Intelligent service execution method based on dialogue mechanism

Abstract

The invention provides an intelligent service execution method based on a dialogue mechanism, and the method comprises the steps: receiving multi-modal input data of a user, and carrying out the preprocessing of the multi-modal input data of the user; a multi-modal graph network is constructed hierarchically to realize cross-modal reasoning of user intentions; constructing a historical attention weight, associating historical request data of the user, and optimizing an answer result of the current dialogue; constructing a task hypergraph to dynamically disassemble the complex task and planning a specific execution path between the subtasks; designing a feedback mechanism based on reinforcement learning to realize self-optimization of the system; interaction logs of the user and the system are maintained, and personalized services are achieved. Through multi-modal data fusion, dynamic task planning and historical feedback optimization technologies, the request understanding accuracy, the task execution efficiency and the system adaptive capacity are remarkably improved, and efficient and flexible service processing support is provided for an intelligent dialogue system.

Description

Technical Field

[0001] This invention relates to the fields of natural language processing, large models, and digital humans, specifically to an intelligent business execution method based on a dialogue mechanism, which improves the accuracy of request understanding, task execution efficiency, and system adaptability, and automatically identifies and plans user requests of different modalities, providing staff with efficient and flexible business processing support. Background Technology

[0002] Natural Language Processing (NLP) is solving an increasing number of practical applications, with the accuracy and efficiency of business execution in the field of intelligent question answering relying heavily on accurate cross-modal NLP technology. The diverse types of business requests and varying input modalities make execution challenging. Effective cross-modal preprocessing of the input request data is crucial for identifying the intent of different modalities and improving the accuracy of intelligent business responses. With the rapid development of artificial intelligence, intelligent dialogue systems are increasingly widely used in finance, customer service, and healthcare. However, existing traditional systems lack longitudinal training specifically for executing business requests, resulting in low efficiency and a certain number of incorrect answers. With the rapid development of NLP and artificial intelligence technologies, dialogue-based intelligent business execution methods can automatically identify the intent of different modalities from user request data and provide optimal answers, offering critical information support for intelligent business execution and NLP.

[0003] Traditional business execution systems still suffer from limited noise reduction and cross-modal intent reasoning capabilities for multimodal inputs, poor semantic understanding accuracy, unclear task scheduling, and insufficient personalized services. Typically, business execution systems use unimodal intent recognition for user requests, making it difficult to capture the correlation between multimodal data. At the same time, task scheduling strategies cannot efficiently handle coupled tasks, and existing dialogue systems mostly rely on predefined templates, making it difficult to achieve adaptive improvement through user feedback.

[0004] To address the shortcomings of the aforementioned intelligent question-answering systems, this invention proposes a multimodal preprocessing technique to accurately process user input data, providing a high-fidelity input data foundation for cross-modal semantic understanding, thereby improving the accuracy of intelligent business execution. Through an innovative cross-modal alignment module and intent enhancement network, heterogeneous data, including but not limited to text, speech, and images, are mapped to a unified semantic space. A dynamically updated multimodal intent graph is constructed using a graph attention mechanism, effectively solving the problems of modality gap and inaccurate semantic understanding in traditional systems. A historical attention mechanism is designed to intelligently aggregate dialogue context and track state by associating historical dialogue information, better maintaining the continuity of dialogue history. A hypergraph decomposition algorithm is introduced to decompose complex business processes into atomic subtasks and their dependencies. A hybrid reward function integrating task completion, semantic alignment, and policy diversity is designed, and end-to-end optimization of multi-task scheduling is achieved through reinforcement learning algorithms. Dynamically updated user profiles enable personalized calibration of dialogue strategies, constructing an efficient, accurate, and personalized intelligent business execution system and method. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent business execution method based on a dialogue mechanism. By using multimodal request data preprocessing operations and efficient task planning algorithms, it can automatically identify and plan user requests of different modalities, and simultaneously maintain interaction logs for self-optimization.

[0006] The objective of this invention is achieved through the following technical solution:

[0007] This intelligent business execution method based on a dialogue mechanism includes the following steps:

[0008] S1: Based on the multimodal data input by the user, construct a multimodal preprocessor to perform data preprocessing operations including word segmentation and noise reduction;

[0009] S2: Utilize a joint reasoning mechanism based on multimodal intent to perform cross-modal intent recognition processing on preprocessed multimodal request data;

[0010] S3: Associate and match the intent data of the current user request with the user's historical request intent data, and optimize the system's response based on the historical information;

[0011] S4: By constructing a task hypergraph, the user's current request is dynamically broken down and the execution paths between subtasks are planned;

[0012] S5: Construct a historical multi-turn dialogue feedback optimization algorithm based on reinforcement learning and optimize the system's business processing capabilities based on user feedback;

[0013] S6: Personalized optimization of the system through maintenance of historical dialogue records and interaction logs.

[0014] Furthermore, step S1 specifically includes:

[0015] S101: The system receives multimodal service requests initiated by users through at least one form of expression: text, voice, or image.

[0016] S102: For input text data, remove data noise in the user's input text request to improve semantic understanding accuracy;

[0017] S103: For input voice data, remove environmental noise from user input voice requests to improve speech recognition accuracy;

[0018] S104: For input image data, remove noise from the user's input image request to improve image recognition accuracy;

[0019] Furthermore, step S102 specifically includes:

[0020] (1) Divide each word in the corpus into basic units, and iterate through all consecutive word pairs in the corpus. i ,w j The frequency of occurrence of word pairs is counted and denoted as: f(w) i ,w j );

[0021] (2) Select the most frequent word pairs (w i ,w j Merge them to generate new subwords w ij And set the merging threshold θ;

[0022] (3) Repeat the above operation to traverse the remaining word pairs until the frequency of all word pairs is less than the threshold θ, and obtain a new word pair w;

[0023] (4) By embedding a new sub-word context window N (w) The semantic weighted information is used to represent the word w as an n-dimensional vector v. The formula for vector embedding is:

[0024] v=L(γ*H(E(w))+(1-γ)*∑ c∈N(w) α c *H(E(c)))

[0025]

[0026] ReLU6 = min(ReLU, 6)

[0027]

[0028] Where L(·) is the normalization function; H(·) is the ReLU6-based nonlinear activation function; E(w), E(c), and E(c') represent the embedding vectors of the target word, its context window word, and candidate word, respectively; γ is a scaling factor used to balance the initial embedding contribution; α c Attention weights of target word w to specific context word c;<E(w),E(c)> This represents the inner product of two word vectors, used to measure similarity; <·> represents the vector inner product operation; τ is a smoothing parameter used to adjust the smoothness of the attention distribution.

[0029] (5) Further optimize the word combination based on the Viterbi algorithm. The method for finding the optimal word combination is as follows:

[0030]

[0031] Where W = {w1, w2, ..., w T} represents the optimized sub-vocabulary, T represents the length of the sub-word sequence in the sub-vocabulary W, and w i P(w) represents the i-th word combination. i |w i-1 ) and P(q i |w i ) represent the transition probability and the corresponding observation unit q, respectively. i The probability of observation;

[0032] (6) Calculate the semantic relevance of the optimal word combination and set a semantic relevance threshold to maximize the overall semantic similarity S of the sentence to obtain the final word combination. The formula for calculating the semantic relevance of sub-words is:

[0033]

[0034] s(w i ,w i+1 )=exp(β· <v i ,v i+1 >)

[0035] Where s(·) is the semantic similarity function; v i and v i+1 They represent the subword w respectively i and w i+1 vector representation; w i ∈W, where W represents the optimized sub-lexicon; β is the adjustment parameter;

[0036] (7) For word combinations that are less than the semantic relevance threshold, the state weights are readjusted using a state weight adjustment algorithm to obtain the final word. The formula for calculating the state weight adjustment algorithm is as follows:

[0037]

[0038] e((w i ,w j ))=αKL(P wi |P wj )+β(1-cos(v i ,v j ))

