Enterprise-level intelligent customer service interaction control system for multi-turn dialogue scenarios

Abstract

This invention proposes an enterprise-level intelligent customer service interaction control system for multi-turn dialogue scenarios, including a multimodal semantic parsing module, a dynamic decision engine module, a closed-loop learning optimization module, and a multi-terminal secure delivery module. The system achieves semantic parsing of diverse texts through a hybrid neural network model, dynamically judges text sensitivity and selects appropriate retrieval paths, optimizes the performance of the hybrid neural network model in real time, and uses national cryptographic algorithms to ensure secure data transmission. It effectively solves the shortcomings of traditional customer service systems in understanding complex contexts, identifying sensitive information, and ensuring secure data transmission, significantly improving the system's accuracy, security, and adaptability. It is particularly suitable for enterprise-level customer service scenarios that require handling dialects, technical terms, and sensitive information, and has significant application value.

Description

Technical Field

[0001] This invention relates to the field of intelligent customer service system technology, specifically to an enterprise-level intelligent customer service interaction control system for multi-turn dialogue scenarios. Background Technology

[0002] Enterprise-level intelligent customer service interaction control systems are a widely used technology solution in the customer service process. They aim to improve the quality and efficiency of customer service through automation and intelligence. These systems typically rely on technologies such as natural language processing, machine learning, and deep learning. They can handle customer requests from different channels and provide personalized services. With the continuous development of technology, intelligent customer service systems have gradually acquired multi-turn dialogue processing capabilities and can respond dynamically based on the user's context.

[0003] However, existing enterprise-level intelligent customer service systems have significant shortcomings in several aspects. First, when handling complex multi-turn dialogues, existing systems struggle to fully understand and parse different types of text content, especially when dialects, industry terms, or proper nouns are involved, resulting in low recognition accuracy and inaccurate responses. Second, existing technologies have weak capabilities in judging and protecting sensitive information. Systems typically rely on a single data retrieval path and cannot dynamically adjust data processing strategies based on the sensitivity of the content, thus easily leading to the risk of data leakage. Finally, many traditional systems rely on public cloud platforms for data storage and processing, facing privacy, security, and compliance issues. Summary of the Invention

[0004] To address the technical problems mentioned in the background section, this invention effectively solves the problems of multi-turn dialogue understanding, sensitive information identification, and secure data transmission by introducing a hybrid neural network model, a dynamic decision engine, and a closed-loop learning optimization mechanism, thereby improving the system's accuracy, security, and adaptability.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] An enterprise-level intelligent customer service interaction control system for multi-turn dialogue scenarios includes:

[0007] M1, a multimodal semantic parsing module, deploys a hybrid neural network model consisting of a Transformer encoder and a BiLSTM network; it calculates scores for diverse text content through the hybrid neural network model, and further calculates the score confidence; it outputs structured semantic information including diverse text content and corresponding score confidence.

[0008] M2, Dynamic Decision Engine Output Module: This module determines whether the diverse text content is sensitive information based on the structured semantic information to select a retrieval path. If it is sensitive information, the retrieval path is the core database; otherwise, the retrieval path is the extended knowledge base. Based on the retrieval path, the core database or the extended knowledge base is selected for retrieval to generate retrieval results.

[0009] M3, a closed-loop learning optimization module, captures in real time the dialect recognition error, emotion response deviation, and terminology database missing alarm signals of the hybrid neural network model and inputs them into the incremental model; the incremental model is used to optimize the hybrid neural network model; the optimized hybrid neural network model is synchronized to each business terminal through a secure channel; the incremental model is established in the enterprise intranet data sandbox;

[0010] M4, the multi-terminal dialogue secure delivery module, generates and feeds back user response text through the optimized neural network model, integrates enterprise-level gateway and private API communication protocol, and uses national cryptographic algorithms to encrypt the channel, realizing end-to-end non-cloud data transmission.

