, requests, complaints, or suggestions. Indicate candidate expression Text type matching score between the text types; This represents the expression level feature corresponding to the j-th context-adaptive skill unit. The expression level feature includes the degree of colloquialism, expression completeness, terminology comprehension deviation, degree of missing reference, or cognitive expression features. Indicate candidate expression The expression level matching score between the expression level features; This represents the allowed replacement boundary of the j-th context-adaptive skill unit. Indicate candidate expression The degree to which the allowed replacement boundary is met; This represents the rule prohibiting substitution for the j-th context-adaptive skill unit. Indicates the risk level. Indicate candidate expression The intensity of risk triggering a rule prohibiting replacement or a risk level; This represents the credibility of the skill history of the j-th context-adapted skill unit; to These represent the weighting coefficients.

6. The method for standardizing civil aviation passenger spoken text using intelligent agents based on multi-dimensional context awareness and dynamic adaptation as described in claim 5, characterized in that: The activation score in the multi-evidence product activation score The expression is as follows: ,in To construct a complete set of evidence types, the complete set of evidence types includes semantic evidence, regional route evidence, business scenario evidence, text type evidence, expression level evidence, replacement boundary evidence, risk evidence, and historical credibility evidence; For multidimensional contextual skills units The corresponding overall data, Present evidence Individual scores; Present evidence The degree of importance in the current candidate expressions; For smoothing terms, For risk mitigation, For the rule of prohibiting substitution Risk level As for the intensity of risk triggering, This represents the risk mitigation coefficient.

7. The method for standardizing spoken text of civil aviation passengers based on multi-dimensional context awareness and dynamic adaptation as described in claim 1, characterized in that: The candidate expression This belongs to standard concepts or related explanations. The probability expression is as follows: , Indicates candidate expression after fusion Belongs to the standard concept The probability, The total number of agents. Represents the set of candidate criteria concepts. Represents intelligent agents The given standard concept of probability, For dynamic reliability weights, the numerator The predictive probability of an agent is based on its own reliability. Multiply the exponents together, and the denominator... To traverse all candidate criterion concepts Summing the products together.

8. The method for standardizing civil aviation passenger spoken text intelligent agent processing based on multi-dimensional context perception and dynamic adaptation according to claim 1 or 7, characterized in that: The intelligent agent In candidate expressions Reliability The expression is as follows: , Represents intelligent agents In the current business scenario Historical accuracy rate Represents intelligent agents In the current business scenario Historical accuracy rate; Represents intelligent agents Under the current risk type The historical error rate, Represents intelligent agents Under the current risk type The historical error rate; Represents intelligent agents Consistency with results verified by other intelligent agents or humans, Represents intelligent agents Consistency with results verified by other intelligent agents or humans; , These are the reliability calibration coefficients.

9. The method for standardizing spoken text of civil aviation passengers based on multi-dimensional context awareness and dynamic adaptation as described in claim 1, characterized in that: The candidate expression Gating score The expression is as follows: , The highest standard concept probability, This represents the probability difference between the highest and second-highest concepts. Information entropy represents the probability distribution of standard concepts. Indicates the overall risk intensity. Indicates the degree of divergence among multiple agents. Indicates the degree of support from external evidence. ~ They represent the weighting coefficients, and σ represents... function.

10. The method for standardizing spoken text of civil aviation passengers based on multi-dimensional context awareness and dynamic adaptation as described in claim 9, characterized in that: The method for determining the transformation set and the set to be confirmed is as follows: when At that time, candidate expressions Set up as a definite transformation set Otherwise, they are grouped into a pending confirmation set. The threshold for gating scores. As the risk intensity threshold, The threshold for divergence is set; the constraint generation module is used to associate the standard expression with the candidate expression in the determined transformation set C; for the candidate expression in the set to be confirmed U, the items to be verified, candidate standard concepts, missing evidence items and risk warnings are generated, and the standard expression is associated with the candidate expression after the supplementary evidence meets the gating conditions. If the gating conditions are still not met after supplementing the evidence, the status will remain pending confirmation.

Images