[0039] Among them, e((w i ,w j )) is used to retrieve subwords (w i ,w j The semantic error is a function, where δ is the step size parameter, α and β are the error parameters, and P is the error parameter. wi and P wj Sub-words (w) i ,w j The frequency of occurrence of ) KL(P) wi |P wj ) is P wi and P wj KL divergence between; v i and v j Sub-words (w) i ,w j Embedded vectors of )

[0040] (8) For each word w i Define a context sliding window of size k, forming a context set C. i :

[0041] C i ={w i-k ,…,w i-1 ,w i+1 ,…,w i+k}

[0042] (9) Calculate the probability p of each word appearing in the window by using the frequency of each word in the context. j ;

[0043]

[0044] in, C i Each word under the window In context set C i The frequency of appearance in;

[0045] (10) Calculation words In context C i Information entropy in And set the word retention threshold τ;

[0046]

[0047] (11) Based on the calculated information entropy, words are dynamically filtered using the word retention function K(·) to obtain a filtered word sequence of length N: The formula for calculating the word retention function K(·) is:

[0048]

[0049] (12) Calculate the standard edit distance D(w,w′) between the current word w and the candidate word w′ as the basis cost for spell correction;

[0050] (13) Calculate the vector of candidate word w′ in context C i Average word semantic similarity S(w′,C) i ):

[0051]

[0052] Wherein, cos(·) is used to calculate the candidate word w′ and the 2k context words w′ before and after it. j A function representing the cosine similarity between them;

[0053] (14) Define the comprehensive scoring function Score(w,w′), and determine the optimal correction word w based on the adjustable parameter α. * The optimal correction calculation method is as follows:

[0054]

[0055] Score(w,w′)=D(w,w′)-α*S(w′,C i )

[0056] Furthermore, step S103 specifically includes:

[0057] (1) Sample the original speech signal to obtain n discrete sample sequences {x0, x1, ..., x n};

[0058] (2) By using probability amplitude encoding, the normalized sample values ​​are mapped to quantum states |ψ>:

[0059]

[0060] (3) Perform a quantum Fourier transform on the obtained quantum state to obtain a new quantum state |ψ′>. The formula for calculating the quantum Fourier transform is:

[0061]

[0062] Where N represents the dimension of the quantum system, and k represents the summation index;

[0063] (4) Statistically calculate the probability distribution p(|ψ′>) of the transformed quantum state |ψ′> at each time slice t, and calculate the information entropy H(t):

[0064] H(t) = -∑p(f)ln p(f);

[0065] p(f) = p(|ψ′(t)>→∣f>)

[0066] Where p(f) represents the probability that the quantum state collapses to the ground state |f> at time t; |ψ′(t)> represents the quantum state at time t; and p(·) represents the probability function.

[0067] (5) Based on the change of entropy with time or frequency band, define an adaptive threshold function τ(t) to determine high-noise sections. The formula for calculating the adaptive threshold function τ(t) is as follows:

[0068] τ(t)=chx(1-H(t))

[0069] Where chx(·) is the hyperbolic cosine activation function with an adaptive threshold;

[0070] (6) Nonlinear sound field modulation is used to eliminate reverberation in the high-noise section t. The calculation formula for nonlinear sound field modulation is:

[0071] G(∣ψ′(t)>)=∣ψ′(t)>*exp(∣ψ′(t)>)*τ(t)

[0072] Furthermore, step S104 specifically includes:

[0073] (1) Perform L-level discrete wavelet decomposition on the original image I to obtain the low-frequency subband A. l and high-frequency subbands in horizontal, high vertical and high diagonal directions The formula for calculating discrete wavelet decomposition is:

[0074] A l =ds(ds(I,G l (I)) T G l (I))

[0075]

[0076] Where l∈{1,2,…,L}, ds(·) is the image downsampling operation, and G l (I) and G h (I) represents low-channel filtering and high-channel filtering operations on image I, respectively;

[0077] (2) Evaluate each high-frequency subband D using local statistics l Noise level:

[0078]

[0079] Ω (x,y) ={Σ(i,j)|i∈[x-Δx,x+Δx],j∈[y-Δy,y+Δy]}

[0080] Among them, D l (i,j) refers to the coefficient value with coordinates (i,j) in the l-th high-frequency sub-band after wavelet decomposition, μ l (x,y) represents an Ω within a window of size m×m centered at (x,y). (x,y) The local mean of the coordinate points, Δx and Δy are the coordinate offsets within the window, and m represents the side length of the noise statistics window;

[0081] (3) Perform inverse wavelet reconstruction on all sub-bands to obtain the coarsely denoised image.

[0082]

[0083] Where O(·) represents the inverse reconstruction operation of discrete wavelet decomposition:

[0084] O(A l )=us(us(A l G h (A l )) T G h (A l ))

[0085]

[0086] Where us(·) is the image upsampling operation;

[0087] (4) Use bilateral filtering to further reduce noise in image I to obtain the final image I'. The formula for calculating bilateral filtering is:

[0088]

[0089] Where I(x) and I(y) are the pixel values ​​of the original image in the x and y directions, respectively, and ω c (x,y) is the spatial kernel of the image in the x and y directions, ω s (I(x)-I(y)) is the intensity kernel;

[0090] Furthermore, step S2 specifically includes:

[0091] S201: Represent the preprocessed multimodal data, including but not limited to text, speech and image data, as a cross-modal semantic graph;

[0092] S202: Hierarchically construct a multimodal graph network to obtain a complete cross-modal representation K * :

[0093]

[0094] K * =K (1) +K (2) +…+K (l)

[0095] Furthermore, step S201 specifically includes:

[0096] First, expand the user input information in the digital human business request scenario, treat each word as a node, and expand the image frame and audio frame into other modal nodes to form a comprehensive multimodal graph;

[0097] First, the processed user input information in the digital human business request scenario is modally expanded, with each word regarded as a node, and image frames and audio frames expanded into other modal nodes to form a comprehensive multimodal graph;

[0098] Then, establish an edge r between adjacent words on the same modal edge to represent their semantic or syntactic relationship;

[0099] Secondly, edges are established between adjacent modal edges that contain the same or similar words to capture the evolutionary patterns of the context.

[0100] Finally, the adjacency matrix M and degree matrix D are constructed to represent the connection relationships between nodes in the graph. The formula for constructing the adjacency matrix M is:

[0101]

[0102] Where, m (i,j) Let $D$ be the word co-occurrence weights for nodes $i$ and $j$. The degree matrix $D$ is a diagonal matrix, where each diagonal element represents the degree of the corresponding node.

[0103] D (i,j) =∑ j M (i,j)

[0104] Furthermore, step S3 specifically includes:

[0105] S301: Query the user's previous N rounds of dialogue history, and use a pre-trained autoencoder to convert the historical dialogue and the current dialogue into fixed dimensions respectively;

[0106] S302: Select the intermediate dialogue as the key dialogue, and calculate the similarity Sim between each round of dialogue and the key dialogue:

[0107]

[0108] Among them, d, μ represents the current dialogue and the intermediate dialogue, respectively. d , These represent the mean values ​​of the vector representations from the two rounds of dialogue. L1 and L2 represent the variances of the two rounds of dialogue vector representations, respectively; cov(·) is the covariance function; and L1 and L2 are stability constants.

[0109] S303: Set an association threshold and identify dialogues with similarity greater than the association threshold as associated dialogues, and introduce historical dialogue information;

[0110] S304: Aggregate all related dialogues through a historical attention mechanism to generate the final answer;

[0111] Furthermore, step S304 specifically includes:

[0112] First, calculate the historical attention weights for the current dialogue and all n rounds of historical dialogues associated with it:

[0113]

[0114] Where, α i In the historical dialogue, d represents the i-th round of dialogue. i The importance weight, where q represents the current conversation;

[0115] Then, the weighted sum of the historical dialogues is calculated to obtain the fused representation of the historical information Θ:

[0116]

[0117] Finally, the current dialogue information q and the historical information Θ are weighted and linearly fused to obtain the final answer A. * :

[0118] A * =λ*q+(1-λ)*Θ

[0119] Wherein, λ is a learnable weight coefficient that controls the fusion ratio of current dialogue and historical information;

[0120] Furthermore, step S4 specifically includes:

[0121] S401: Perform semantic analysis on user requests that have undergone cross-modal processing and break them down into multiple sub-tasks;

[0122] S402: Treat each subtask as a supernode o∈V, connect them with hyperedges based on the coupling dependencies between subtasks, and assign different dependency weights:

[0123]

[0124] Among them, dep(o i ,o j ) represents a subtask o i and o j The coupling dependency is σ(·), where σ is the Sigmoid function, β is the trainable dependency sensitivity coefficient, and b is the bias term.

