An ai-based airport intelligent service question and answer method and system

Abstract

The application discloses an AI-based airport intelligent service question and answer method and system, relates to the technical field of artificial intelligence, and comprises the following steps: the application utilizes time sequence and multi-head attention mechanism to jointly model, extracts key behavior characteristics, fuses intention recognition and behavior prediction, improves the accuracy of answer reasoning, constructs deep semantic representation, and accurately matches the semantic knowledge base; in combination with a sliding window mechanism, high-weight phrases are extracted, a Karp-Rabin hash function is used to generate unique semantic fingerprints, the speed and accuracy of answer matching are improved, a rough set tolerance model is used to control the fuzziness of candidate answers and remove redundant information, the semantic similarity is judged according to the intersection and union ratio between the upper and lower approximate sets, the problems of semantic drift and repeated candidates in traditional matching are effectively alleviated, the question and answer strategy is continuously optimized through reinforcement learning and A / B testing, and the response effect of the system and the user satisfaction are enhanced.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to an AI-based intelligent airport service question-and-answer method and system. Background Technology

[0002] In recent years, with the rapid development of artificial intelligence technology, especially breakthroughs in natural language processing and computer vision, intelligent service systems have been widely applied in various scenarios, with intelligent question-and-answer systems becoming an important component. In the specific environment of airports, the application of airport intelligent service question-and-answer systems is gradually increasing, aiming to provide passengers with real-time and accurate consultation services. Existing technologies typically acquire various input data through speech recognition, image recognition, and text input, responding in real-time to various passenger needs in the airport environment, and using pre-trained models and rule matching methods to provide automated question-and-answer services, offering passengers a more convenient query channel.

[0003] While existing intelligent service question-answering systems can match user queries using certain rules, they still fall short in accurately understanding and predicting user needs. First, current technologies have weak capabilities in fusing multimodal data, relying heavily on single-modal input (such as text or speech) for semantic understanding. They fail to fully utilize auxiliary information from multimodal data such as images and videos, resulting in the system often failing to fully capture the user's true intent when the user's input data contains multiple modalities. This is especially problematic in complex scenarios where traditional methods exhibit poor robustness. Second, existing technologies have limitations in building user profiles and providing personalized services. Although they have begun to use historical user data for personalized answers, they rely on static user data, lacking real-time updates and failing to capture changes in user behavior during actual interactions, leading to poor personalized answer performance. Furthermore, existing technologies still have limitations in generating and optimizing question-answer content. Many existing question-answering systems use simple rule-based or template-based matching to generate answers. While this can handle some common questions, it often fails to provide satisfactory answers to complex and ever-changing user queries. Therefore, existing intelligent question answering technologies mainly face problems such as insufficient ability to process multimodal data, inaccurate user profiles, and insufficient optimization of answer generation. Summary of the Invention

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

[0005] Therefore, this invention provides an AI-based intelligent airport service question-answering method and system to solve the problems of insufficient multimodal data capabilities, inaccurate user profiles, and insufficient optimization of answer generation.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides an AI-based intelligent airport service question-answering method and system, which includes,

[0008] Receive and parse multimodal data input by users, and build user profiles in real time by combining users' historical query sequences, browsing behavior and personalized information;

[0009] The multimodal data includes text, voice, images, and location information;

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

[0011] Based on the constructed user profile, the system identifies the user's current intent and predicts future behavior paths through a temporal attention mechanism model. Based on the user's input data and current intent, it uses deep learning to perform TF-IDF weighted calculation and semantic matching, and obtains the final candidate set of answers through a rough set model and pruning mechanism.

[0012] By calculating the confidence level of the responses and selecting the response with the highest confidence level, the question-answering strategy is continuously optimized through reinforcement learning.

[0013] As a preferred embodiment of the AI-based airport smart service question-answering method described in this invention, the step of receiving and parsing the multimodal data input by the user refers to removing noise, aligning the temporal sequence, and normalizing the dimension of the multimodal data, extracting its features and converting them into a unified text semantic representation, performing semantic alignment based on a multimodal fusion model with a self-attention network architecture, using a multimodal encoder to convert the features of different modalities into vector representations, and concatenating them to obtain a multimodal semantic representation vector A′.

[0014] As a preferred embodiment of the AI-based airport smart service question-and-answer method described in this invention, the step of constructing a user profile in real time 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 records, and page dwell time from the database to construct a behavior sequence vector H;

[0015] The behavior sequence vector H is weighted using a time decay function to obtain the final behavior representation vector H′;

[0016] By utilizing the basic information provided by users during registration, personalized preference features are analyzed and obtained, and user preference vector B is represented by rule vectorization.

[0017] The final behavior representation vector H′ and the user preference vector B are concatenated using a multilayer perceptron to obtain the complete dynamic representation vector U of the user profile.

[0018] As a preferred embodiment of the AI-based airport smart service question-answering method described in this invention, the step of identifying the user's current intent and predicting future behavior paths based on the constructed user profile and through a temporal attention mechanism model refers to concatenating the generated user profile dynamic representation vector U and the multimodal semantic representation vector A′ into a joint representation vector X.

[0019] The joint representation vector X is linearly transformed using the three sets of weight matrices in the temporal self-attention mechanism to generate the query vector Q, key vector K, and value vector V, thereby obtaining the key behavioral features in each time step.

[0020] A multi-head attention mechanism is adopted to map the query, key, and value vectors of key behavioral features to each head. At the same time, multiple attentions are calculated and analyzed to capture feature representations of different subspaces and map them into comprehensive behavioral features MultiHead(Q,K,V).