[0011] Furthermore, the diverse text content includes natural language, dialect features, and industry terminology;

[0012] The diverse text content input by the user is transformed into word vector representations for each word through word embedding. ; This represents the word vector of the i-th word;

[0013] Then each of the word vectors is... Perform a linear transformation to extract the query vector. Key vector Sum value vector ;

[0014] The Transformer encoder optimizes the representation of each word vector by calculating attention weights between the word vectors through a self-attention mechanism. The formula for calculating the attention weights is as follows:

[0015]

[0016] in, This represents the self-attention mechanism; The dimension of the key vector; This represents the Softmax function;

[0017] The Transformer encoder generates an optimized word vector sequence through calculation. , ,in, It is the first in the text Optimized word vectors for each word.

[0018] Furthermore, the optimized word vector sequence H is passed to the BiLSTM network to further model the time-series dependencies in the word vector sequence. The calculation formula is as follows:

[0019]

[0020] in, It is the hidden state from the previous moment; It represents the hidden state at the current moment, indicating the contextual information for optimizing the word vector; It is the optimized word vector sequence at time t, i.e. From ; Represents the LSTM function;

[0021] Hidden state of the Transformer encoder output Scores corresponding to the natural language, dialect features, and industry terms are generated through linear transformation;

[0022] The formula for calculating the score of the natural language is as follows:

[0023]

[0024] in, Represents natural language score; This represents the weight matrix related to natural language. Represents bias terms related to natural language;

[0025] The scoring formula for the dialect features is as follows:

[0026]

[0027] in, Indicates the score for dialect features; Represents the weight matrix related to dialect features; Indicates bias terms related to dialect features;

[0028] The scoring formula for the industry terms is as follows:

[0029]

[0030] in, Indicates the score for industry terminology; This represents a weight matrix related to industry terminology; Indicates bias terms related to industry terminology.

[0031] Furthermore, the scores corresponding to natural language, dialect features, and industry terms are passed to the Softmax function to calculate the confidence score of each natural language, dialect feature, and industry term.

[0032] The confidence level of the natural language score The calculation formula is:

[0033] in, It is the normalization factor for all scores, ensuring that the final probability sum is 1;

[0034] The confidence level of the dialect feature score The calculation formula is: ;

[0035] Confidence level of the industry terminology score The calculation formula is: .

[0036] Furthermore, the confidence-weighted evaluation of the structured semantic information is specifically expanded using the following formula:

[0037]

[0038] in, It is the weighted natural language confidence score; Weighted confidence level of dialect features; It is a weighted confidence level for industry-specific terms; , , It is a context adjustment factor; It is a factor that normalizes all scores, ensuring that the final probability sum is 1;

[0039] Let the initial structured semantic information be denoted as I; let the weighted structured semantic information be denoted as... ,in This includes confidence levels for diverse text content and weighted natural language, dialect features, and industry terminology.

[0040] The search path is determined by judging whether diverse text content is sensitive information based on a weighted confidence value: First, the sensitivity threshold is set to 0.8; when the weighted confidence value is greater than or equal to the preset sensitivity threshold, it is marked as sensitive information, and the core database is selected as the search path; when the weighted confidence value is lower than the set sensitivity threshold, it is marked as non-sensitive information, and the extended knowledge base is selected as the search path.

[0041] Furthermore, the retrieval process of the core database involves inputting the sensitive information into the core database, encrypting the input sensitive information using an encryption retrieval function, and then retrieving the information in the database using indexing technology to obtain the retrieval results.

[0042] Furthermore, the dialect recognition error refers to the errors made by the hybrid neural network model when processing accents, vocabulary, or grammar from different regions, and is calculated using the following formula:

[0043]

[0044] in, This indicates dialect recognition error; This represents the predicted value of the hybrid neural network system for the i-th dialect sample; This represents the actual label value, i.e., the correct dialect recognition result; The number of dialect samples;

[0045] The sentiment response deviation value represents the difference between the prediction result of the hybrid neural network model and the actual user sentiment when performing sentiment analysis. The calculation formula is as follows:

[0046]

[0047] in, Indicates the deviation value of the emotional response; This indicates that the hybrid neural network model is for the first... Predicted sentiment values ​​for each sentiment sample; Indicates the actual sentiment value; The number of emotion samples;

[0048] The terminology database missing alarm signal indicates that a missing industry term has been detected during the processing of the hybrid neural network system. The calculation formula is as follows:

[0049]

[0050] in, Ter indicates a missing term alarm signal; Indicates the first A term; Indicates the current terminology database; 1 is an indicator function, when the term... Not in the terminology database If the condition is met, return 1; otherwise, return 0. For the number of terms.