[0125] S403: Use a dynamic update algorithm to update the state of hypergraph nodes. The specific formula is as follows:

[0126]

[0127]

[0128] in, Let S(o) represent the state of subtask o after the l-th round of subtask updates, and let S(o) represent the set of neighboring supernodes of subtask o. r and W o Let α represent the trainable parameters of the hyperedge and the hypernode, respectively. Let AG(·) be the hypergraph aggregation function. i Attention weights;

[0129] S404: Calculate the feasibility score and execution cost of a task, and dynamically adjust the task schedule;

[0130] Furthermore, step S404 specifically includes:

[0131] First, calculate the executability score of the task:

[0132] Score(o) = shx(W·h) o +b)

[0133] Where S(o) represents the executability score of task o, reflecting whether the task is suitable for the current path; h o This indicates the state of subtask o after the task is updated; shx(·) are hyperbolic sine activation functions for executability scoring; W and b are learning parameters;

[0134] Then, based on the dynamics of execution costs between tasks, a greedy algorithm is used to search for the task path P with the highest executability, where the formula for calculating the optimal task execution path is:

[0135] c(o i ,o j )=η1*(1-dep(o i ,o j ))+η2*res(o i ,o j )+η3*delay(o i ,o j )+η4*Score(o i )

[0136]

[0137] Wherein, c(o) i ,o j ) indicates task o i and o j The execution costs between tasks are as follows: dep(·), res(·), and delay(·), which represent the execution costs of task o. i and o j The weights of the super-edges, resource consumption items, and real-time latency items are defined by η1, η2, η3, and η4, which are the weight coefficients for each item.

[0138] Furthermore, step S5 specifically includes:

[0139] S501: Embed the user's current dialogue q into the historical dialogue Θ for historical encoding, and construct a dialogue memory matrix set Z:

[0140] Z = {Z1, Z2, ..., Z} k}

[0141] Among them, Z k This represents the memory matrix for the k-th round of dialogue;

[0142] S502: Based on the KL divergence term, a dialogue reward function is constructed in multi-turn dialogues, and its calculation formula is as follows:

[0143]

[0144] Where Q(s,a) and R(s,a) are the dialogue reward function and user feedback score for performing action a in dialogue state s, respectively; KL(p||q) is the KL divergence between the current policy p and the expected goal q; C(·) is the stability score of the current dialogue state; and v s Let m be the embedding vector of the dialogue state s, m be the vector dimension, λ1, λ2, λ3 and λ4 be the weight values ​​of each part, and T be the matrix transpose operation.

[0145] S503: The dialogue strategy is stably optimized through a dual-delay update mechanism, while an entropy regularization term is introduced to maintain strategy diversity.

[0146] S504: Write the user input, system response, and execution result of this round into the user profile;

[0147] Furthermore, step S503 specifically includes:

[0148] First, for each state s and the next state s′, define the timing difference of the double-delay update mechanism:

[0149]

[0150] Where r represents the immediate reward for performing action a in state s, and ξ is the double-delay update parameter. The policy parameters of networks N1 and N2 are evaluated as a function to determine the value of the next state s′;

[0151] Then, for each state-action pair (s, a), the loss function of the double-delay update mechanism is calculated:

[0152]

[0153] Where B represents the batch of dialogues processed by the double-delay update mechanism;

[0154] Finally, after the loss function converges, the state-action pair (s′,a′) is updated with a delay.

[0155] Furthermore, step S6 specifically includes:

[0156] S601: Maintain historical dialogue records and multi-turn interaction logs, record user expression habits and behavioral preferences and perform statistical analysis to form user profiles;

[0157] S602: The system dynamically adjusts the response style based on the user's acceptance of different types of responses, optimizing the length, detail, and information organization of the responses to ensure that each conversation meets the user's personalized needs, thereby improving the interactive experience.

[0158] The beneficial effects of this invention include:

[0159] (1) The multimodal preprocessing technology of the present invention, the semantic sub-word segmentation and optimization algorithm, the quantum speech coding based on chaotic entropy and the discrete wavelet algorithm respectively perform word segmentation, noise reduction and other preprocessing operations on the user's multimodal input data, significantly eliminating noise interference in multimodal input, providing a good data guarantee for subsequent semantic understanding work, and thus achieving the goal of improving the accuracy of intelligent business execution;

[0160] (2) This invention proposes a new cross-modal joint inference method for intent. First, nodes at different modal levels are constructed based on text words, image frames, and audio frames. Then, cross-modal representations are generated by aggregating multimodal network hierarchical features, breaking through the limitations of traditional single-modal intent recognition. Multimodal information such as speech and images are complemented to enhance intent inference capabilities, solve intent ambiguity problems in complex scenarios, ensure the accuracy of the answer results, and improve the user experience.

[0161] (3) This invention proposes a historical dialogue intent association method, which encodes historical dialogues into fixed-dimensional vectors, uses a dialogue similarity calculation function to calculate the similarity between the current dialogue and the historical multi-turn dialogue, and uses historical attention weights to dynamically aggregate historical related information, which can better maintain the continuity of dialogue history, solve the problem of information forgetting or confusion in multi-turn dialogues of intelligent question answering systems, and track the current dialogue status in real time.

[0162] (4) This invention innovatively proposes a dynamic task planning method based on hypergraph theory. By intelligently decomposing user business requests into hypergraph nodes and hyperedges, a dynamically evolving task topology network is constructed. The dynamic update algorithm is used to adjust the task status in real time, enabling the technology to support the flexible decomposition and dynamic scheduling of complex business, avoid resource conflicts, and is especially suitable for high-concurrency and multi-dependency work scenarios.

[0163] (5) This invention deeply integrates the advantages of reinforcement learning technology and constructs a dynamic policy optimization framework based on deep reinforcement learning. The framework takes the dual-delay deep deterministic policy gradient algorithm as the core architecture to realize the adaptive evolution and risk quantification and control of the intelligent business execution system. By suppressing the policy evaluation bias through the dual network structure and combining KL divergence constraints to dynamically limit the update step size of the state-action space, the stability of the policy optimization process is guaranteed, and the local optimum trap can be avoided through the trust domain mechanism. This mechanism can accurately control the parameter update amplitude in the policy iteration process. Under the premise of ensuring that the system risk is controllable (such as the illusion problem generated by large models), the policy performance can be continuously self-optimized, providing a robust and adaptive solution for decision optimization in complex business scenarios.

[0164] (6) This invention makes full use of the advantages of the attention mechanism and designs a variety of different attention to be applied to different steps in the process of intelligent dialogue of digital human. It can better complete the following tasks from different perspectives: the fusion of multimodal inputs, the association of multi-round historical dialogues, and the scheduling between sub-tasks. The attention mechanism enables intelligent terminal devices, including computers, to imitate human thinking and attention patterns for automated and autonomous learning, ensuring the comprehensiveness and accuracy of the intelligent dialogue process of digital human.

[0165] Other advantages, objectives, and features of the present invention will be set forth in detail in the following description. Through a thorough study of the following text, those skilled in the art will be able to clearly recognize these advantages and features and gain valuable lessons from the practice of the invention, whose objectives and other advantages can be realized and embodied in the following description and the previously mentioned claims. Attached Figure Description

[0166] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will now be described in further detail with reference to the accompanying drawings, wherein:

[0167] Figure 1 This is a schematic diagram of the intelligent business system and method based on a dialogue mechanism according to the present invention;

[0168] Figure 2 This is an example diagram illustrating how the semantic word segmentation and optimization algorithm of this invention segments user input requests.