[0021] By capturing feature representations from different subspaces, the user's current behavior pattern is identified, and the comprehensive behavioral features MultiHead(Q,K,V) are mapped to the corresponding current intent category vector y through a fully connected layer;

[0022] Based on the current intent category vector y, an attention mechanism is used to select the position in the behavior sequence, and a pointer mechanism in a pointer network is used to predict the next behavior path P of the generating user. next ;

[0023] The predicted next behavioral path P next Merge 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.

[0024] As a preferred embodiment of the AI-based airport smart service question-answering method described in this invention, the step of using deep learning fusion calculation based on user input data and current intent to perform TF-IDF weighted and semantic matching, and obtaining the final answer candidate set through rough set model and pruning mechanism refers to inputting the multimodal semantic representation vector A′ and the future path intent vector y′ into the deep learning model for fusion calculation to obtain the final joint semantic representation X′;

[0025] Based on the matching of the final joint semantic representation X' with the knowledge base, the final joint semantic representation X' is segmented into individual words using the BERT word segmentation tool, and the corresponding parts of speech are labeled. Nouns and verbs are retained as candidate keywords, and TF-IDF weighted calculation is performed to obtain the TF-IDF weight w of word i. i Calculate the weight of each keyword and generate a sequence of feature keywords based on the weights;

[0026] Set the window size to g, and select g keywords consecutively as a phrase. Extract all triples from the feature keyword sequence. Calculate the comprehensive weight of the feature phrase by weighted summing of 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 feature phrase set a. β ;

[0027] The feature phrase set a β The words are decomposed into independent words, and the similarity Sim(c1,c2) ​​between each independent word and its semantic concept is calculated using the HowNet semantic knowledge base for matching.

[0028] A similarity threshold ψ is set. When the similarity Sim(c1,c2) ​​is greater than or equal to ψ, the two phrases are semantically close enough to be replaced, generating a new semantic feature phrase a containing the concept. β Otherwise, no replacement will be performed;

[0029] Each semantic feature phrase is converted into a unique and fixed-length fingerprint hash using the Karp-Rabin hash function. β '), for semantic feature phrase a β Each phrase in ' is hashed using a Karp-Rabin function and mapped to a unique digital fingerprint, forming a semantic fingerprint set F. input ;

[0030] By using a dual-tower network encoding to ultimately jointly encode the semantic representation X' and each knowledge item in the knowledge base, a feature vector representation is generated and combined with the semantic fingerprint set F. input Calculate semantic similarity Sim(X',G) j );

[0031] By calculating semantic similarity, candidate answers that match the joint semantic representation X' in the knowledge base are obtained and set into a candidate answer set W;

[0032] The candidate response set W is further optimized using a probabilistic tolerance rough set model by constructing upper and lower approximation sets U(S,θ). i ), E(S,θ i ), and based on the semantic feature phrase a β 'and candidate answer item W ξ The tolerance similarity Sim(a) is calculated by measuring the size of the intersection and union of the upper and lower approximate sets. β ',W ξ );

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

[0034] The selected answer candidates are used to generate search-based answers and decode them. During the decoding stage, a dynamic pruning mechanism is used to calculate the pruning decision for each attention channel, eliminate redundant answer candidates, decode the attention channel, and output the answer candidate set W'.

[0035] Generative models generate φ candidate answers from the final joint semantic representation X', and Beam Search is used to select the answer from the generative model. Generative candidate answers are generated, and the answers in the candidate answer set W' are labeled with their source, semantic similarity between the input and the answer, and the degree of pruning under the current input, to output the final candidate answer set.

[0036] As a preferred embodiment of the AI-based airport smart service question-answering method described in this invention, the step of calculating the confidence level of the answer and selecting the answer with the optimal confidence level refers to calculating a candidate set using a multimodal semantic representation vector A'. The Middle The semantic matching confidence, contextual consistency score, and generation quality score of each candidate answer;

[0037] The semantic matching confidence score is obtained by using a pre-trained BERT model to combine A' and... Encode the code and calculate the cosine similarity to obtain the first... Semantic matching confidence of each candidate answer;

[0038] The context consistency score is calculated using the BERTS score. Each candidate answer's contextual consistency score;

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

[0040] The comprehensive score for each candidate answer is calculated by weighting the results of semantic matching confidence, context consistency score and generation quality score.

[0041] Set a threshold q, retain candidate answers with scores greater than the threshold q and sort them in descending order, retain the top C candidate answers with the highest scores as backup suggestions to show to the user, otherwise prompt the user to re-enter the question.

[0042] As a preferred embodiment of the AI-based airport smart service question-answering method described in this invention, the step of continuously optimizing the question-answering strategy by combining reinforcement learning refers to formalizing the question-answering interaction into a Markov decision process modeled through state, action, and reward after completing the confidence screening and outputting the answer.

[0043] Use policy networks to optimize the question-answering decision-making process. Output the action distribution and use a proximal strategy to train and optimize the policy network. The set of learnable parameters in

[0044] Based on policy network optimization, a high-scoring feedback trajectory is specifically stored through an experience replay pool, and a threshold is set. After each interaction, if a reward is given Greater than the threshold Current trajectory A high-quality replay pool is added. Every v training iterations, a batch of high-quality trajectories is sampled from the pool, and the gradients of the policy parameters are updated using the target reinforcement policy. For parameter set Perform iterative updates;

[0045] By combining A / B testing, multiple trained strategy versions are assigned to different user subgroups. Feedback data from both strategies is collected with an equal number of users. The difference in feedback values ​​between the new and old strategies is calculated. If the feedback value of the new strategy is greater than that of the old strategy, then the new strategy is switched to live.