[0051] Furthermore, the training steps of the incremental model include training data preprocessing, incremental model initialization, incremental training loop, real-time signal monitoring, and security auditing and rollback;

[0052] The incremental model initialization uses the hybrid neural network as the basic architecture for incremental training, and sets the weights and bias parameters using the Xavier initialization method. The calculation formula is as follows:

[0053]

[0054] in, These are weight parameters, representing the weights from the first... layer to the first Layer weights; This indicates that random values ​​are drawn uniformly from the interval; It is the number of neurons in the input layer; It is the number of neurons in the output layer;

[0055] The incremental training loop includes forward propagation, loss calculation, back propagation, and gradient calculation.

[0056] The real-time signal monitoring evaluates the accuracy of the incremental model by monitoring the dialect recognition error, emotional response deviation, and terminology database missing alarm signals of the hybrid neural network model, and continues the incremental training loop when the accuracy is lower than a set first threshold.

[0057] The security audit and rollback mechanism is activated when the accuracy of the incremental model falls below the set second threshold, restoring the incremental model to its state before the last incremental training cycle.

[0058] Furthermore, the forward propagation refers to the training process using an incremental model to perform forward propagation and calculate the prediction result. The calculation formula for the forward propagation is as follows:

[0059]

[0060] in, This is the predicted output of the incremental model; It is a word vector; These are the parameters of the incremental model;

[0061] The loss calculation uses a loss function to calculate the difference between the incremental model's predicted value and the actual label, and employs the cross-entropy loss function. The specific calculation formula is as follows:

[0062]

[0063] in, It is a loss function; , , These are the weighting coefficients; This represents dialect recognition error; This represents the deviation value of the emotional response; A warning signal indicating a missing terminology database is displayed.

[0064] The backpropagation and gradient calculation involve calculating the gradient of the loss function with respect to the incremental model parameters through backpropagation, and then updating the incremental model parameters using gradient descent. The formula for the update process is as follows:

[0065]

[0066] in, The parameters represent the updated incremental model; This represents the incremental model parameters before the update; It is the learning rate, which controls the step size for updating parameters; It is the gradient of the loss function with respect to the parameters of the incremental model.

[0067] Compared with the prior art, the advantages of the present invention are as follows:

[0068] 1. This invention employs a hybrid neural network model, combining a Transformer encoder and a BiLSTM network, to accurately process diverse text content. The Transformer encoder optimizes the semantic representation of text through a self-attention mechanism, while the BiLSTM network effectively captures dependencies in time series, ensuring that the system can accurately understand the context of each turn in multi-turn dialogues. Through this technological innovation, the accuracy of the system in handling different contexts and complex dialogues is significantly improved, making the interaction between the customer and the system smoother and more precise.

[0069] 2. This invention utilizes a dynamic decision engine to intelligently determine whether text contains sensitive information based on its confidence level, and selects an appropriate retrieval path accordingly. Specifically, the system performs a weighted evaluation based on the confidence values ​​of diverse text content. When the confidence value exceeds a set threshold, the sensitive information is encrypted and retrieved from the core database. When the confidence value is below the threshold, the system retrieves from an expanded knowledge base. This innovative mechanism ensures the intelligent identification and secure processing of sensitive information, avoiding the data leakage risks associated with relying on a single retrieval path in traditional systems. Furthermore, it enhances data security through non-cloud data transmission and a private API communication protocol.