[0169] Figure 3 This is a schematic diagram of the multimodal graph network fusion user request method of the present invention;

[0170] Figure 4 This is an example diagram showing the final user intent recognition result of the present invention;

[0171] Figure 5 This is an example diagram illustrating the subtask division and planning of user requests according to the present invention;

[0172] Figure 6 An example diagram showing the user background and preference settings detected by this invention. Detailed Implementation

[0173] The preferred embodiments of the present invention will now be described in detail. It should be understood that the preferred embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0174] Considering the cross-modal nature of input requests, this invention utilizes comprehensive multimodal data preprocessing and efficient semantic recognition algorithms to identify and classify request data from different modalities. By constructing a hypergraph, it splits and redistributes task flows into subtasks, optimizing task execution paths. Furthermore, it makes targeted optimizations to the current response based on the similarity to the user's historical dialogues. In addition, this invention also incorporates the advantages of reinforcement learning to optimize the response content according to user preferences.

[0175] Attention mechanisms are an important component in deep learning. By combining cross-modal attention mechanisms with reinforcement learning, feature weights can be dynamically optimized, significantly improving the model's ability to extract key business features while suppressing irrelevant noise. By constructing a hierarchical attention network, high-level semantic nodes can focus on low-level detailed features, while low-level features are enhanced through gating attention (integrating high-level context), thus improving the efficiency of multimodal feature representation. This is particularly relevant for multimodal business scenarios.

[0176] like Figure 1 As shown, the present invention provides an intelligent business execution system and method based on a dialogue mechanism, comprising the following steps:

[0177] Step S1: Based on the multimodal data input by the user, construct a multimodal preprocessor to perform data preprocessing operations including word segmentation and noise reduction;

[0178] Step S2: Utilize a joint reasoning mechanism based on multimodal intent to perform cross-modal intent recognition processing on the preprocessed multimodal request data;

[0179] Step S3: Associate and match the intent data of the current user request with the user's historical request intent data, and optimize the system's response based on the historical information;

[0180] Step S4: By constructing a task hypergraph, dynamically break down the user's current request and plan the execution paths between subtasks;

[0181] Step S5: Construct a historical multi-turn dialogue feedback optimization algorithm based on reinforcement learning and optimize the system's business processing capabilities based on user feedback;

[0182] Step S6: Perform personalized optimization of the system by maintaining historical dialogue records and interaction logs.

[0183] It should be noted that any process or method description in the flowcharts of this invention or otherwise described herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process, and the scope of the preferred embodiments of the invention includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order according to the functions involved, as should be understood by those skilled in the art to which the embodiments of the invention pertain.

[0184] like Figures 2 to 6 As shown, the specific steps of the above method will be further explained below through a specific embodiment.

[0185] Step S1 includes the following sub-steps:

[0186] Step S101: The user initiates a credit card loss reporting service through the mobile APP, and inputs the following: voice: "My credit card is lost and needs to be reported lost" (16kHz sampling rate, average sound intensity detected 65dB, speech rate 4.2 words / second), text: "last four digits of card number 3087", image: ID card photo (resolution 1920×1080, JPEG compression rate 85%).

[0187] Step S102: For the input text data, remove data noise from the user's input text request to improve the accuracy of semantic understanding. Specific operations for removing text request noise include:

[0188] First, a semantic word segmentation and optimization algorithm is constructed to segment the initially denoised text data. The steps for constructing the semantic word segmentation and optimization algorithm are as follows:

[0189] (1) Break down each word in the corpus into basic units: {“my”, “credit”, “using card”, “lost”, “lost”, “need”, “to hang”, “to report lost”}, and traverse all consecutive word pairs in the corpus (“credit”, “using card”, “lost”, “lost”, “need”, “to hang”, “to report lost”);

[0190] (2) Select the most frequent sub-word pairs ("credit", "use card", "lost", "lost", "need", "need to report", "report lost") and merge them to generate new sub-words "credit card" and "need to report lost", and set the merging threshold θ = 150;

[0191] (3) Repeat the above operation to traverse the remaining word pairs until the frequency of all word pairs is less than the threshold θ, and obtain new word pairs. Figure 2 This demonstrates the process of using semantic word segmentation and optimization algorithms to segment user input requests.

[0192] (4) By embedding the semantic weighted information of the new sub-word context window, the sub-word "report lost" is represented as a 32-dimensional vector, where the calculation formula for vector embedding is:

[0193] v=L(γ*H(E(w))+(1-γ)*∑ c∈N(w) α c *H(E(c)))

[0194]

[0195] ReLU6 = min(ReLU, 6)

[0196]

[0197] Among them, L(·) is a normalization function, H(·) is a non-linear activation function based on ReLU. E(w), E(c), and E(c’) respectively represent the embedding vectors of the target sub-word: "needs to be reported lost", the sub-word within its context window: "lost", and the candidate sub-word: "credit card": E("needs to be reported lost") = [0.21, -0.33, 0.47, -0.12, 0.85,..., 0.02], E("lost") = [0.15, 0.28, -0.41, 0.37, -0.09,..., 0.18], E("credit card") = [-0.23, 0.44, 0.07, 0.59, 0.12,...,-0.31]; γ is a scaling factor, taken as 0.34 in this example, used to balance the contribution of the initial embedding; <E(w), E(c)> represents the inner product of two word vectors to measure similarity. In this example: <E("needs to be reported lost"), E("credit card")> = 0.21×0.15 + (-0.33)×0.28 +... = 1.27, <E("needs to be reported lost"), E("lost")> = 0.21×(-0.23) + (-0.33)×0.44 +... = 0.88, <E("needs to be reported lost"), E("lost")> = 0.21×0.38 + (-0.33)×(-0.17) +... = 1.05; α c The attention weight of the target sub-word w to the specific context sub-word c. In this example, α ( “ 信用卡”) = 0.38, α ( “ 丢失”) [[ID=IO]]= 0.31; τ is a smoothing parameter, used to adjust the smoothness of the attention distribution, taken as 0.7 in this example;

[0198] The target sub-word: "needs to be reported lost", the 32-dimensional vector v("needs to be reported lost") = [0.26, 0.18, 0.31, -0.05, 0.42,..., 0.12] after normalization;

[0199] (5) Further optimize the sub-word combination based on the Viterbi algorithm. The method to find the optimized sub-word combination is as follows:

[0200]

[0201] Among them, W = {w1, w2,…, w T} represents the optimized sub-word library, T represents the length of the sub-word sequence in the sub-word library W, w i represents the i-th sub-word combination, P(w i |w i-1 ) and P(q i |w i ) respectively represent the transition probability and the observation probability of the corresponding observation unit q[[ID=S5]] i ;

[0202] (6) Calculate the semantic relevance of the optimal word combination and set a semantic relevance threshold to maximize the overall semantic similarity S of the sentence to obtain the final word combination. The formula for calculating the semantic relevance of sub-words is:

[0203]

[0204] s(w i ,w i+1 )=exp(β· <v i ,v i+1 >)

[0205] Where s(·) is the semantic similarity function; v i and v i+1 They represent the subword w respectively i and w i+1 vector representation; w i ∈W, where W represents the optimized sub-vocabulary; β is an adjustment parameter, which is 0.38 in this example;

[0206] (7) For word combinations that are less than the semantic relevance threshold, the state weights are readjusted using a state weight adjustment algorithm to obtain the final word. The formula for calculating the state weight adjustment algorithm is as follows:

[0207]

[0208] e((w i ,w j ))=αKL(P wi |P wj )+β(1-cos(v i ,v j ))

[0209] Among them, e((w i ,w j )) is used to retrieve subwords (w i ,w j The semantic error is a function, which is 0.098 in this example; δ is the step size parameter, which is 0.74 in this example; α and β are the error parameters, which are 0.4 and 0.6 respectively in this example; P wi and P wj Sub-words (w) i ,w j The frequency of occurrence of ) KL(P) wi |P wj ) is P wi and P wj KL divergence between; v i and v jSub-words (w) i ,w j Embedded vectors of )