[0046] Secondly, this invention provides an AI-based intelligent airport service question-and-answer system, including:

[0047] The multimodal data processing module is used to perform denoising, temporal alignment, and dimensionality normalization on multimodal data, extract features, and transform them into a unified text semantic representation.

[0048] The user profile building module is used to build user profiles based on historical queries, browsing behavior, and personalized information, and generate dynamic representation vectors.

[0049] The intent recognition and behavior prediction module is used to identify the current user intent and predict future behavior paths based on user profiles and multimodal semantic representations, using a temporal self-attention mechanism.

[0050] The semantic matching module is used to match semantic representations with the knowledge base, perform rough set optimization and dynamic pruning, and generate the final set of candidate answers.

[0051] The answer optimization and strategy module is used to calculate confidence and optimize strategies, thereby improving response performance by optimizing question-answering decisions through reinforcement learning.

[0052] Thirdly, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the AI-based airport smart service question-answering method as described in the first aspect of the present invention.

[0053] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the AI-based airport smart service question-answering method as described in the first aspect of the present invention.

[0054] The beneficial effects of this invention are as follows: This invention utilizes a joint modeling mechanism of temporal sequence and multi-head attention to extract key behavioral features, integrates intent recognition and behavior prediction to improve the accuracy of answer reasoning, constructs a deep semantic representation and performs accurate matching with a semantic knowledge base, extracts high-weight phrases using a sliding window mechanism, generates unique semantic fingerprints using the Karp-Rabin hash function to improve the speed and accuracy of answer matching, and employs a rough set tolerance model to perform fuzzy control and redundant information removal on candidate answers. Semantic similarity is judged based on the intersection and union ratio between upper and lower approximate sets, effectively alleviating the problems of semantic drift and duplicate candidates in traditional matching. Through reinforcement learning and A / B testing, the question answering strategy is continuously optimized to enhance the system's response effect and user satisfaction. Attached Figure Description

[0055] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0056] Figure 1 This is a flowchart of an AI-based smart airport service question-and-answer method in Example 1.

[0057] Figure 2 This is a structural diagram of an AI-based smart airport service question-and-answer system in Example 1.

[0058] Figure 3 A structural diagram for constructing the user profile in Example 1 is provided.

[0059] Figure 4 This is a flowchart of the semantic matching optimization process in Example 1. Detailed Implementation

[0060] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0061] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0062] Secondly, the term "one embodiment" or "embodiment" as used 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 different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0063] Example 1, referring to Figures 1-4 This is the first embodiment of the present invention, which provides an AI-based intelligent airport service question-and-answer method, including the following steps:

[0064] S1. Receive and parse multimodal data input by the user, and build a user profile in real time by combining the user's historical query sequence, browsing behavior and personalized information;

[0065] Specifically, receiving and parsing multimodal data input by users involves removing noise, aligning temporally, and normalizing dimensions of the multimodal data, extracting its features and converting them into a unified text semantic representation. To ensure comparability between different modalities, semantic alignment is performed based on a multimodal fusion model with a self-attention network architecture. A multimodal encoder is then used to convert the features of different modalities into vector representations, which are then concatenated to obtain the multimodal semantic representation vector A'.

[0066] By performing denoising, temporal alignment, and dimensionality normalization on multimodal data, consistency of the original input signal in both time and numerical space is achieved, effectively eliminating interference and scale differences between different modal inputs. Semantic alignment is performed using a multimodal fusion model based on a self-attention network architecture, enabling automatic modeling of implicit semantic relationships between modalities such as text, speech, image, and location information. This enhances the model's ability to understand and associate cross-modal semantics, overcoming the weaknesses of traditional feature concatenation methods in terms of weak semantic connections and poor robustness. Furthermore, a multimodal encoder performs unified mapping of feature vectors from different modalities and completes vector-level concatenation to generate multimodal semantic representation vectors, achieving high-dimensional fusion expression of user input data in a unified semantic space and improving the model's ability to handle complex cross-semantic scenarios.

[0067] Furthermore, combining users' historical query sequences, browsing behavior, and personalized information to build user profiles in real time refers to obtaining users' historical query texts, query timestamps, page access records, and page dwell time from the database to capture users' behavioral patterns and construct behavioral sequence vectors H.

[0068] The behavior sequence vector H is weighted using a time decay function to avoid static user profiles and obtain the final behavior representation vector H'.

[0069] The calculation formula is:

[0070]

[0071] in, The time weight of the action in the ζ-th row, Δε ζ It 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] in, It is the maximum number of rows. h is the time weight of the action in the ζ-th row. ζ It is the row vector of the ζ-th row;

[0075] Using the basic information (age, gender, and language preference) provided by users during registration, we analyze and obtain personalized preference features (user's selected language, preferred service category, and recent query information), and represent the user preference vector B through rule vectorization.

[0076] The final behavior representation vector H' and the user preference vector B are concatenated using a multilayer perceptron to obtain the complete dynamic representation vector U of the user profile.

[0077] By retrieving users' historical query records from the database, a systematic collection and sequence modeling of users' historical interaction behaviors is achieved. This enables the reconstruction of users' behavioral paths and the evolution of their information needs over time. By employing a time decay function to weight the constructed behavioral sequence vector H, enhanced attention is paid to recent behaviors, and the weights of long-term behaviors are adjusted. This effectively avoids the problem of static user profiles and enhances the model's adaptability to changes in user behavior preferences over time, making it more aligned with users' current intentions. By analyzing user registration information and extracting personalized preferences based on recent query service categories, a vectorized expression of users' long-term stable preferences is achieved. Background semantic features that supplement the behavioral sequence are constructed, further enriching the individual dimension of the user profile. The time-weighted behavioral representation vector and the static preference vector B are nonlinearly fused using a multilayer perceptron (MLP), achieving deep coupling between behavior-driven features and user preference features in the semantic space. This results in a more real-time, individualized, and context-aware dynamic representation vector for the user profile.

