AI-based airport intelligent service question and answer method and system

Abstract

The invention discloses an airport intelligent service question answering method and system based on AI, relates to the technical field of artificial intelligence, and comprises the steps of performing combined modeling by using a time sequence and a multi-head attention mechanism, extracting key behavior characteristics, fusing intention recognition and behavior prediction, and improving the answer reasoning accuracy. The method comprises the following steps: selecting a semantic knowledge base, constructing deep semantic representation and accurately matching the deep semantic representation with the semantic knowledge base, extracting high-weight phrases in combination with a sliding window mechanism, generating unique semantic fingerprints by utilizing a Karp-Rabin hash function, improving the speed and accuracy of answer matching, and performing fuzzy control and redundant information elimination on candidate answers by adopting a rough set tolerance model. Semantic similarity judgment is carried out according to the intersection and union set proportion between the upper and lower approximate sets, the problems of semantic drift and repeated candidates in traditional matching are effectively relieved, and through reinforcement learning and A / B testing, the question and answer strategy is continuously optimized, and the response effect of the system and the user satisfaction degree are enhanced.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and particularly to an AI-based intelligent service Q&A method and system for airports. Background Art

[0002] In recent years, with the rapid development of artificial intelligence technology, especially the breakthroughs in the fields of natural language processing and computer vision, intelligent service systems have been widely applied in various scenarios, and intelligent Q&A systems have become an important part. In the specific environment of airports, the application of airport intelligent service Q&A systems has gradually increased, aiming to provide real-time and accurate consultation services for passengers. Existing technologies usually obtain various input data through forms such as speech recognition, image recognition, and text input, respond to various needs of passengers in the airport environment in real time, and use pre-trained models and rule matching methods to provide automated Q&A services, providing a more convenient query channel for passengers.

[0003] Although existing intelligent service Q&A systems can match user queries through certain rules, they are still insufficient in accurately understanding and predicting user needs. First, the existing technologies have weak capabilities in fusing and processing multi-modal data, still relying on single-modal inputs (such as text or speech) for semantic understanding, and failing to fully utilize the auxiliary information of multi-modal data such as images and videos. As a result, when the input data provided by users contains multiple modalities, the system often cannot fully capture the true intentions of users, especially in complex scenarios, and the robustness of traditional methods is poor. Second, existing technologies have limitations in constructing user portraits and providing personalized services. Although existing technologies have started to use user historical data for personalized answers, they all rely on static user data, lack real-time nature, and are difficult to capture the behavioral changes of users during actual interactions in a timely manner, resulting in poor personalized answer effects of the system. In addition, existing technologies still have limitations in generating and optimizing Q&A content. Many existing Q&A systems use simple matching methods based on rules or templates to generate answers. Although they can handle some common questions, they often cannot provide satisfactory answers when facing complex and changing user queries. Therefore, existing intelligent Q&A technologies mainly face problems such as insufficient capabilities in processing multi-modal data, inaccurate user portraits, and insufficient optimization of answer generation. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides an AI-based intelligent service Q&A method and system for airports, which solve the problems of insufficient multi-modal data capabilities, inaccurate user portraits, and insufficient optimization of answer generation.

[0006] To solve the above technical problems, the present invention provides the following technical solutions:

[0007] In a first aspect, the present invention provides an AI-based intelligent service Q&A method and system for airports, which includes

[0008] Receiving and parsing multi-modal data input by a user, and constructing a user profile in real time by combining the user's historical query sequence, browsing behavior, and personalized information;

[0009] The multi-modal data includes text, voice, image, and location information;

[0010] The personalized information includes language preference and user type;

[0011] Based on the constructed user profile, identifying the user's current intention through a temporal attention mechanism model, predicting the future behavior path, and performing TF-IDF weighting and semantic matching through deep learning fusion calculation according to the user input data and the current intention, and obtaining the final answer candidate set through a rough set model and a pruning mechanism;

[0012] Calculating the answer confidence and selecting the answer with the optimal confidence, and continuously optimizing the Q&A strategy by combining reinforcement learning.

[0013] As a preferred solution of the AI-based intelligent service Q&A method for airports according to the present invention, wherein: the receiving and parsing of the multi-modal data input by the user refers to performing noise removal, temporal alignment, and dimensional normalization processing on the multi-modal data, extracting its features and converting them into a unified text semantic representation, and performing semantic alignment based on a multi-modal fusion model with a self-attention network architecture, using a multi-modal encoder to convert the features of different modalities into vector representations, and splicing them to obtain a multi-modal semantic representation vector A'.

[0014] As a preferred solution of the AI-based intelligent service Q&A method for airports according to the present invention, wherein: the real-time construction of the user profile by combining the user's historical query sequence, browsing behavior, and personalized information refers to obtaining the user's historical inquiry text, inquiry timestamp, page access record, and page stay time from the database, and constructing a behavior sequence vector H;

[0015] Weighting the behavior sequence vector H using a time decay function to obtain a final behavior representation vector H';

[0016] Analyzing and obtaining personalized preference features using the basic information provided by the user during registration, and representing the user preference vector B through rule vectorization;

[0017] Splicing the final behavior representation vector H' and the user preference vector B through a multi-layer perceptron to obtain a complete dynamic representation vector U of the user profile.

[0018] As a preferred solution of the AI-based intelligent service Q&A method for airports described in the present invention, wherein: based on the constructed user profile, the user's current intention is recognized through the temporal attention mechanism model, and predicting the future behavior path means splicing the generated dynamic representation vector U of the user profile and the multi-modal semantic representation vector A' into a joint representation vector X;

[0019] Using three groups of weight matrices in the temporal self-attention mechanism to perform a linear transformation on the joint representation vector X to generate a query vector Q, a key vector K, and a value vector V, obtaining the key behavior features within each time step;

[0020] Adopting the multi-head attention mechanism, mapping the query, key, and value vectors of the key behavior features to each head, calculating multiple attentions simultaneously and analyzing them to capture the feature representations in different subspaces, and mapping them to the comprehensive behavior feature MultiHead(Q, K, V);

[0021] Through the captured feature representations in different subspaces, identifying the user's current behavior pattern, and mapping the comprehensive behavior feature MultiHead(Q, K, V) to the corresponding current intention category vector y through a fully connected layer;

[0022] Based on the current intention category vector y, indicating the position in the behavior sequence through the attention mechanism, and using the pointer mechanism in the pointer network to predict and generate the user's next behavior path P next ;

[0023] Combining the predicted next behavior path P next and the current intention category vector y to generate an intention vector y' containing the current user's intention and future behavior prediction information.