[0210] Then, based on word segmentation, a chaotic entropy stop word filtering algorithm is constructed to adaptively perceive and analyze the lexical context. The specific steps for constructing the chaotic entropy stop word filtering algorithm include:

[0211] (1) For each word w i Define a context sliding window of size 5, forming a context set C. i The context set for "reporting loss" is C. i = {"Need", "Immediately", "Report Loss", "My", "Credit"};

[0212] (2) Calculate the probability p of each word appearing in the window by using the frequency of each word in the context. j ;

[0213]

[0214] in, For words In the context set C i The frequency of occurrence of p in this embodiment j The values ​​are: {"Need": 0.18, "Immediately": 0.22, "Report Loss": 0.35, "My": 0.15, "Credit": 0.10};

[0215] (3) Calculation words In context C i Information entropy in And set the word retention threshold τ = 1.2;

[0216]

[0217] In this embodiment, F(“report lost”) = 1.52 > τ = 1.2, so it is retained;

[0218] (4) Based on the calculated information entropy, words are dynamically filtered using the word retention function K(·) to obtain a filtered word sequence of length N: The formula for calculating the word retention function K(·) is:

[0219]

[0220] Finally, the text data is spelled correctly using context-aware technology to obtain the final text request. The specific steps of the context-aware technology are as follows:

[0221] First, the standard edit distance D(w,w′) between the current word w and the candidate word w′ is calculated as the basic cost for spell correction;

[0222] Then, the vector of candidate word w′ in context C is calculated. i Average word semantic similarity S(w′,C) i The calculation formula is as follows:

[0223]

[0224] Wherein, cos(·) is used to calculate the candidate word w′ and the 2k context words w′ before and after it. j A function representing the cosine similarity between them;

[0225] Finally, a comprehensive scoring function Score(w,w′) is defined, and the optimal correction word w is determined based on the adjustable parameter α. * The optimal correction calculation method is as follows:

[0226]

[0227] Score(w,w′)=D(w,w′)-α*S(w′,C i )

[0228] In this embodiment, Score(“report lost”,“report lost”) = D(0) - 0.6*S(“report lost”,“report lost”) = 1, Score(“report lost”,“report lost”) = D(1) - 0.6*S(“report lost”,“report lost”) = 0.46;

[0229] Step S103: For the input voice data, remove environmental noise from the user's input voice request to improve the accuracy of voice recognition. The specific steps for removing environmental noise from the voice request include:

[0230] (1) The original speech signal was sampled at 16kHz to obtain a discrete sample sequence {0.32, -0.15, 0.28, ..., -0.66};

[0231] (2) By using probability amplitude encoding, each sample value is normalized and mapped to a quantum state |ψ>:

[0232]

[0233] (3) Perform a quantum Fourier transform on the obtained quantum state to obtain a new quantum state |ψ′>. The formula for calculating the quantum Fourier transform is:

[0234]

[0235] Where N represents the dimension of the quantum system, and k represents the summation index;

[0236] (4) Statistically calculate the probability distribution p(|ψ′>) of the transformed quantum state |ψ′> at each time slice t, and calculate the information entropy H(t):

[0237] H(t) = -∑p(f)ln p(f);

[0238] p(f) = p(|ψ′(t)>→∣f>)

[0239] Where p(f) represents the probability that the quantum state collapses to the ground state |f> at time t; |ψ′(t)> represents the quantum state at time t; and p(·) represents the probability function.

[0240] (5) Based on the change of entropy with time or frequency band, define an adaptive threshold function τ(t) to determine high-noise sections. The formula for calculating the adaptive threshold function τ(t) is as follows:

[0241] τ(t)=chx(1-H(t))

[0242] Where chx(·) is the hyperbolic cosine activation function with an adaptive threshold;

[0243] (6) In the high-noise section t, nonlinear sound field modulation is used to eliminate reverberation. The calculation formula for nonlinear sound field modulation is:

[0244] G(∣ψ′(t)>)=∣ψ′(t)>*exp(∣ψ′(t)>)*τ(t)

[0245] In this embodiment, the entropy value H(t) = 2.1 > τ = 1.8 of frames 50-53 is detected, triggering nonlinear sound field modulation:

[0246] G(t) = 0.28 * exp(0.28) = 0.37

[0247] Step S104: For the input image data, remove noise from the user's input image request. Specific steps include:

[0248] (1) Perform L-level discrete wavelet decomposition on the original image I to obtain the low-frequency subband A. l and high-frequency subbands in horizontal, high vertical and high diagonal directions The formula for calculating discrete wavelet decomposition is:

[0249] A l =ds(ds(I,G l (I)) T G l (I))

[0250]

[0251] Where l∈{1,2,…,L}, ds(·) is the image downsampling operation, and G l (I) and G h (I) represents low-channel filtering and high-channel filtering operations on image I, respectively. In this example, L=3, the downsampling ratio is 66.7%, and the convolution kernel used is:

[0252] The first convolutional kernel: N1 = 256, size 7×7, stride 1;

[0253] The second convolutional kernel: N2 = 128, size 5×5, stride 1;

[0254] The third convolutional kernel: N3 = 64, size 3×3, stride 1;

[0255] (2) Evaluate each high-frequency subband D using local statistics l Noise level:

[0256]

[0257] Ω (x,y) ={Σ(i,j)|i∈[x-Δx,x+Δx],j∈[y-Δy,y+Δy]}

[0258] Among them, D l (i,j) refers to the coefficient value with coordinates (i,j) in the l-th high-frequency sub-band after wavelet decomposition, μ l (x,y) represents an Ω within a window of size m×m centered at (x,y). (x,y) The local mean of the coordinate point, Δx and Δy are the coordinate offsets within the window, and m represents the side length of the noise statistics window; in this embodiment, m is 7, and the noise estimation on the second-level high-frequency sub-band of the pixel at coordinate (128, 128) is as follows:

[0259] (3) Perform inverse wavelet reconstruction on all sub-bands to obtain the coarsely denoised image.

[0260]

[0261] Where O(·) represents the inverse reconstruction operation of discrete wavelet decomposition:

[0262] O(A l )=us(us(A l G h (A l )) T G h (A l ))

[0263]

[0264] Where us(·) represents the image upsampling operation. In this embodiment, the upsampling rate is 166.7%, and the upsampling convolution kernel is:

[0265] The first convolutional kernel: N1 = 64, size 3×3, stride 1;

[0266] The second convolutional kernel: N2 = 128, size 5×5, stride 1;

[0267] The third convolutional kernel: N3 = 256, size 7×7, stride 1;

[0268] (4) Use bilateral filtering to further reduce noise in image I to obtain the final image I'. The formula for calculating bilateral filtering is:

[0269]

[0270] Where I(x) and I(y) are the pixel values ​​of the original image in the x and y directions, respectively, and ω c (x,y) is the spatial kernel of the image in the x and y directions, ω s (I(x)-I(y)) is the intensity kernel;

[0271] Step S2 includes the following sub-steps:

[0272] Step S201: Represent the preprocessed multimodal data, including but not limited to text, speech, and image data, as a cross-modal semantic graph. Figure 3 This demonstrates how a multimodal graph network fuses user requests across modalities. Specific operational steps include:

[0273] First, the processed user input information in the digital human business request scenario is modally expanded, with each word regarded as a node, and image frames and audio frames expanded into other modal nodes to form a comprehensive multimodal graph;

[0274] Then, establish an edge r between adjacent words on the same modal edge to represent their semantic or syntactic relationship;

[0275] Secondly, edges are established between adjacent modal edges that contain the same or similar words to capture the evolutionary patterns of the context.

[0276] Finally, the adjacency matrix M and degree matrix D are constructed to represent the connection relationships between nodes in the graph. The formula for constructing the adjacency matrix M is:

[0277]

[0278] Where, m (i,j) Let $D$ be the word co-occurrence weights for nodes $i$ and $j$. The degree matrix $D$ is a diagonal matrix, where each diagonal element represents the degree of the corresponding node.