[0078] S2. Based on the constructed user profile, the system identifies the user's current intent and predicts future behavior paths through a temporal attention mechanism model. Based on the user's input data and current intent, the system uses deep learning to perform TF-IDF weighted calculation and semantic matching. Finally, the system obtains the candidate set of answers through a rough set model and pruning mechanism.

[0079] Specifically, based on the constructed user profile, the user's current intent is identified through a temporal attention mechanism model, and the future behavior path is predicted by concatenating the generated dynamic representation vector U of the user profile and the multimodal semantic representation vector A' into a joint representation vector X.

[0080] The joint representation vector X is linearly transformed using the three sets of weight matrices in the temporal self-attention mechanism to generate the query vector Q, the key vector K, and the value vector V.

[0081]

[0082] in, It is a learnable linear transformation matrix;

[0083] By querying the dot product (similarity) of the query vector Q and the key vector K, and scaling the key vector by the square root of its dimension, a weight is assigned to each input feature (the dynamic representation vector U of the user profile is a dynamic feature, and the multimodal semantic representation vector A' is a static feature) based on the calculated dot product. This weight is then used to weight the value vector V to obtain the weighted value vector. Finally, the weighted values ​​are summed to obtain the key behavioral features for each time step (the input features with the highest weights, which play a major role in user behavior prediction):

[0084]

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

[0086] A multi-head attention mechanism is employed to enhance the model's comprehensive expressive power. Key behavioral features (query, key, and value vectors) are mapped to each head (multiple subspaces), and multiple attention mechanisms are simultaneously computed and analyzed to capture feature representations from different subspaces.

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

[0088] Among them, MultiHead(Q,K,V) represents the comprehensive behavioral features (the different importance of user profiles and multimodal semantic features in multiple subspaces), and head m It is the calculated m-th head, D O It is an output mapping matrix that connects the various headers and generates the final output;

[0089] By capturing feature representations from different subspaces, the user's current behavioral pattern is identified, and the comprehensive behavioral features are mapped to the corresponding current intent category vector y through a fully connected layer:

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

[0091] Where y is the current intent category vector, and W intent The weights of y are set through machine learning and regression models, b intent It is the bias term of y, and MultiHead(Q,K,V) is the comprehensive behavioral feature;

[0092] Based on the current intent category vector y, an attention mechanism is used to select the position in the behavior sequence (the behavior that will affect future behavior) to indicate, and a pointer mechanism in a pointer network is used to predict the next behavior path of the generating user.

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

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

[0095] The predicted next behavioral path P next Merge 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.

[0096] By concatenating dynamic vectors of user profiles with multimodal semantic vectors, a joint representation is constructed, achieving unified modeling of users' historical behavior and current input. This provides complete contextual information for intent recognition. A temporal self-attention mechanism is employed to calculate the importance weights of input features, extract key behavioral features, effectively mine the core information driving user intent, and enhance the model's temporal sensitivity and behavior discrimination ability. Combined with a multi-head attention mechanism, semantic differences are captured from multiple subspaces, improving the expressiveness of users' behavioral preferences in different dimensions and making intent recognition results more refined and accurate. By mapping comprehensive behavioral features to intent vectors and combining them with pointer networks to predict users' next behavioral path, the ability to infer future user intent is strengthened. Ultimately, end-to-end modeling from understanding users' current intent to predicting future behavior is achieved, significantly improving the semantic understanding depth, responsiveness, and personalized interactive experience of question answering.

[0097] Furthermore, based on user input data and current intent, deep learning is used for fusion computation to perform TF-IDF weighted summation and semantic matching. A rough set model and pruning mechanism are then used to obtain the final candidate answer set. This involves inputting the multimodal semantic representation vector A' and the future path intent vector y' into the deep learning model for fusion computation to obtain the final joint semantic representation X'.

[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 multimodal semantic representation vector A' after processing by a convolutional neural network, and Z... intent (y' intent ) is the vector representation of the future path intention vector y' after processing by a convolutional neural network;

[0100] Based on the matching of the final joint semantic representation X' with the knowledge base, the final joint semantic representation X' is segmented into individual words using the BERT word segmentation tool. Part-of-speech tagging is then used to label each word in the text with its corresponding part of speech (noun, verb, adjective). Nouns and verbs are retained as candidate keywords. TF-IDF weighted calculation is then performed on the candidate keywords to accurately assess the importance of each keyword in the text.

[0101]

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

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

[0104] A sliding window mechanism is used, with a window size of g. Each time, g keywords are selected consecutively as a phrase. The window is continuously slid across the entire feature keyword sequence to generate a continuous set of feature phrases until all possible triples (feature phrases) in the sequence are extracted. The TF-IDF weights of all keywords in each feature phrase are weighted and summed to obtain the comprehensive weight of the feature phrase, reflecting its semantic importance in the entire text. The comprehensive weights of all feature phrases are calculated and sorted in descending order to obtain the feature phrase set a. β ;

[0105] The feature phrase set a β The words are decomposed into independent words. The similarity between each independent word and its semantic concept is calculated using the HowNet semantic knowledge base. The root node of the semantic tree is constructed with a depth of 0. Starting from the root node, each semantic concept is represented as a node in the tree structure. The depth of the direct child node under the root node is 1. The depth of the child node of each child node is 1 based on the depth of the previous node. The path distance is calculated based on the number of edges required from each concept node to the root node.