[0024] As a preferred solution of the AI-based intelligent service Q&A method for airports described in the present invention, wherein: according to the user input data and the current intention, using deep learning fusion calculation, performing TF-IDF weighting and semantic matching, and obtaining the final answer candidate set through the rough set model and the pruning mechanism means inputting the multi-modal semantic representation vector A' and the future path intention vector y' into the deep learning model for fusion calculation to obtain the final joint semantic representation X';

[0025] Based on the final joint semantic representation X' and the knowledge base for matching, using the BERT tokenization tool to tokenize the final joint semantic representation X' into individual words, annotating the corresponding part of speech, retaining nouns and verbs as candidate keywords, and performing TF-IDF weighting calculation to obtain the TF-IDF weight w of the word i i , calculating the weight of each keyword, and generating a feature keyword sequence according to the weight;

[0026] Set the window size to g, and each time continuously select g keywords as a phrase. Extract all triples from the sequence of feature keywords. Calculate the comprehensive weight of the feature phrase by weighted summing the TF-IDF weights of all keywords in the feature phrase. Sort the feature phrases in descending order according to the comprehensive weight to obtain the set of feature phrases a β ;

[0027] Decompose the set of feature phrases a β into independent words, and use the HowNet semantic knowledge base to calculate the similarity Sim(c1, c2) between each independent word and its lexical semantic concept for matching;

[0028] Set the similarity threshold ψ. When the similarity Sim(c1, c2) is greater than or equal to ψ, they are semantically close enough to be replaced, generating a new semantic feature phrase a β ' containing concepts. Otherwise, no replacement is made;

[0029] Use the Karp-Rabin hash function to convert each semantic feature phrase into a unique and fixed-length fingerprint Hash(a β '), and perform the Karp-Rabin hash function on each phrase in the semantic feature phrase a β ' and map it to a unique digital fingerprint, forming a set of semantic fingerprints F input ;

[0030] Use a two-tower network to encode the final joint semantic representation X' and each knowledge item in the knowledge base, generate a feature vector representation, and combine it with the set of semantic fingerprints F input Calculate the semantic similarity Sim(X', G j );

[0031] Through the calculation of semantic similarity, obtain the candidate answer items in the knowledge base that match the joint semantic representation X', and aggregate them into the candidate answer set W;

[0032] Use the probabilistic tolerance rough set model to further optimize the candidate answer set W. By constructing the upper and lower approximation sets U(S, θ i ), E(S, θ i ), and calculate the tolerance similarity Sim(a β ', W ξ ) according to the sizes of the intersections and unions of the upper and lower approximation sets of the semantic feature phrase a β [[ID=4 + 4]]', W ξ );

[0033] Based on the calculation of tolerance similarity, set the similarity threshold Γ. When the tolerance similarity of the candidate item is greater than or equal to the threshold Γ, the candidate answer item is retained in the candidate answer set W and sorted in descending order. Otherwise, it is screened out from the candidate answer set W;

[0034] From the selected answer candidates, a retrieval-based answer is generated and decoded. During the decoding stage, a dynamic pruning mechanism is used to calculate the pruning decision for each attention channel, eliminating redundant answer candidate decoding attention channels, and outputting the answer candidate set W'.

[0035] The generative model generates φ candidate answers for the final joint semantic representation X'. The Beam Search method is used to select generative candidate answers from the generative model, and the answers in the candidate answer set W' are combined to annotate the source, the semantic similarity between the input and the answer, and the degree of pruning under the current input, and the final answer candidate set is output.

[0036] As a preferred solution of the AI-based intelligent airport service Q&A method described in the present invention, wherein: the calculation of the answer confidence and the selection of the answer with the optimal confidence refer to using the multi-modal semantic representation vector A' to calculate the candidate set in the semantic matching confidence, context consistency score, and generation quality score of the

[0037] The semantic matching confidence encodes A' and using a pre-trained BERT model, and calculates the semantic matching confidence of the th candidate answer using cosine similarity;

[0038] The context consistency score calculates the context consistency score of the th candidate answer using BERTScore;

[0039] The generation quality score uses ROUGE to calculate the overlap degree between the generated text and the reference text;

[0040] Based on the calculation results of the semantic matching confidence, context consistency score, and generation quality score, the comprehensive score of each candidate answer is calculated through weighted sum;

[0041] A threshold q is set, the candidate answers with scores greater than the threshold q are retained and sorted in descending order, and the top C candidate answers with scores are retained as alternative suggestions to be presented to the user, otherwise the user is prompted to re-enter the query.

[0042] As a preferred solution of the AI-based intelligent airport service Q&A method described in the present invention, wherein: the continuous optimization of the Q&A strategy by combining reinforcement learning means that after the confidence screening and answer output are completed, the Q&A interaction is formalized into a Markov decision process modeled by state, action, and reward;

[0043] Optimize the Q&A decision-making process using a policy network, and the policy network outputs an action distribution, and uses proximal policy optimization training to optimize the learnable parameter set in the policy network

[0044] Based on the optimization of the policy network, through an experience replay pool, specifically store high-score feedback trajectories and set a threshold After each interaction, if the reward is greater than the threshold add the current trajectory to the preferred replay pool. Every v times of training, sample a batch of high-quality trajectories from the pool, and use the target reinforcement policy to update the gradient of the policy parameters for the parameter set to perform iterative updates;

[0045] Combine the A / B test mechanism to assign multiple trained policy versions to different user subgroups, collect feedback data of both under the same number of users, and calculate the difference in feedback values between the new policy and the old policy. If the feedback value of the new policy is greater than that of the old policy, switch it online.