[0279] D (i,j) =∑ j M (i,j)

[0280] In this embodiment, the text node is: reported lost / card number / 3087 (weight 0.83), the voice node is: lost / urgent (weight 0.75), and the image node is: ID card border (weight 0.68).

[0281] Step S202: Construct a multimodal graph network hierarchically to obtain a complete cross-modal representation K. * :

[0282]

[0283] K * =K (1) +K (2) +…+K (l)

[0284] Where l represents the number of multimodal data types, in this embodiment, the ReLU activation function is used, the number of layers l = 3, and the complete cross-modal representation H * After decoding, it reads: "My credit card is lost. The last four digits of the card number are 3087. I need to report the loss using my ID." Figure 4 The final user intent recognition results after cross-modal fusion are shown;

[0285] Step S3 includes the following sub-steps:

[0286] Step S301: Query the user's previous N rounds of dialogue history, and use a pre-trained autoencoder to convert the historical dialogue and the current dialogue into fixed dimensions respectively.

[0287] In this embodiment, the current dialogue content is: "My credit card is lost. The last four digits of the card number are 3087. I need to report the loss based on my ID document." The previous two rounds of historical dialogue content were: "I need to apply for sick leave tomorrow because I have a cold" and "Please help me summarize the following documents." The dialogue codes are: [0.32,-0.15,...,0.28].

[0288] Step S302: Select the intermediate dialogue as the key dialogue, and calculate the similarity Sim between each round of dialogue and the key dialogue:

[0289]

[0290] in, These represent the current conversation and the conversation in between, respectively. These represent the mean values ​​of the vector representations from the two rounds of dialogue. represents the variance of the two rounds of dialogue vector representation, cov(·) is the covariance function, and L1 and L2 are stability constants, which are taken as 0.23 and 0.85 respectively in this embodiment. The similarity between the current dialogue and the previous two rounds of dialogue is 0.12 and 0.03 respectively.

[0291] Step S303: Set the association threshold And similarity greater than the association threshold The dialogue is confirmed as a related dialogue, and historical dialogue information is introduced. In this embodiment, since the similarity between the current dialogue and the previous two rounds of dialogue is less than the connection threshold, All judgments determine that the current dialogue is independent dialogue content;

[0292] Step S304: Aggregate all related dialogues through a historical attention mechanism to generate the final answer;

[0293] In step S304, the steps of historical attention aggregation include:

[0294] First, calculate the historical attention weights for the current dialogue and all n rounds of historical dialogues associated with it:

[0295]

[0296] Where, α i In the historical dialogue, d represents the i-th round of dialogue. i The importance weight, where q represents the current conversation;

[0297] Then, the weighted sum of the historical dialogues is calculated to obtain the fused representation of the historical information Θ:

[0298]

[0299] Finally, the current dialogue information q and the historical information Θ are weighted and linearly fused to obtain the final answer A. * :

[0300] A * =λ*q+(1-λ)*Θ

[0301] Wherein, λ is a learnable weight coefficient that controls the fusion ratio of the current dialogue and historical information, and is set to 0.6 in this embodiment;

[0302] Step S4 includes the following sub-steps:

[0303] Step S401: Perform semantic analysis on the user request that has undergone cross-modal processing and break it down into multiple sub-tasks;

[0304] In this embodiment, the subtask set V = {identity verification, card loss registration, card replacement application};

[0305] Step S402: Treat each subtask as a supernode o∈V, connect them with hyperedges based on the coupling dependencies between subtasks, and assign different dependency weights:

[0306]

[0307] Among them, dep(o i ,o j ) represents a subtask o i and o j The coupling dependency is σ(·), which is the Sigmoid function, β is the trainable dependency sensitivity coefficient, and b is the bias term. In this example, β = 0.65, b = 0.2, and the hyperedge E = (verification → loss reporting, w = 0.9), (loss reporting → card replacement, w = 0.85).

[0308] Step S403: Update the state of the hypergraph nodes using a dynamic update algorithm. The specific formula is as follows:

[0309]

[0310] in, Let S(o) represent the state of subtask o after the l-th round of subtask updates, and let S(o) represent the set of neighboring supernodes of subtask o. r and W o Let α represent the trainable parameters of the hyperedge and the hypernode, respectively. Let AG(·) be the hypergraph aggregation function. i In this embodiment, attention weights are used.

[0311] Step S404: Calculate the feasibility score of the task and dynamically adjust the task schedule. The specific operations for adjusting the task schedule include:

[0312] First, calculate the executability score of the task:

[0313] Score(o) = shx(W·h o +b)

[0314] Where S(o) represents the executability score of task o, reflecting whether the task is suitable for the current path; h o The state of subtask o after task update is represented; shx(·) are hyperbolic sine activation functions for executability scoring; W and b are learning parameters. In this embodiment, the W matrix follows a uniform distribution: U(-0.1,0.1), b=0.01, S(complementary card)=shx(0.45*512+0.32)=0.78;

[0315] Then, based on the dynamics of execution costs between tasks, a greedy algorithm is used to search for the task path P with the highest executability, where the formula for calculating the optimal task execution path is:

[0316] c(o i ,o j )=η1*(1-dep(o i ,o j ))+η2*res(o i ,o j )+η3*delay(o i ,o j )+η4*Score(o i )

[0317]

[0318] Wherein, c(o) i ,o j ) indicates task o i and o j The execution costs between tasks are as follows: dep(·), res(·), and delay(·), which represent the execution costs of task o. i and o j The weights of the superedges, resource consumption, and real-time latency are defined by η1, η2, η3, and η4, which are the weight coefficients for each item. In this example, η1, η2, η3, and η4 are taken as 0.22, 0.23, 0.35, and 0.2, respectively. Figure 5 Example of subtask division and planning for user request tasks;

[0319] Step S5 includes the following sub-steps:

[0320] Step S501: Embed the user's current dialogue q into the historical dialogue Θ for historical encoding, and construct a dialogue memory matrix set Z:

[0321] Z = {Z1, Z2, ..., Z} k}

[0322] Among them, Z k This represents the memory matrix for the k-th round of dialogue;

[0323] Step S502: Based on the KL divergence term, construct the dialogue reward function in multi-turn dialogues. The calculation formula is as follows:

[0324]

[0325] Where Q(s,a) and R(s,a) are the dialogue reward function and user feedback score for performing action a in dialogue state s, respectively; KL(p||q) is the KL divergence between the current policy p and the expected goal q; C(·) is the stability score of the current dialogue state; and v s Let m be the embedding vector of the dialogue state s, m be the vector dimension, λ1, λ2, λ3 and λ4 be the weight values ​​of each part, and T be the matrix transpose operation, which is 0.3, 0.3, 0.2 and 0.2 respectively in this embodiment.

[0326] Step S503: Optimize the dialogue strategy stably through a dual-delay update mechanism, while introducing an entropy regularization term to maintain strategy diversity. The specific steps of the dual-delay update strategy include:

[0327] First, for each state s and the next state s′, define the timing difference of the double-delay update mechanism:

[0328]

[0329] Where r represents the immediate reward for performing action a in state s, and ξ is the double-delay update parameter. The policy parameters of networks N1 and N2 are evaluated as a function to determine the value of the next state s′;

[0330] Then, for each state-action pair (s, a), the loss function of the double-delay update mechanism is calculated:

[0331]

[0332] Where B is the number of dialogues processed by the double-delay update mechanism, and in this example, the number of dialogues processed is B = 10;

[0333] Finally, after the loss function converges, the state-action pair (s′,a′) is updated with a delay.

[0334] Step S504: Write the user input, system response, and execution result of this round into the user profile;

[0335] Step S6 includes the following sub-steps:

[0336] Step S601: Maintain historical dialogue records and multi-turn interaction logs, record user expression habits and behavioral preferences, and perform statistical analysis to form user profiles;

[0337] Step S602: The system dynamically adjusts the response style based on the user's acceptance of different types of responses, optimizes the length, detail, and information organization of the responses, and ensures that each conversation meets the user's personalized needs, thereby improving the interactive experience.