[0106] The formula for calculating the semantic concept Sim(c1,c2) ​​is as follows:

[0107]

[0108] Here, c1 and c2 are the two concepts corresponding to the words, and o1 and o2 are the positions of these two concepts in the semantic tree. O1 and O2 are the depths of these two concept nodes (i.e., their levels in the semantic tree), and distance(o1, o2) is the path distance between the two concept nodes in the semantic tree, representing the semantic relationship distance between them. It is an adjustment factor used to control the weights of semantic depth and path distance;

[0109] By heuristically setting a similarity threshold ψ using a knowledge base, when the similarity Sim(c1,c2) ​​is greater than or equal to ψ, the terms are semantically close enough to be replaced, generating a new semantic feature phrase a' containing the concept. e Otherwise, no replacement will be performed;

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

[0111]

[0112] in, It is a semantic feature phrase a β The first in ' A phrase, yes The encoding is p, 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, and x is the index;

[0113] For semantic feature phrase a β Each phrase in ' is hashed using a Karp-Rabin function and mapped to a unique digital fingerprint, forming a semantic fingerprint set F. input This represents the core semantic features of the text, which can be used to quickly match knowledge items in the knowledge base;

[0114] By using a dual-tower network to encode the final joint semantic representation X' and each knowledge item in the knowledge base, a feature vector representation is generated and the similarity is calculated. During the similarity calculation process, the semantic fingerprint set F is utilized. input Influences the generation of feature vectors, indirectly improving the accuracy of similarity calculation:

[0115]

[0116] Where Sim(X',G j ) represents semantic similarity, G j V is the j-th knowledge item in knowledge base G. X' , It is a feature vector representation generated through a dual-tower network. It is the semantic fingerprint of the final joint semantic representation X'. It is G j The semantic fingerprint, where Π and Ω are weighting factors, is set using a genetic algorithm;

[0117] By calculating semantic similarity, candidate answers that match the joint semantic representation X' in the knowledge base are obtained and set into a candidate answer set W;

[0118] The candidate answer set W is further optimized using a probabilistic tolerance rough set model. The rough set model uses tolerance relations to handle ambiguity and redundancy between text and candidate knowledge items, enhancing the system's robustness. Its upper and lower approximation sets are constructed as follows:

[0119]

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

[0121] Among them, U(S,w i L(S,w) is an upper approximate set. i ) is the lower approximation set, S is the similarity measure, and θ is the lower approximation set. i The currently being processed words i and θ f Is with θ i The word f for comparison δ is the tolerance level, set using rules of thumb; δ is the confidence level, set using cross-validation. δ is used to control the flexibility and accuracy of similarity;

[0122] Based on semantic feature phrase a β 'and candidate answer item W ξ The semantic feature phrase a is measured by the size of the intersection and union of the upper and lower approximate sets. β 'and candidate answer item W ξ The tolerance similarity effectively handles semantic ambiguity and redundant information between words, improving the robustness and accuracy of similarity calculation:

[0123]

[0124] Among them, a β 'Semantic feature phrase, W' ξ It is the ξ-th candidate option in the candidate answer set W, U(S,a) β ') and E(S,a β ') is the upper and lower approximation set of the feature phrase, U(S,W) ξ ) and E(S,W ξ () is the upper and lower approximate sets of candidate knowledge items;

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

[0126] The selected answer candidates are used to generate search-based answers and decode them. During the decoding stage, a dynamic pruning mechanism is used to remove redundant answer candidates and reduce the computational burden of the decoding attention channel, thereby accelerating the output of the answer candidate set W'.

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

[0128]

[0129] Among them, I γ It is the pruning decision for the γth attention channel, λ γ The weight of the γth attention channel is set through multi-objective optimization. τ is the average weight of the channel, and τ is the pruning threshold, which is set by the weight magnitude.

[0130] λ γ The calculation formula is:

[0131]

[0132] Among them, Q γ and K γ These are the query vector and key vector of the γth attention channel. It is the square root of the dimension of the key vector;

[0133] The calculation formula is:

[0134]

[0135] in, Is the γth channel in the candidate answer The weight Λ is the total number of candidate answers in the candidate answer set W;

[0136] Generative models are used to generate φ generative candidate answers from the final joint semantic representation X' (generative models include generative adversarial networks, autoregressive models, variational autoencoders, stream models, and diffusion models; here, an autoregressive model is chosen) and the Beam Search method is used to select from the generative models. The final candidate answer set is output by combining the source of the answers (retrieval-based and generative) in the candidate answer set W', the semantic similarity between the input and the answers, and the degree of pruning under the current input.

[0137] By fusing multimodal semantic representation vectors with user future path intent vectors as input to a deep model, a joint expression of current input semantics and behavior prediction results is achieved, enhancing the question-answering system's ability to grasp the user's comprehensive intent. Convolutional neural networks are used to encode the two types of vectors separately and generate a unified joint semantic representation, strengthening the model's ability to capture contextual relationships and making subsequent knowledge matching processes more semantically consistent. BERT segmentation and TF-IDF weighting are employed to extract high-value keywords, and a sliding window is used to construct semantic phrases, enabling automatic extraction of representative semantic segments from the text. This provides sparse semantic representations for the matching process, effectively improving computational efficiency. Combining HowNet semantic trees and path distance calculations, structural modeling of deep semantic relationships between words is achieved; based on this, a semantic index is constructed. After being processed by Karp-Rabin hashing, the fingerprint can be quickly indexed and compared, significantly reducing the search overhead of knowledge item matching. Semantic fingerprints are introduced into the feature representation generation through a dual-tower network, and candidate answers are selected based on vector similarity, improving the accuracy of semantic matching and the relevance of answers. This makes the candidate set more focused on the semantic target. A rough set tolerance model is used to construct upper and lower approximate sets and handle ambiguity tolerance for the candidate set, effectively alleviating the ambiguity and overlap problems in natural language semantic expression and enhancing the robustness and practicality of the system. Finally, combined with a dynamic pruning mechanism, redundant answer paths are identified and eliminated in the decoding stage, realizing structural compression and efficiency optimization of the candidate answer decoding process. At the same time, BeamSearch is used to generate the optimal candidate set, ensuring that the output results have the best performance in terms of relevance, simplicity and semantic coverage.