[0046] In a second aspect, the present invention provides an AI-based intelligent service Q&A system for airports, including

[0047] a multi-modal data processing module for denoising, time series alignment, and dimensional normalization of multi-modal data, extracting features and converting them into a unified text semantic representation;

[0048] a user profile construction module for constructing a user profile based on historical queries, browsing behaviors, and personalized information, and generating a dynamic representation vector;

[0049] an intention recognition and behavior prediction module for using a temporal self-attention mechanism to recognize the current user's intention and predict the future behavior path based on the user profile and multi-modal semantic representation;

[0050] a semantic matching module for matching through semantic representation with a knowledge base, performing rough set optimization and dynamic pruning, and generating a final answer candidate set;

[0051] an answer optimization and policy module for calculating confidence and optimizing the policy, and optimizing the Q&A decision-making through reinforcement learning to improve the response effect.

[0052] In a third aspect, the present invention provides a computer device, including a memory and a processor, where the memory stores a computer program, and: when the computer program is executed by the processor, any step of the AI-based intelligent service Q&A method for airports described in the first aspect of the present invention is implemented.

[0053] ​Fourthly, the present invention provides a computer-readable storage medium with a computer program stored thereon, wherein: when the computer program is executed by a processor, any step of the AI-based intelligent airport service Q&A method described in the first aspect of the present invention is implemented.

[0054] The beneficial effects of the present invention are as follows: The present invention uses the combined modeling of time series and multi-head attention mechanism to extract key behavior features, integrates intention recognition and behavior prediction, improves the accuracy of answer reasoning, constructs a deep semantic representation and precisely matches it with the semantic knowledge base, combines the sliding window mechanism to extract high-weight phrases, generates a unique semantic fingerprint using the Karp-Rabin hash function, improves the speed and accuracy of answer matching, and uses the rough set tolerance model to control the ambiguity of candidate answers and eliminate redundant information, and judges the semantic similarity based on the intersection and union ratio between the upper and lower approximation sets, effectively alleviating the problems of semantic drift and repeated candidates in traditional matching. Through reinforcement learning and A / B testing, the Q&A strategy is continuously optimized to enhance the response effect of the system and user satisfaction. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention, and those of ordinary skill in the art can obtain other drawings without creative efforts based on these drawings.

[0056] Figure 1 It is a flowchart of an AI-based intelligent airport service Q&A method in Embodiment 1.

[0057] Figure 2 It is a structural diagram of an AI-based intelligent airport service Q&A system in Embodiment 1.

[0058] Figure 3 It is a structural diagram of user portrait construction in Embodiment 1.

[0059] Figure 4 It is a flowchart of semantic matching optimization in Embodiment 1. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0060] In order to make the above-mentioned objects, features, and advantages of the present invention more obvious and understandable, the following detailed description of the specific embodiments of the present invention will be made with reference to the accompanying drawings of the specification.

[0061] Many specific details are set forth in the following description to facilitate a thorough understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein. Those skilled in the art can make similar generalizations without departing from the spirit of the present invention, so the present invention is not limited by the specific embodiments disclosed below.

[0062] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0063] Example 1, with reference to Figures 1 to 4 , which is the first embodiment of the present invention, provides an AI-based airport smart service question-answering method, comprising the following steps:

[0064] S1. Receive and parse multimodal data input by users, and build user profiles in real time based on their historical query sequences, browsing behavior, and personalized information.

[0065] Specifically, receiving and parsing multimodal data input by users means removing noise, time series alignment and dimension normalization of the multimodal data, extracting its features and converting them into a unified text semantic representation. To ensure the comparability between different modalities, a multimodal fusion model based on the self-attention network architecture is used for semantic alignment, and a multimodal encoder is used to convert the features of different modalities into vector representations, which are then spliced together to obtain the multimodal semantic representation vector A'.

[0066] By performing denoising, time series alignment and dimensional normalization on multimodal data, the consistency of the original input signal in the time dimension and numerical space is achieved, thereby effectively eliminating the interference and scale differences between different modal inputs. Through semantic alignment based on the multimodal fusion model of the self-attention network architecture, the implicit semantic relationship between modalities such as text, speech, image, and location information is automatically modeled, which enhances the model's understanding and association capabilities of cross-modal semantics, and compensates for the weak semantic connection and poor robustness of traditional feature splicing methods. The multimodal encoder is used to uniformly map the feature vectors of different modalities, and completes vector-level splicing to generate multimodal semantic representation vectors, realizing high-dimensional fusion expression of user input data in a unified semantic space, and improving the model's representation capabilities in complex cross-semantic scenarios.

[0067] Furthermore, combining the user's historical query sequence, browsing behavior and personalized information to build a user profile in real time means obtaining the user's historical query text, query timestamp, page visit record and page dwell time from the database to capture the user's behavior pattern and build a behavior sequence vector H;

[0068] Use the time decay function to weight the behavior sequence vector H to avoid static user portraits and obtain the final behavior representation vector H';

[0069] The calculation formula is:

[0070]

[0071] Among them, is the behavior time weight of the ζ-th row, Δε ζ is the time difference from the current time, and ζ is the time decay coefficient, which is set through experiments;

[0072] The final behavior representation vector H' is:

[0073]

[0074] Among them, is the maximum number of behaviors, is the behavior time weight of the ζ-th row, h ζ is the behavior vector of the ζ-th row;

[0075] Using the basic information (age, gender, and language preference) registered when the user makes an inquiry, analyze and obtain personalized preference features (the language selected by the user, the preferred service category, and the recently queried information), and vectorize the user preference vector B through rules;

[0076] Concatenate the final behavior representation vector H' and the user preference vector B through a multi-layer perceptron to obtain a complete dynamic representation vector U of the user portrait.

[0077] By obtaining the user's historical inquiry records from the database, the systematic collection and sequence modeling of the user's historical interaction behaviors are realized, and the behavior path and the evolution process of information needs of the user can be restored in the time dimension. By using a time decay function to weight the constructed behavior sequence vector H, the enhanced attention to recent behaviors and the weight adjustment of long-term behaviors are realized, effectively avoiding the problem of static user portraits, enhancing the adaptability of the model to the changes of user behavior preferences over time, being more in line with the user's current intention. By analyzing the user registration information and combining the recently queried service categories to extract personalized preferences, the vectorized expression of the user's long-term stable preferences is realized, constructing and supplementing the background semantic features of the behavior sequence, further enriching the individual dimension of the user portrait. Non-linearly fuse the time-series weighted behavior representation vector and the static preference vector B through a multi-layer perceptron (MLP) to realize the deep coupling of behavior-driven features and user preference features in the semantic space, thereby generating a more real-time, individual-differentiated, and behavior context-aware dynamic representation vector of the user portrait.