[0338] Figure 6In this embodiment, the user profile features are defined based on the user background and preferences detected by the system: {

[0339] "Interaction Habits":

[0340] Average speaking speed: 4.2 words per second;

[0341] "Preferred Channel": Voice (73%) vs. Text (27%);

[0342] Information density: 8.3 information points per sentence;

[0343] "Business Records":

[0344] "Recent Services": ["Report Loss", "Balance Inquiry"];

[0345] Risk sensitivity: 0.68;

[0346] }

[0347] Specific embodiments have been used to illustrate the principles and implementation methods of this invention. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

[0348] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A method for intelligent business execution based on a dialogue mechanism, characterized in that: Includes the following steps: Step S1: Based on the multimodal data input by the user, construct a multimodal preprocessor and perform data preprocessing operations; Step S2: Utilize a joint reasoning mechanism based on multimodal intent to perform cross-modal intent recognition processing on the preprocessed multimodal request data; Step S3: Associate and match the intent data of the current user request with the user's historical request intent data, and optimize the system's response based on the historical information; Step S4: By constructing a task hypergraph, the user's current request is broken down and the execution paths between subtasks are planned; Step S5: Construct a historical multi-turn dialogue feedback optimization algorithm based on reinforcement learning and optimize the system's business processing capabilities based on user feedback; Step S6: Perform personalized optimization of the system by maintaining historical dialogue records and interaction logs.

2. The intelligent business execution method based on a dialogue mechanism according to claim 1, characterized in that: Step S1 specifically includes: Step S101: The system receives a multimodal service request initiated by the user through at least one of the following forms of expression: text, voice, and image; Step S102: For input text data, remove data noise from the user's input text request to improve semantic understanding accuracy; Step S103: For the input voice data, remove environmental noise from the user's voice input request to improve the accuracy of voice recognition. Step S104: For the input image data, remove noise from the user's input image request.

3. The intelligent business execution method based on a dialogue mechanism according to claim 2, characterized in that: In step S102, the specific operations for removing text request noise include: First, a semantic word segmentation and optimization algorithm is constructed to segment the initially denoised text data. The steps for constructing the semantic word segmentation and optimization algorithm are as follows: (1) Divide each word in the corpus into basic units, and iterate through all consecutive word pairs in the corpus. i ,w j The frequency of occurrence of word pairs is counted and denoted as: f(w) i ,w j ); (2) Select the most frequent word pairs (w i ,w j Merge them to generate new subwords w ij And set the merging threshold θ; (3) Repeat the above operation to traverse the remaining word pairs until the frequency of all word pairs is less than the threshold θ, and obtain a new word pair w; (4) By embedding a new sub-word context window N (w) The semantic weighted information is used to represent the word w as an n-dimensional vector v. The formula for vector embedding is: v=L(γ*H(E(w))+(1-γ)*∑ c∈N(w) a c *H(E(c))) ReLU6 = min(ReLU, 6) Where L(·) is the normalization function; H(·) is the ReLU6-based nonlinear activation function; E(w), E(c), and E(c') represent the embedding vectors of the target word, its context window word, and candidate word, respectively; γ is a scaling factor used to balance the initial embedding contribution; α c The attention weights of the target subword w for a specific context subword c;<E(w),E(c)> The inner product of two word vectors is used to measure similarity; <·> is the inner product operation; τ is a smoothing parameter used to adjust the smoothness of the attention distribution; the expression and subsequent exp(·) are exponential functions with the natural constant e as the base; ReLU(·) and ReLU6(·) are the ReLU and ReLU6 nonlinear activation functions, respectively. (5) Further optimize the word combination based on the Viterbi algorithm. The method for finding the optimal word combination is as follows: Where W = {w1, w2, ..., w T } represents the optimized sub-vocabulary, T represents the length of the sub-word sequence in the sub-vocabulary W, and w i P(w) represents the i-th word combination. i |w i-1 ) and P(q i |w i ) represent the transition probability and the corresponding observation unit q, respectively. i The probability of observation; (6) Calculate the semantic relevance of the optimal word combination and set a semantic relevance threshold to maximize the overall semantic similarity S of the sentence to obtain the final word combination. The formula for calculating the semantic relevance of sub-words is: s(w i ,w i+1 )=exp(β·<v i ,v i+1 >) Where s(·) is the semantic similarity function; v i and v i+1 They represent the subword w respectively i and w i+1 Vector representation; w i ∈W, where W represents the optimized sub-lexicon; β is the adjustment parameter; (7) For word combinations that are less than the semantic relevance threshold, the state weights are readjusted using a state weight adjustment algorithm to obtain the final word. The formula for calculating the state weight adjustment algorithm is as follows: e((in i ,In j ))=α*KL(P wi |P wj )+β*(1-cos(v i ,v j )) Among them, e((w i ,w j )) is used to retrieve subwords (w i ,w j The semantic error is a function, where δ is the step size parameter, α and β are the error parameters, and P is the error parameter. wi and P wj Sub-words (w) i ,w j The frequency of occurrence of ) KL(P) wi |P wj ) is P wi and P wj KL divergence between; v i and v j Sub-words (w) i ,w j Embedded vectors of ) Then, based on word segmentation, a chaotic entropy stop word filtering algorithm is constructed to adaptively perceive and analyze the lexical context. The specific steps for constructing the chaotic entropy stop word filtering algorithm include: (1) For each word w i Define a context sliding window of size k, forming a context set C. i : C i ={w i-k ,…,w i-1 ,w i+1 ,…,w i+k } (2) Calculate the probability p of each word appearing in the window by using the frequency of each word in the context. j ; in, C i Each word under the window In context set C i Frequency of appearance in; (3) Calculation words In context C i Information entropy in And set the word retention threshold τ; (4) Based on the calculated information entropy, words are dynamically filtered using the word retention function K(·) to obtain a filtered word sequence of length N: The formula for calculating the word retention function K(·) is: Finally, context-aware technology is used to correct the spelling of the text data, resulting in a preprocessed text request. The specific steps of context-aware technology include: First, the standard edit distance D(w,w′) between the current word w and the candidate word w′ is calculated as the basic cost for spell correction; Then, the vector of candidate word w′ in context C is calculated. i Average word semantic similarity S(w′,C) i The calculation formula is as follows: Wherein, cos(·) calculates the candidate word w′ and the 2k context words w′ before and after it. j A function representing the cosine similarity between them; Finally, a comprehensive scoring function Score(w,w′) is defined, and the optimal correction word w is determined based on the adjustable parameter α. * The optimal correction calculation method is as follows: Score(w,w′)=D(w,w′)-α*S(w′,C i )。 4. The intelligent business execution method based on a dialogue mechanism according to claim 2, characterized in that: In step S103, the specific steps for removing environmental noise from the voice request include: (1) Sample the original speech signal to obtain n discrete sample sequences {x0, x1, ..., x n }; (2) By using probability amplitude encoding, the normalized sample values ​​are mapped to quantum states |ψ>: (3) Perform a quantum Fourier transform on the obtained quantum state to obtain a new quantum state |ψ′>. The formula for calculating the quantum Fourier transform is: Where N represents the dimension of the quantum system, and k represents the summation index; (4) Statistically calculate the probability distribution p(|ψ′>) of the transformed quantum state |ψ′> at each time slice t, and calculate the information entropy H(t): H(t) = -∑p(f)ln p(f); p(f)=p(|ψ′(t)>→∣f>) Where p(f) represents the probability that the quantum state collapses to the ground state |f> at time t; |ψ′(t)> represents the quantum state at time t; and p(·) represents the probability function. (5) Based on the change of entropy with time or frequency band, define an adaptive threshold function τ(t) to determine high-noise sections. The formula for calculating the adaptive threshold function τ(t) is as follows: τ(t)=chx(1-H(t)) Where chx(·) is the hyperbolic cosine activation function with an adaptive threshold; (6) Nonlinear sound field modulation is used to eliminate reverberation in the high-noise section t. The calculation formula for nonlinear sound field modulation is: G(∣ψ′(t)>)=∣ψ′(t)>*exp(∣ψ′(t)>)*τ(t).