[0138] S3. Calculate the confidence level of the answer and select the answer with the best confidence level, and continuously optimize the question-answering strategy by combining reinforcement learning.

[0139] Specifically, by calculating the confidence score of the responses and selecting the response with the highest confidence score, the candidate set is calculated using the multimodal semantic representation vector A'. The Middle The semantic matching confidence, contextual consistency score, and generation quality score of each candidate answer;

[0140] The semantic matching confidence calculation uses a pre-trained BERT model to combine A' and... Encode the code and calculate the cosine similarity to obtain the first... Semantic matching confidence of each candidate answer;

[0141] The context consistency score is calculated using the BERTS score. Each candidate response has a contextual consistency score, which is used to evaluate whether the candidate response 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] The comprehensive score for each candidate answer is calculated by weighting the results of semantic matching confidence, context consistency score and generation quality score.

[0144] By dynamically distributing and adaptively setting a threshold q, candidate answers with scores greater than the threshold q are retained and sorted in descending order. The top C candidate answers with the highest scores are retained as backup suggestions and displayed to the user; otherwise, the user is prompted to re-enter the question.

[0145] By performing semantic matching, contextual consistency, and multi-dimensional scoring of candidate responses based on multimodal semantic representation vectors, a fine-grained evaluation mechanism for candidate response quality is achieved. This mechanism not only identifies semantic fit but also judges the coherence and quality of the response, improving question-and-answer accuracy and contextual adaptability. By calling a pre-trained BERT model and using cosine similarity to calculate semantic matching confidence, a deep semantic relevance between user input and candidate responses is measured, effectively filtering out low-quality responses with semantic deviations and ensuring that the generated results are semantically closely related. The BERTScore is introduced to evaluate contextual consistency, enabling a more nuanced assessment of the response quality. The logical connection assessment between the answer and the dialogue history avoids semantic jumps or contextual breaks, which helps generate more continuous and fluent answers for users, thus improving the overall dialogue quality. Combined with the ROUGE metric to calculate the quality score, it realizes the assessment of the structural overlap between the answer and the standard reference answer, ensuring the integrity and coverage of the generated results. In generative question answering, it effectively improves readability and content coverage. Finally, by weighting and summarizing the scores of each dimension and setting a threshold filtering mechanism, it ensures that only candidates with high overall quality are displayed. When the scores are generally insufficient, the user can be prompted to re-enter, which enhances response control and user experience consistency.

[0146] Furthermore, by continuously optimizing the question-answering strategy using reinforcement learning, after completing confidence screening and outputting the answer, the question-answering interaction is formalized as a Markov decision process modeled through states, actions, and rewards:

[0147]

[0148] Wherein, state μ t The environmental state at time t is jointly composed of historical question-and-answer pairs, the current dialogue context, and user profile vectors, with action ε. t In state μ t Below, the model selects an answer strategy, such as rewarding a selected candidate answer. It is a feedback signal provided by the user after time t;

[0149] Use policy networks to optimize the question-answering decision-making process. Output action distribution (generating a series of possible response strategies based on the current state), and use proximal policy optimization to train and optimize the policy network. The set of learnable parameters in

[0150]

[0151] in, It is the loss function for near-end policy optimization, used to evaluate the difference between the current policy and the old policy, and to guide the optimization of the policy. t It is the time step operator, representing an operation on time t. It is the strategy update ratio, M t is the advantage function, estimated by GAE, used to evaluate the superiority of each action choice to guide the policy update direction, ∈ is the clipping coefficient, set by an adaptive policy to prevent drastic policy updates and ensure training stability;

[0152] For loss function Minimize the negative numbers and use the Adam optimizer on the set of learnable parameters in the policy network. Update;

[0153] Building upon policy network optimization, an experience replay pool is used to specifically store high-scoring feedback trajectories, enhancing the system's ability to memorize excellent answer paths. A threshold is set using a fixed-ratio screening method. After each interaction, if a reward is given Greater than the threshold Current trajectory Add a high-quality replay pool. Every v training iterations, sample a batch of high-quality trajectories from the pool and update them using a target reinforcement strategy:

[0154]

[0155] in, It is the gradient used to update the policy parameters. It is the probability of choosing an action in state μ, and the gradient is... For strategy parameters The derivative, When the parameter is strategy Below, the expected total reward obtained from the environment, It is a reward value;

[0156] Based on the gradient of the calculated update policy parameters The gradient ascent strategy is used to optimize the parameter set of the policy network. Perform iterative updates;

[0157] By combining A / B testing, multiple trained strategy versions are assigned to different user subgroups. Feedback data (such as like rate and bounce rate) is collected for both strategies with an equal number of users. The difference in feedback values ​​between the new and old strategies is calculated. If the feedback value of the new strategy is greater than that of the old strategy, then the new strategy is switched to live.