[0078] S2. Based on the constructed user profile, identify the user's current intention through the temporal attention mechanism model, predict the future behavior path, and according to the user input data and the current intention, use deep learning fusion calculation to perform TF-IDF weighting and semantic matching, and obtain the final candidate answer set through the rough set model and pruning mechanism;

[0079] Specifically, based on the constructed user profile, identifying the user's current intention through the temporal attention mechanism model and predicting the future behavior path means concatenating the generated dynamic representation vector U of the user profile and the multi-modal semantic representation vector A' into a joint representation vector X;

[0080] Use three groups of weight matrices in the temporal self-attention mechanism to perform a linear transformation on the joint representation vector X to generate a query vector Q, a key vector K, and a value vector V;

[0081]

[0082] Among them, is a learnable linear transformation matrix;

[0083] Through the dot product (similarity) of the query vector Q and the key vector K, and scaling by the square root of the dimension of the key vector K, according to the calculated dot product, assign a weight to each input feature (the dynamic representation vector U of the user profile is a dynamic feature, and the multi-modal semantic representation vector A' is a static feature) to weight the value vector V to obtain a weighted value vector, and sum the weighted values to obtain the key behavior feature within each time step (the input feature with the highest weight, which plays a major role in user behavior prediction):

[0084]

[0085] Among them, Attention(Q, K, V) is the weighted value vector, Q, K, V are the query vector, the key vector, and the value vector respectively, is the square root of the dimension of the key vector K, used to scale the dot product result to prevent the value from being too large, K T is the transpose of the key vector K, QK T is the dot product of the query vector Q and the key vector K;

[0086] Adopt the multi-head attention mechanism to increase the comprehensive expression ability of the model, map the query, key, and value vectors of the key behavior feature to each head (multiple subspaces), calculate multiple attentions simultaneously and analyze them to capture the feature representations in different subspaces:

[0087] MultiHead(Q, K, V) = Concat(head1, head2,..., head m )D O

[0088] Among them, MultiHead(Q, K, V) is the comprehensive behavioral feature (the different importance of user portraits and multimodal semantic features in multiple subspaces), and head m is the m-th head calculated, and D O is the output mapping matrix, which connects each head and generates the final output;

[0089] By capturing the feature representations of different subspaces, the current behavior pattern of the user is recognized, and the comprehensive behavioral feature is mapped to the corresponding current intention category vector y through a fully connected layer:

[0090] y = softmax(W intent ·MultiHead(Q, K, V) + b intent )

[0091] Among them, y is the current intention category vector, and W intent is the weight of y, which is set through machine learning and regression models, and b intent is the bias term of y, and MultiHead(Q, K, V) is the comprehensive behavioral feature;

[0092] Based on the current intention category vector y, the attention mechanism is used to select the positions in the behavior sequence (the behaviors that affect future behaviors), and the pointer mechanism in the pointer network is used to predict and generate the next behavior path of the user:

[0093] P next = Pointer(y, X)

[0094] Among them, P next is the predicted next behavior path, y is the intention category vector, and X is the joint representation vector;

[0095] The predicted next behavior path P next and the current intention category vector y are merged to generate an intention vector y' that contains the current user's intention and future behavior prediction information.

[0096] By splicing the dynamic vector of the user profile and the multi-modal semantic vector, a joint representation is constructed, realizing the unified modeling of the user's historical behavior and the current input, providing complete context information for intent recognition. The importance weights of the input features are calculated using the temporal self-attention mechanism to extract key behavior features, effectively mining the core information driving the user's intent, enhancing the model's temporal sensitivity and behavior discrimination ability. Combining the multi-head attention mechanism, semantic differences are captured from multiple subspaces, improving the expressiveness of the user's behavior preferences in different dimensions, making the intent recognition results more delicate and accurate. By mapping the comprehensive behavior features to intent vectors and combining pointer networks to predict the user's next behavior path, the ability to infer future user intents is strengthened. Finally, a full-link modeling from understanding the user's current intent to predicting future behavior is achieved, significantly enhancing the semantic understanding depth, response forward-lookingness, and personalized interaction experience of the question and answer.

[0097] Furthermore, according to the user input data and the current intent, using deep learning fusion calculation, TF-IDF weighting and semantic matching are performed, and the final answer candidate set is obtained through the rough set model and the pruning mechanism. That is, the multi-modal semantic representation vector A' and the future path intent vector y' are input into the deep learning model for fusion calculation to obtain the final joint semantic representation X': <s

[0098] X' = MLP([Z(A'), Z intent (y' intent )])

[0099] where X' is the final joint semantic representation, Z(A') is the vector representation of the multi-modal semantic representation vector A' after being processed by the convolutional neural network, and Z intent (y' intent ) is the vector representation of the future path intent vector y' after being processed by the convolutional neural network;

[0100] Based on the final joint semantic representation X' and the knowledge base for matching, the final joint semantic representation X' is segmented into individual words through the BERT tokenizer, and through part-of-speech tagging, each word in the text is marked with its corresponding part-of-speech (noun, verb, adjective), and nouns and verbs are retained as candidate keywords. The candidate keywords are subjected to TF-IDF weighting calculation to accurately evaluate the importance of each keyword in the text:

[0101]

[0102] where, w i is the TF-IDF weight of the word i, tf i is the frequency of occurrence of the word i in the text, N is the total number of documents, n i is the number of documents containing the word i, and s is a weighting constant set through smoothed IDF calculation;

[0103] Through TF-IDF weighted calculation, the weight value of each keyword is obtained, and a sequence of feature keywords {u1, u2,..., u e} is generated, where u e is the e-th keyword in the text;

[0104] Using the sliding window mechanism, the window size is set to g. Each time, g consecutive keywords are selected as a phrase, and the window is continuously slid over the entire sequence of feature keywords to generate a set of consecutive feature phrases until all possible triples (feature phrases) in the sequence are extracted. Then, the TF-IDF weights of all keywords in the feature phrase are weighted and summed to obtain the comprehensive weight of the feature phrase, which reflects the semantic importance of the feature phrase in the entire text. Calculate the comprehensive weights of all feature phrases and sort them in descending order to obtain the set of feature phrases a β ;