5. The intelligent business execution method based on a dialogue mechanism according to claim 2, characterized in that: The specific steps of step S104 include: (1) Perform L-level discrete wavelet decomposition on the original image I to obtain the low-frequency subband A. l and high-frequency subbands in horizontal, high vertical and high diagonal directions The formula for calculating discrete wavelet decomposition is: A l =ds(ds(I,G l (I)) T ,G l (I)) Where l∈{1,2,…,L}, ds(·) is the image downsampling operation, and G l (I) and G h (I) represents low-channel filtering and high-channel filtering operations on image I, respectively; (2) Evaluate each high-frequency subband D using local statistics l Noise level: Ω (x,y) ={Σ(i,j)|i∈[x-Δx,x+Δx],j∈[y-Δy,y+Δy]} Among them, D l (i,j) refers to the coefficient value with coordinates (i,j) in the l-th high-frequency sub-band after wavelet decomposition, μ l (x,y) represents an Ω within a window of size m×m centered at (x,y). (x,y) The local mean of the coordinate points, Δx and Δy are the coordinate offsets within the window, and m represents the side length of the noise statistics window; (3) Perform inverse wavelet reconstruction on all sub-bands to obtain the coarsely denoised image. Where O(·) represents the inverse reconstruction operation of discrete wavelet decomposition: O(A l )=us(us(A l ,G h (A l )) T ,G h (A l )) Where us(·) is the image upsampling operation; (4) Use bilateral filtering to further reduce noise in image I to obtain the final image I'. The formula for calculating bilateral filtering is: Where I(x) and I(y) are the pixel values ​​of the original image in the x and y directions, respectively, and ω c (x,y) is the spatial kernel of the image in the x and y directions, ω s (I(x)-I(y)) is the intensity kernel.

6. The intelligent business execution method based on a dialogue mechanism according to claim 1, characterized in that: Step S2 specifically includes: Step S201: Represent the preprocessed multimodal data, including but not limited to one or more combinations of text, speech, and image data, as a cross-modal semantic graph. Specific steps include: First, the processed user input information in the digital human business request scenario is modally expanded, with each word regarded as a node, and image frames and audio frames expanded into other modal nodes to form a comprehensive multimodal graph; Then, establish an edge r between adjacent words on the same modal edge to represent their semantic or syntactic relationship; Secondly, edges are established between adjacent modal edges that are the same or similar words to capture the evolution of context; Finally, the adjacency matrix M and degree matrix D are constructed to represent the connection relationships between nodes in the graph. The formula for constructing the adjacency matrix M is: Where, m (i,j) Let $D$ be the word co-occurrence weights for nodes $i$ and $j$. The degree matrix $D$ is a diagonal matrix, where each diagonal element represents the degree of the corresponding node. D (i,j) =∑ j M (i,j) Step S202: Construct a multimodal graph network hierarchically to obtain the complete cross-modal representation K. * : K * =K (1) +K (2) +…+K (l) Where l represents the number of multimodal data types.

7. The intelligent business execution method based on a dialogue mechanism according to claim 1, characterized in that: Step S3 specifically includes: Step S301: Query the user's previous N rounds of dialogue history, and use a pre-trained autoencoder to convert the historical dialogue and the current dialogue into fixed dimensions respectively; Step S302: Select the intermediate dialogue as the key dialogue, and calculate the similarity Sim between each round of dialogue and the key dialogue: Among them, d, μ represents the current dialogue and the intermediate dialogue, respectively. d , These represent the mean values ​​of the vector representations from the two rounds of dialogue. L1 and L2 represent the variances of the two rounds of dialogue vector representations, respectively; cov(·) is the covariance function; and L1 and L2 are stability constants. Step S303: Set an association threshold and identify dialogues with a similarity greater than the association threshold as associated dialogues, and introduce historical dialogue information; Step S304: Aggregate all related dialogues through a historical attention mechanism to generate the final answer.

8. The intelligent business execution method based on a dialogue mechanism according to claim 7, characterized in that: In step S304, the historical attention aggregation step includes: First, calculate the historical attention weights for the current dialogue and all n rounds of historical dialogues associated with it: Where, α i In the historical dialogue, d represents the i-th round of dialogue. i The importance weight, where q represents the current conversation; Then, the weighted sum of the historical dialogues is calculated to obtain the fused representation of the historical information Θ: Finally, the current dialogue information q and the historical information Θ are weighted and linearly fused to obtain the final answer A. * : A * =λ*q+(1-λ)*Θ Here, λ is a learnable weight coefficient that controls the fusion ratio of current dialogue and historical information.

9. The intelligent business execution method based on a dialogue mechanism according to claim 1, characterized in that: Step S4 specifically includes: Step S401: Perform semantic analysis on the user request that has undergone cross-modal processing and break it down into multiple sub-tasks; Step S402: Treat each subtask as a supernode o∈V, connect them with hyperedges based on the coupling dependencies between subtasks, and assign different dependency weights: Among them, dep(o i ,o j ) represents a subtask o i and o j The coupling dependency is σ(·), where σ is the Sigmoid function, β is the trainable dependency sensitivity coefficient, and b is the bias term. Step S403: Update the state of the hypergraph nodes using a dynamic update algorithm. The specific formula is as follows: in, Let S(o) represent the state of subtask o after the l-th round of subtask updates, and let S(o) represent the set of neighboring supernodes of subtask o. r and W o Let α represent the trainable parameters of the hyperedge and the hypernode, respectively. Let AG(·) be the hypergraph aggregation function. i Attention weights; Step S404: Calculate the feasibility score of the task and dynamically adjust the task schedule. The specific operations for adjusting the task schedule include: First, calculate the executability score of the task: Score(o)=shx(W·h o +b) Where S(o) represents the executability score of task o, reflecting whether the task is suitable for the current path; h o This indicates the state of subtask o after the task is updated; shx(·) are hyperbolic sine activation functions for executability scoring; W and b are learning parameters; Then, based on the dynamics of the execution costs between tasks, a greedy algorithm is used to search for the task path P with the highest executability, where the formula for calculating the optimal task execution path is: c(o i ,o j )=η1*(1-dep(o i ,o j ))+η2*res(o i ,o j )+η3*delay(o i ,o j )+η4*Score(o i ) Wherein, c(o) i ,o j ) indicates task o i and o j The execution costs between tasks are as follows: dep(·), res(·), and delay(·), which represent the execution costs of task o. i and o j The weights of the superedges, resource consumption, and real-time latency are defined by η1, η2, η3, and η4, which are the weight coefficients for each item.

10. The intelligent business execution method based on a dialogue mechanism according to claim 1, characterized in that: Step S5 specifically includes: Step S501: Embed the user's current dialogue q into the historical dialogue Θ for historical encoding, and construct a dialogue memory matrix set Z: Z={Z1,Z2,…,Z k } Among them, Z k This represents the memory matrix for the k-th round of dialogue; Step S502: Based on the KL divergence term, construct the dialogue reward function in multi-turn dialogues. The calculation formula is as follows: Where Q(s,a) and R(s,a) are the dialogue reward function and user feedback score for performing action a in dialogue state s, respectively; KL(p||q) is the KL divergence between the current policy p and the expected goal q; C(·) is the stability score of the current dialogue state; and v s Let m be the embedding vector of the dialogue state s, m be the vector dimension, λ1, λ2, λ3 and λ4 be the weight values ​​of each part, and T be the matrix transpose operation. Step S503: Optimize the dialogue strategy stably through a dual-delay update mechanism, while introducing an entropy regularization term to maintain strategy diversity. The specific steps of the dual-delay update strategy include: First, for each state s and the next state s′, define the timing difference of the double-delay update mechanism: Where r represents the immediate reward for performing action a in state s, and ξ is the double-delay update parameter. The policy parameters of networks N1 and N2 are evaluated as a function to determine the value of the next state s′; Then, for each state-action pair (s, a), the loss function of the double-delay update mechanism is calculated: Where B represents the batch of dialogues processed by the double-delay update mechanism; Finally, after the loss function converges, the state-action pair (s′,a′) is updated with a delay. Step S504: Write the user input, system response, and execution result of this round into the user profile.

Images