[0158] By formalizing the question-answering process as a Markov decision process, an optimizable state-action-reward mechanism is implemented, enabling the model to adaptively select response strategies based on historical question-answer pairs, current context, and user profiles. By employing a policy network and using the Proximal Policy Optimization (PPO) algorithm for training, policy performance is effectively improved while maintaining policy update stability, avoiding instability in question-answer quality due to drastic policy changes. This enhances reliability and learning ability in real-world interaction scenarios. Through the design of a dominance function and clipping loss, the policy network prioritizes learning action selection paths with superior performance, improving the identification and learning efficiency of high-value answer strategies, thus enhancing the policy network's ability to respond effectively. During the update process, it can continuously converge towards a better response, increasing the upper limit of answer quality. By designing an experience replay pool and storing historical question-and-answer trajectories with good user feedback, and periodically sampling for optimization of the target strategy, it achieves long-term memory and reuse of high-quality interaction experiences, improves the utilization efficiency of training samples and the generalization ability of the model, and avoids the model getting stuck in local optimal strategy paths. Combined with the A / B testing mechanism, multiple strategy versions are deployed to different user groups for comparative experiments, realizing the quantitative evaluation of the performance differences between the old and new strategies and the construction of the basis for dynamic switching, ensuring that the new strategy performs better than the old version, significantly improving the scientific nature of question-and-answer strategy upgrades and the stability of user experience.

[0159] This embodiment also provides an AI-based intelligent airport service question-and-answer system, including:

[0160] The multimodal data processing module is used to perform denoising, temporal alignment, and dimensionality normalization on multimodal data, extract features, and transform them into a unified text semantic representation.

[0161] The user profile building module is used to build user profiles based on historical queries, browsing behavior, and personalized information, and generate dynamic representation vectors.

[0162] The intent recognition and behavior prediction module is used to identify the current user intent and predict future behavior paths based on user profiles and multimodal semantic representations, using a temporal self-attention mechanism.

[0163] The semantic matching module is used to match semantic representations with the knowledge base, perform rough set optimization and dynamic pruning, and generate the final set of candidate answers.

[0164] The answer optimization and strategy module is used to calculate confidence and optimize strategies, thereby improving response performance by optimizing question-answering decisions through reinforcement learning.

[0165] This embodiment also provides a computer device applicable to the AI-based airport smart service question-and-answer 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 smart service question-and-answer method proposed in the above embodiment.

[0166] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0167] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the AI-based airport smart service question-answering method proposed in the above embodiments. 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), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0168] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An AI-based airport intelligent service question and answer method, characterized in that: include, Receive and parse multimodal data input by users, and build user profiles in real time by combining users' historical query sequences, browsing behavior and personalized information; The multimodal data includes text, voice, images, and location information; The personalized information includes language preferences and user type; Based on the constructed user profile, the system identifies the user's current intent and predicts future behavior paths through a temporal attention mechanism model. Based on the user's input data and current intent, it uses deep learning to perform TF-IDF weighted calculation and semantic matching, and obtains the final candidate set of answers through a rough set model and pruning mechanism. By calculating the confidence level of the answer and selecting the answer with the best confidence level, the question-answering strategy is continuously optimized by combining reinforcement learning. The constructed user portrait is used to identify the current intention of the user through a time sequence attention mechanism model, and a future behavior path is predicted to generate a joint representation vector X; The three sets of weight matrices in the time sequence self-attention mechanism are used to linearly transform the joint representation vector X to generate a query vector , a key vector , and a value vector , to obtain the key behavior features at each time step. A multi-head attention mechanism is employed to map the query, key, and value vectors of key behavioral features to each head, while simultaneously calculating and analyzing multiple attentions to capture feature representations from different subspaces and map them into comprehensive behavioral features. ; By capturing feature representations from different subspaces, the user's current behavioral patterns are identified, and the comprehensive behavioral features are integrated through a fully connected layer. Mapped to the corresponding current intent category vector y; Based on the current intent category vector It selects the position in the action sequence through an attention mechanism and uses the pointer mechanism in a pointer network to predict the next action path of the generating user. ; Predict the next behavioral path This is combined with the current intent category vector y to generate an intent vector that contains the current user's intent and future behavior prediction information. ; Based on user input data and current intent, deep learning is used for fusion computation, TF-IDF weighted summation and semantic matching are performed, and the final candidate answer set is obtained through rough set model and pruning mechanism, which represents the multimodal semantic representation vector. and future path intent vector The input is fused into a deep learning model to obtain the final joint semantic representation. ; Based on the final joint semantic representation The data is matched against a knowledge base, and the final joint semantic representation is obtained using the BERT word segmentation tool. The word segmentation process divides the word into individual words, labels their corresponding parts of speech, retains nouns and verbs as candidate keywords, and performs TF-IDF weighted calculation to obtain the TF-IDF weight of word i. Calculate the weight of each keyword and generate a sequence of feature keywords based on the weights; Set the window size to g, and select g keywords consecutively as a phrase. Extract all triples from the feature keyword sequence. Calculate the comprehensive weight of the feature phrase by weighted summing of 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 feature phrase set. ; Set of feature phrases The words are broken down into independent words, and the similarity between each independent word and its semantic concept is calculated using the HowNet semantic knowledge base. Perform a match; Set similarity threshold When similarity Greater than or equal to If they are semantically close enough, they can be replaced to generate new semantic feature phrases containing the concept. Otherwise, no replacement will be performed; Each semantic feature phrase is converted into a unique, fixed-length fingerprint using the Karp-Rabin hash function. For semantic feature phrases Each phrase in the dataset is hashed using a Karp-Rabin function and mapped to a unique digital fingerprint, forming a semantic fingerprint set. ; Using a dual-tower network to encode the final joint semantic representation For each knowledge item in the knowledge base, generate a feature vector representation and combine it with a semantic fingerprint set. Calculate semantic similarity ; By calculating semantic similarity, we obtain the joint semantic representation in the knowledge base. Matched answer candidates are grouped into a candidate answer set W; The candidate answer set W is further optimized using a probabilistic tolerance rough set model by constructing upper and lower approximation sets. , And based on semantic feature phrases and candidate answers The tolerance similarity is calculated by measuring the size of the intersection and union of the upper and lower approximate sets. ; Based on tolerance similarity calculation, a similarity threshold is set. When the tolerance similarity of the candidates is greater than or equal to the threshold If the answer is correct, the candidate answer is retained in the candidate answer set W and sorted in descending order; otherwise, it is removed from the candidate answer set W. The selected candidate answers are used to generate search-based answers and decode them. During the decoding stage, a dynamic pruning mechanism is used to calculate the pruning decision for each attention channel, eliminating redundant candidate answers and decoding the attention channels to output a set of candidate answers. ; The final joint semantic representation is obtained through a generative model. generate 10 candidate answers were selected from the generative model using the Beam Search method. A generative candidate answer, combined with the candidate answer set The system annotates the sources of the answers, the semantic similarity between the input and the answer, and the degree of pruning under the current input, and outputs the final set of candidate answers. .