[0105] Decompose the set of feature phrases a β into independent words, and use the HowNet semantic knowledge base to calculate the similarity between each independent word and its lexical semantic concept for matching. By constructing the root node of the semantic tree with a depth of 0, starting from the root node, each lexical semantic concept is represented as a node in the tree structure. The depth of the direct child nodes under the root node is 1, and the depth of the child nodes of each child node is incremented by 1 based on the previous node. Calculate the path distance according to the number of edges required to pass from each concept node to the root node;

[0106] The calculation formula for the lexical semantic concept Sim(c1, c2) is:

[0107]

[0108] where c1 and c2 are two concepts corresponding to the words, o1 and o2 are the positions of these two concepts in the semantic tree, is the depth of these two concept nodes (i.e., their levels in the semantic tree), distance(o1, o2) is the path distance between the two concept nodes in the semantic tree, representing the semantic relationship distance between them, is the adjustment factor used to control the weights of semantic depth and path distance;

[0109] By heuristically setting the similarity threshold ψ in the knowledge base, when the similarity Sim(c1, c2) is greater than or equal to ψ, they are semantically close enough to be replaced, generating a new semantic feature phrase a' containing concepts e , otherwise, no replacement is made;

[0110] Each semantic feature phrase is converted into a unique and fixed-length fingerprint representation using the Karp-Rabin hash function:

[0111]

[0112] where is the β -th phrase in the semantic feature phrase a ', is the encoding of , p is a selected prime number used to prevent hash collisions and ensure the uniqueness of hash values, Θ is the number of words in the phrase, x is the index;

[0113] Apply the Karp-Rabin hash function to each phrase in the semantic feature phrase a β ' and map it to a unique digital fingerprint, forming a semantic fingerprint set F input , which represents the core semantic features of the text and can be used for quick matching with knowledge items in the knowledge base;

[0114] Use a two-tower network to encode the final joint semantic representation X' and each knowledge item in the knowledge base, generate a feature vector representation and calculate the similarity. During the similarity calculation process, use the semantic fingerprint set F input to affect the generation of feature vectors and indirectly improve the accuracy of similarity calculation:

[0115]

[0116] where Sim(X', G j ) is the semantic similarity, G j is the j-th knowledge item in the knowledge base G, V X' , is the feature vector representation generated by the two-tower network, is the semantic fingerprint of the final joint semantic representation X', is the semantic fingerprint of G j , Π and Ω are weighting factors set by genetic algorithms;

[0117] Through the calculation of semantic similarity, obtain the candidate answer items that match the joint semantic representation X' in the knowledge base and aggregate them into a candidate answer set W;

[0118] Use the probabilistic tolerance rough set model to further optimize the candidate answer set W. The rough set model deals with the ambiguity and redundant information between the text and candidate knowledge items through the tolerance relation, enhances the robustness of the system, and constructs its upper and lower approximation sets:

[0119]

[0120] E(S, θi ) = {θ f | sim(θ i , θ f ) ≥ δ}

[0121] where U(S, w i ) is the upper approximation set, L(S, w i ) is the lower approximation set, S is the similarity measure, θ i is the currently processed vocabulary i, θ f is the vocabulary f compared with θ i ; is the tolerance, set by the rule of thumb, δ is the confidence level, set by cross - validation, and δ are used to control the flexibility and accuracy of the similarity;

[0122] According to the sizes of the intersection and union of the upper and lower approximation sets of the semantic feature phrase a β ' and the candidate answer item W ξ , measure the tolerance similarity of the semantic feature phrase a β ' and the candidate answer item W ξ , effectively handle the semantic ambiguity and redundant information between vocabulary, and improve the robustness and precision of the similarity calculation:

[0123]

[0124] where a β ' is the semantic feature phrase, W ξ is the ξ - th candidate in the candidate answer item set W, U(S, a β ) and E(S, a β ) are the upper and lower approximation sets of the feature phrase, U(S, W ξ ) and E(S, W ξ ) are the upper and lower approximation sets of the candidate knowledge item;

[0125] Based on the tolerance similarity calculation, set the similarity threshold Γ through statistical analysis. When the tolerance similarity of the candidate item is greater than or equal to the threshold Γ, then the answer candidate item is retained in the candidate answer set W and sorted in descending order. Otherwise, it is screened out from the candidate answer set W;

[0126] Through the screened answer candidates, generate a retrieval - type answer and decode it. In the decoding stage, use a dynamic pruning mechanism to eliminate the redundant answer candidate decoding attention channels to reduce the computational burden and accelerate the output of the candidate answer set W';

[0127] The formula for the dynamic pruning mechanism is:

[0128]

[0129] Among them, I γ is the pruning decision of the γ-th attention channel, and λ γ is the weight of the γ-th attention channel, which is set through multi-objective optimization. is the average weight of this channel, and τ is the pruning threshold, which is set according to the weight magnitude;

[0130] λ γ The calculation formula is:

[0131]

[0132] Among them, Q γ and K γ are the query vector and key vector of the γ-th attention channel, is the square root of the dimension of the key vector;

[0133] The calculation formula is:

[0134]

[0135] Among them, is the weight of the γ-th channel in the candidate answer , and Λ is the total number of candidate answers in the candidate answer set W;

[0136] Generate φ generative candidate answers through a generative model (the generative model includes generative adversarial networks, autoregressive models, variational autoencoders, flow models, and diffusion models, and here the autoregressive model is selected), and use the Beam Search method to select generative candidate answers from the generative model, and combine the answer annotation sources (retrieval-based, generative), the semantic similarity between the input and the answer, and the pruning degree under the current input in the candidate answer set W' to output the final answer candidate set

[0137] By fusing the multi-modal semantic representation vector with the user's future path intention vector and inputting them into a deep model, the joint expression of the current input semantics and the behavior prediction result is achieved, which is used to improve the ability of the question-answering system to grasp the user's comprehensive intention. The convolutional neural network is used to encode the two types of vectors respectively and generate a unified joint semantic representation, enhancing the model's ability to capture the relevance between contexts and making the subsequent knowledge matching process more semantically consistent. The BERT tokenization and TF-IDF weighting mechanism are used to extract high-value keywords, and semantic phrases are constructed through a sliding window, realizing the automatic extraction of representative semantic fragments in the text and providing a sparse semantic representation for the matching process, effectively improving the computational efficiency. Combining HowNet semantic tree and path distance calculation, the structural modeling of the deep semantic relationship between words is realized; on this basis, the semantic fingerprint constructed can be quickly indexed and compared after Karp-Rabin hashing, greatly reducing the search overhead of knowledge item matching. The semantic fingerprint is introduced into the feature representation generation through a two-tower network, and candidate answers are screened based on vector similarity, improving the semantic matching accuracy and the relevance of the answers, making the candidate set more focused on semantic targets. The rough set tolerance model is used to construct the upper and lower approximation sets and perform fuzziness tolerance processing on the candidate set, effectively alleviating the ambiguity and overlap problems in natural language semantic expressions and enhancing the robustness and practicality of the system. Finally, combined with the dynamic pruning mechanism, redundant answer paths are identified and removed in the decoding stage, realizing the structural compression and efficiency optimization of the candidate answer decoding process. At the same time, BeamSearch is used to generate the optimal candidate set to ensure the best performance of the output results in terms of relevance, conciseness, and semantic coverage.

[0138] S3. By calculating the answer confidence and selecting the answer with the optimal confidence, and continuously optimizing the question-answering strategy in combination with reinforcement learning.

[0139] Specifically, calculating the answer confidence and selecting the answer with the optimal confidence means using the multi-modal semantic representation vector A' to calculate the semantic matching confidence, context consistency score, and generation quality score of the th candidate answer in the candidate set;

[0140] The calculation of the semantic matching confidence encodes A' and through a pre-trained BERT model and calculates it using cosine similarity to obtain the semantic matching confidence of the th candidate answer;

[0141] The context consistency score calculates the context consistency score of the th candidate answer through BERTScore, which is used to evaluate whether the candidate answer is consistent with the entire conversation or context;

[0142] The generated quality score uses ROUGE to calculate the overlap between the generated text and the reference text;

[0143] Based on the semantic matching confidence, context consistency score, and the calculation result of the generated quality score, the comprehensive score of each candidate answer is calculated through weighted summation;

[0144] The threshold q is set through dynamic distribution adaptation. The candidate answers with scores greater than the threshold q are retained and sorted in descending order. The top C candidate answers with scores are retained as alternative suggestions for display to the user. Otherwise, the user is prompted to re-enter the query.

[0145] Through multi-modal semantic representation vectors for semantic matching, context consistency, and multi-dimensional scoring of the generated quality of candidate answers, a fine-grained evaluation mechanism for the quality of candidate answers is realized. It can not only identify whether the semantics fit, but also judge whether the answer is coherent and the content is of high quality, improving the accuracy of question answering and context adaptability. By calling the pre-trained BERT model and using cosine similarity to calculate the semantic matching confidence, the measurement of the deep semantic relevance between the user input and the candidate answer is realized, effectively filtering out low-quality answers with semantic deviation and ensuring that the generated results are semantically closely related. Introducing BERTScore to evaluate context consistency realizes the evaluation of the logical connection between the answer and the dialogue history, avoiding semantic jumps or context breaks, helping to generate more continuous and smoother answer content for users to understand, and improving the overall dialogue quality. Combining the ROUGE index to calculate the generated quality score realizes the evaluation of the structural overlap degree between the answer and the standard reference answer, ensuring the integrity and coverage of the generated results, and effectively improving the readability and content coverage in generative question answering. Finally, through weighted aggregation of scores in each dimension and setting a threshold screening mechanism, only the candidate items with high comprehensive quality are retained for display. When the scores are generally insufficient, the user can be prompted to re-enter, enhancing the response control ability and user experience consistency.

[0146] Furthermore, continuously optimizing the question answering strategy by combining reinforcement learning means that after confidence screening and outputting answers, the question answering interaction is formalized as a Markov decision process modeled by states, actions, and rewards:

[0147]

[0148] Among them, the state μ t is the environmental state at time t, jointly composed of historical question-answer pairs, the current dialogue context, and the user profile vector. The action ε t is the answer strategy selected by the model in the state μ t , such as selecting a certain candidate answer. The reward is the feedback signal provided by the user after time t;

[0149] Optimize the question - answering decision - making process using a policy network. The policy network outputs an action distribution (generates a series of possible answer strategies based on the current state), and uses Proximal Policy Optimization (PPO) to train and optimize the learnable parameter set in the policy network

[0150]

[0151] Among them, is the loss function of Proximal Policy Optimization, which is used to evaluate the difference between the current policy and the old policy, and guide the optimization of the policy, Ξ t is the time - step operator, indicating the operation at time step t, is the policy update ratio, M t is the advantage function, estimated by Generalized Advantage Estimation (GAE), which is used to evaluate the superiority of each action selection to guide the policy update direction. ∈ is the clipping coefficient, set through an adaptive policy to prevent the policy from updating violently and ensure stable training;

[0152] For the loss function take the negative value for minimization, and use the Adam optimizer to update the learnable parameter set in the policy network ;

[0153] On the basis of policy network optimization, through an experience replay pool, which is specifically used to store high - score feedback trajectories, enhance the system's memory of excellent answer paths, and set the threshold through a fixed - ratio screening method After each interaction, if the reward is greater than the threshold add the current trajectory to the preferred replay pool. Every v times of training, sample a batch of high - quality trajectories from the pool and use the target reinforcement policy to update:

[0154]

[0155] Among them, is the gradient used to update the policy parameters, is the probability of selecting an action under state μ, and the gradient is the derivative of the policy parameter ; is under the policy with parameter the expected total reward obtained from the environment, is the reward value;

[0156] Based on the calculated gradient used to update the policy parameters,adopt the gradient ascent strategy to update the parameter set of the policy network ​Perform iterative updates;

[0157] Combine the A / B test mechanism to assign multiple trained policy versions to different user subgroups, collect feedback data (such as like rate, bounce rate) of both under the same number of users, and calculate the feedback value difference between the new policy and the old policy. If the feedback value of the new policy is greater than that of the old policy, switch it online.