2. The AI-based airport smart service question-answering method as described in claim 1, characterized in that: The process of receiving and parsing multimodal data input from users involves removing noise, aligning temporal sequences, and normalizing dimensions from the multimodal data. Features are extracted and transformed into a unified text semantic representation. Semantic alignment is then performed using a multimodal fusion model based on a self-attention network architecture. A multimodal encoder is employed to convert features from different modalities into vector representations, which are then concatenated to obtain a multimodal semantic representation vector. .

3. The AI-based airport smart service question-answering method as described in claim 2, characterized in that: The real-time construction of user profiles by combining user historical query sequences, browsing behavior and personalized information refers to obtaining user historical query text, query timestamps, page access records and page dwell time from the database to construct a behavior sequence vector H. The behavior sequence vector H is weighted using a time decay function to obtain the final behavior representation vector. ; By utilizing the basic information provided by users during registration, personalized preference features are analyzed and obtained, and user preference vector B is represented by rule vectorization. The final behavior representation vector The user preference vector B is concatenated with the user preference vector B through a multilayer perceptron to obtain a complete dynamic representation vector of the user profile. .

4. The AI-based airport smart service question-answering method as described in claim 3, characterized in that: The step of calculating the confidence score of the response and selecting the response with the optimal confidence score refers to using a multimodal semantic representation vector. Compute candidate set The Middle The semantic matching confidence, contextual consistency score, and generation quality score of each candidate answer; The semantic matching confidence is obtained through a pre-trained BERT model. and Encode the code and calculate the cosine similarity to obtain the first... Semantic matching confidence of each candidate answer; The context consistency score is calculated using the BERTS score. Each candidate answer's contextual consistency score; The generated quality score uses ROUGE to calculate the overlap between the generated text and the reference text; The comprehensive score for each candidate answer is calculated by weighting the results of semantic matching confidence, context consistency score and generation quality score. Set a threshold q, retain candidate answers with scores greater than the threshold q and sort them in descending order, retain the top C candidate answers with the highest scores as backup suggestions to show to the user, otherwise prompt the user to re-enter the question.

5. The AI-based intelligent airport service question-answering method as described in claim 4, characterized in that: The aforementioned combination of reinforcement learning to continuously optimize the question-answering strategy refers to formalizing the question-answering interaction into a Markov decision process modeled through states, actions, and rewards after completing confidence screening and outputting answers. Use policy networks to optimize the question-answering decision-making process. Output the action distribution and use a proximal strategy to train and optimize the policy network. The set of learnable parameters in ; Based on policy network optimization, a high-scoring feedback trajectory is specifically stored through an experience replay pool, and a threshold is set. After each interaction, if a reward is given Greater than the threshold , the current trajectory A high-quality replay pool is added. Every v training iterations, a batch of high-quality trajectories is sampled from the pool, and the gradients of the policy parameters are updated using the target reinforcement policy. For parameter set Perform iterative updates; By combining A / B testing, multiple trained strategy versions are assigned to different user subgroups. Feedback data from both strategies is collected with an equal number of users. The difference in feedback values ​​between the new and old strategies is calculated. If the feedback value of the new strategy is greater than that of the old strategy, then the new strategy is switched to live.

6. An AI-based intelligent airport service question-and-answer system, based on the AI-based intelligent airport service question-and-answer method according to any one of claims 1 to 5, characterized in that: include, The multimodal data processing module is used to perform denoising, temporal alignment, and dimensionality normalization on multimodal data, extract features, and transform them into a unified text semantic representation. The user profile building module is used to build user profiles based on historical queries, browsing behavior, and personalized information, and generate dynamic representation vectors. The intent recognition and behavior prediction module is used to identify the current user intent and predict future behavior paths based on user profiles and multimodal semantic representations, using a temporal self-attention mechanism. The semantic matching module is used to match semantic representations with the knowledge base, perform rough set optimization and dynamic pruning, and generate the final set of candidate answers. The answer optimization and strategy module is used to calculate confidence and optimize strategies, thereby improving response performance by optimizing question-answering decisions through reinforcement learning.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the AI-based airport smart service question-answering method according to any one of claims 1 to 6.

8. 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 the AI-based airport smart service question-answering method as described in any one of claims 1 to 6.

Images