[0158] By formalizing the Q&A process into a Markov decision process, an optimizable state-action-reward mechanism is realized, enabling the model to adaptively select response strategies based on historical Q&A pairs, the current context, and the user profile. By adopting a policy network and using the Proximal Policy Optimization (PPO) algorithm for training, while maintaining the stability of policy updates, the policy performance is effectively improved, avoiding the problem of unstable Q&A quality caused by drastic policy changes, enhancing the reliability and learning ability in real interaction scenarios. Through the design of the advantage function and clipped loss, the policy network preferentially learns superior action selection paths, improving the identification and learning efficiency of high-value answer strategies, enabling continuous convergence towards a better response direction during the policy update process, raising the upper limit of answer quality. By designing an experience replay pool and storing historical Q&A trajectories with good user feedback, and regularly sampling for optimizing the target policy, long-term memory and reuse of high-quality interaction experiences are realized, improving the utilization efficiency of training samples and the generalization ability of the model, avoiding the model from falling into local optimal policy paths. Combining the A / B test mechanism, deploying multiple policy versions to different user groups for comparative experiments, realizing the quantitative evaluation of the performance differences between the new and old policies and the construction of the basis for dynamic switching, ensuring that the performance of the new policy online is better than that of the old version, significantly improving the scientific nature of Q&A policy upgrades and the stability of the user experience.

[0159] This embodiment also provides an AI-based intelligent airport service Q&A system, including:

[0160] A multimodal data processing module for denoising, temporal alignment, and dimensional normalization of multimodal data, extracting features and converting them into a unified text semantic representation;

[0161] A user profile construction module for constructing a user profile based on historical queries, browsing behaviors, and personalized information, generating a dynamic representation vector;

[0162] An intention recognition and behavior prediction module for using a temporal self-attention mechanism to recognize the current user's intention and predict the future behavior path based on the user profile and multimodal semantic representation;

[0163] A semantic matching module for matching through semantic representations with a knowledge base, performing rough set optimization and dynamic pruning to generate a final set of answer candidates;

[0164] Answer optimization and strategy module, which is used to calculate confidence and optimize strategies, optimize Q&A decisions through reinforcement learning, and improve response effects.

[0165] This embodiment also provides a computer device applicable to the case of the AI-based airport intelligent service Q&A method, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the AI-based airport intelligent service Q&A method as proposed in the above embodiment.

[0166] This computer device can be a terminal. The computer device includes a processor, a memory, a communication interface, a display screen, and an input device connected through a system bus. Among them, the processor of this computer device is used to provide computing and control capabilities. The memory of this computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of this computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be implemented through WIFI, a carrier network, NFC (Near Field Communication), or other technologies. The display screen of this computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of this computer device can be a touch layer covering the display screen, or a button, a trackball, or a touchpad provided on the computer device housing, or an external keyboard, a touchpad, or a mouse, etc.

[0167] This embodiment also provides a storage medium, on which a computer program is stored. When the program is executed by a processor, it implements the AI-based airport intelligent service Q&A method as proposed in the above embodiment; the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM for short), Electrically Erasable Programmable Read-Only Memory (EEPROM for short), Erasable Programmable Read Only Memory (EPROM for short), Programmable Red-Only Memory (PROM for short), Read-Only Memory (ROM for short), magnetic memory, flash memory, a magnetic disk, or an optical disc.

[0168] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered within the scope of the claims of the present invention.

Claims

1. An AI-based intelligent service Q&A method for airports, characterized in that: include, Receive and analyze multimodal data input by users, and build user profiles in real time based on their historical query sequences, browsing behaviors, and personalized information; The multimodal data includes text, voice, image and location information; The personalized information includes language preference and user type; Based on the constructed user profile, the temporal attention mechanism model is used to identify the user's current intent and predict future behavior paths. Based on the user's input data and current intent, deep learning fusion calculations are used to perform TF-IDF weighting and semantic matching. The final set of answer candidates is obtained through a rough set model and pruning mechanism. By calculating the answer confidence and selecting the answer with the best confidence, the question-answering strategy is continuously optimized in combination with reinforcement learning.

2. The AI-based intelligent service Q&A method for airport according to claim 1, wherein: The receiving and parsing of multimodal data input by the user refers to removing noise, time series alignment and dimension normalization of the multimodal data, extracting its features and converting them into a unified text semantic representation, and performing semantic alignment based on a multimodal fusion model of a self-attention network architecture. A multimodal encoder is used to convert features of different modalities into vector representations, and the features are spliced together to obtain a multimodal semantic representation vector A'.

3. The method for answering questions about intelligent services at an airport based on AI according to claim 2, wherein: The real-time construction of a user profile by combining the user's historical query sequence, browsing behavior, and personalized information refers to obtaining the user's historical query text, query timestamp, page access record, and page dwell time from the database to construct a behavior sequence vector H; Use the time decay function to weight the behavior sequence vector H to obtain the final behavior representation vector H'; Using the basic information provided by the user during registration, we analyze and obtain personalized preference features, and then represent the user preference vector B through rule vectorization. The final behavior representation vector H' and the user preference vector B are spliced through a multi-layer perceptron to obtain a complete user portrait dynamic representation vector U.

4. The method for answering questions about intelligent airport services based on AI according to claim 3, characterized in that: The method of identifying the user's current intention based on the constructed user portrait through the temporal attention mechanism model and predicting the future behavior path refers to splicing the generated user portrait dynamic representation vector U and the multimodal semantic representation vector A' into a joint representation vector X; The three sets of weight matrices in the temporal self-attention mechanism are used to linearly transform the joint representation vector X to generate the query vector Q, key vector K, and value vector V, thereby obtaining the key behavioral features within each time step. A multi-head attention mechanism is used to map the query, key, and value vectors of key behavioral features to each head. Multiple attentions are calculated and analyzed simultaneously to capture feature representations of different subspaces and map them into comprehensive behavioral features MultiHead(Q, K, V). By capturing the feature representations of different subspaces, the user's current behavior pattern is identified, and the comprehensive behavior feature MultiHead (Q, K, V) is mapped to the corresponding current intent category vector y through the fully connected layer; Based on the current intention category vector y, the attention mechanism is used to select the positions in the behavior sequence for indication, and the pointer mechanism in the pointer network is used to predict and generate the user's next behavior path P next ; Merge the predicted next-step behavior path P next with the current intent category vector y to generate an intent vector y' that contains the current user's intent and future behavior prediction information.

5. The method for answering questions about intelligent airport services based on AI according to claim 4, characterized in that: According to the user input data and the current intention, using deep learning fusion computing, perform TF-IDF weighting and semantic matching, and obtain the final answer candidate set through the rough set model and pruning mechanism, which means inputting the multi-modal semantic representation vector A' and the future path intention vector y' into the deep learning model for fusion computing to obtain the final joint semantic representation X'; Match the final joint semantic representation X' with the knowledge base. Use the BERT tokenizer to tokenize the final joint semantic representation X' into individual words, annotate the corresponding part of speech, retain nouns and verbs as candidate keywords, and perform TF-IDF weighted calculation to obtain the TF-IDF weight w of word i i , calculate the weight of each keyword, and generate a sequence of feature keywords according to the weights; Set the window size to g, and each time continuously select g keywords as a phrase. Extract all triples from the sequence of feature keywords, calculate the comprehensive weight of the feature phrase by weighted summation of the TF-IDF weights of all keywords in the feature phrase, sort the feature phrases in descending order according to the comprehensive weight, and obtain the set a of feature phrases β ; Decompose the feature phrase set a β into independent words, and use the HowNet semantic knowledge base to calculate the similarity Sim(c1, c2) between each independent word and its lexical semantic concept for matching; Set the similarity threshold ψ. When the similarity Sim(c1, c2) is greater than or equal to ψ, they are semantically close enough to be replaced, generating a new semantic feature phrase a that contains the concept. β ', otherwise, no replacement is made. Convert each semantic feature phrase into a unique and fixed-length fingerprint using the Karp-Rabin hash function Hash(a β '), and apply the Karp-Rabin hash function to each phrase in the semantic feature phrase a β ' and map it to a unique digital fingerprint to form the semantic fingerprint set F input ; Use a two-tower network to encode the final joint semantic representation X' and each knowledge item in the knowledge base to generate a feature vector representation and combine it with the semantic fingerprint set F input Calculate the semantic similarity Sim(X', G j ) Through the calculation of semantic similarity, obtain the answer candidates in the knowledge base that match the joint semantic representation X', and aggregate them into the candidate answer set W; Further optimize the candidate answer set W using the probabilistic tolerance rough set model by constructing the upper and lower approximation sets U(S,θ i ), E(S,θ i ), and calculate the tolerance similarity Sim(a β ', W ξ ) according to the size measures of the intersection and union of the upper and lower approximation sets of the semantic feature phrase a β ' and the candidate answer item W ξ ; Based on the tolerance similarity calculation, set the similarity threshold Γ. When the tolerance similarity of the candidate item is greater than or equal to the threshold Γ, the answer candidate item is retained in the candidate answer set W and sorted in descending order. Otherwise, it is screened out from the candidate answer set W; Through the selected answer candidates, generate a retrieval-based answer and decode it. In the decoding stage, use the dynamic pruning mechanism to calculate the pruning decision for each attention channel, eliminate the redundant answer candidate decoding attention channels, and output the answer candidate set W'; Generate φ candidate answers for the final joint semantic representation X' through the generation model, and use the BeamSearch method to select generative candidate answers from the generation model, combine the answers in the candidate answer set W', annotate the source, the semantic similarity between the input and the answer, and the degree of pruning under the current input, and output the final candidate answer set 6. The AI-based intelligent service Q&A method for airport according to claim 5, wherein: The step of calculating the answer confidence and selecting the optimal confidence answer refers to calculating the semantic matching confidence, context consistency score, and generation quality score of the th candidate answer in the candidate set by using the multimodal semantic representation vector A'. The semantic matching confidence is obtained by encoding A' and using a pre-trained BERT model, and calculating the semantic matching confidence of the th candidate answer using cosine similarity; The context consistency score is calculated by BERTScore for the th candidate answer's context consistency score; The generated quality score uses ROUGE to calculate the overlap degree between the generated text and the reference text; Based on the semantic matching confidence, context consistency score, and generated quality score calculation results, calculate the comprehensive score of each candidate answer through weighted summation; Set the threshold q, retain the candidate answers with scores greater than the threshold q and sort them in descending order, retain the top C candidate answers with scores as alternative suggestions to be presented to the user, otherwise prompt the user to re-enter the query.

7. The method for answering questions about intelligent services at airports based on AI according to claim 6, characterized in that: The continuous optimization of the question-answering strategy by combining reinforcement learning means that after the confidence screening and answer output are completed, the question-answering interaction is formalized into a Markov decision process modeled by state, action, and reward; Optimize the Q&A decision-making process using a policy network, where the policy network outputs an action distribution and optimizes the policy network using proximal policy optimization to train the set of learnable parameters in Based on the optimization of the policy network, through the experience replay pool, specifically store high-score feedback trajectories and set a threshold After each interaction, if the reward is greater than the threshold add the current trajectory to the preferred replay pool. Every v times of training, sample a batch of high-quality trajectories from the pool and use the target reinforcement policy to update the gradient of the policy parameters Perform iterative updates on the parameter set ; Combined with the A / B test mechanism, assign multiple trained policy versions to different user subgroups, collect the feedback data of both under the same number of users, and calculate the feedback numerical difference between the new policy and the old policy. If the feedback value of the new policy is greater than that of the old policy, switch it online.

8. A question-and-answer system for intelligent airport services based on AI, based on a method for intelligent airport service question and answer according to any one of claims 1 to 7, characterized in that: Including, A multi-modal data processing module for denoising, temporal alignment, and dimensional normalization of multi-modal data, extracting features and converting them into a unified text semantic representation; A user portrait construction module for constructing a user portrait based on historical queries, browsing behaviors, and personalized information, and generating a dynamic representation vector; An intention recognition and behavior prediction module for identifying the current user intention and predicting the future behavior path using the temporal self-attention mechanism based on the user portrait and multi-modal semantic representation; A semantic matching module for matching through semantic representation with the knowledge base, performing rough set optimization and dynamic pruning, and generating the final answer candidate set; An answer optimization and strategy module for calculating confidence and optimizing the strategy, and optimizing the question-answering decision through reinforcement learning to improve the response effect.

9. A computer device, comprising a memory and a processor, the memory storing a computer program, characterized in that: When the processor executes the computer program, it implements the steps of an AI-based intelligent airport service question-answering method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of an AI-based intelligent airport service question-answering method according to any one of claims 1 to 7.

Images