[0070] 3. By introducing a closed-loop learning optimization module, this invention can capture and correct problems such as dialect recognition errors, emotional response deviations, and missing terminology databases in real time. When the system identifies errors in dialects or industry terms, it inputs the error signal into the incremental model for optimization and improves the accuracy of the hybrid neural network model through incremental learning. This mechanism enables the system to continuously adapt to new language environments and business scenarios, improve its accuracy and responsiveness in long-term operation, and ensure that it can provide high-quality customer service in different contexts. Attached Figure Description

[0071] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0072] Figure 1 This is a schematic diagram of the system workflow of the present invention;

[0073] Figure 2 This is a schematic diagram of the workflow of the multimodal semantic parsing module of the present invention;

[0074] Figure 3 This is a schematic diagram of the incremental model training process of the present invention. Detailed Implementation

[0075] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0076] To achieve the above objectives, the present invention provides an enterprise-level intelligent customer service interaction control system for multi-turn dialogue scenarios, such as... Figures 1-3 As shown, the system includes:

[0077] M1, a multimodal semantic parsing module, deploys a hybrid neural network model consisting of a Transformer encoder and a BiLSTM network; it calculates scores for diverse text content through the hybrid neural network model, and further calculates the confidence level of the scores; it outputs structured semantic information including diverse text content and corresponding score confidence levels.

[0078] The diverse text content includes natural language, dialect features, and industry terminology;

[0079] 1) The diverse text content input by the user is transformed into word vector representations for each word through word embedding. ; This represents the word vector of the i-th word;

[0080] Then each of the word vectors is... Perform a linear transformation to extract the query vector. Key vector Sum value vector

[0081] 2) The Transformer encoder optimizes the representation of each word vector by calculating the attention weights between the word vectors through a self-attention mechanism. The formula for calculating the attention weights is as follows:

[0082]

[0083] in, This represents the self-attention mechanism; The dimension of the key vector; This represents the Softmax function;

[0084] The Transformer encoder generates an optimized word vector sequence through calculation. , ,in, It is the first in the text Optimize word vectors for each word;

[0085] 3) The optimized word vector sequence H is passed to the BiLSTM network to further model the time series dependencies in the word vector sequence. The calculation formula is as follows:

[0086]

[0087] in, It is the hidden state from the previous moment; It represents the hidden state at the current moment, indicating the contextual information for optimizing the word vector; It is the optimized word vector sequence at time t, i.e. From ; Represents the LSTM function;

[0088] 4) Hidden states of the Transformer encoder output Scores corresponding to the natural language, dialect features, and industry terms are generated through linear transformation;

[0089] The formula for calculating the score of the natural language is as follows:

[0090]

[0091] in, Represents natural language score; This represents the weight matrix related to natural language. This represents bias terms related to natural language, used to adjust the natural language score;

[0092] The scoring formula for the dialect features is as follows:

[0093]

[0094] in, Indicates the score for dialect features; Represents the weight matrix related to dialect features; A bias term related to dialect features is used to adjust the dialect feature scores;

[0095] The scoring formula for the industry terms is as follows:

[0096]

[0097] in, Indicates the score for industry terminology; This represents a weight matrix related to industry terminology; This indicates bias terms related to industry terms, used to adjust the scores for those industry terms;

[0098] 5) Pass the scores corresponding to natural language, dialect features, and industry terms to the Softmax function to calculate the confidence score of each natural language, dialect feature, and industry term.

[0099] The confidence level of the natural language score The calculation formula is:

[0100] in, It is the normalization factor for all scores, ensuring that the final probability sum is 1;

[0101] The confidence level of the dialect feature score The calculation formula is: ;

[0102] Confidence level of the industry terminology score The calculation formula is: .

[0103] M2, Dynamic Decision Engine Output Module: This module determines whether the diverse text content is sensitive information based on the structured semantic information to select a retrieval path. If it is sensitive information, the retrieval path is the core database; otherwise, the retrieval path is the extended knowledge base. Based on the retrieval path, the core database or the extended knowledge base is selected for retrieval to generate retrieval results.

[0104] 1) A weighted evaluation is performed on the confidence level within the structured semantic information. Specifically, the evaluation formula for the confidence level is extended as follows:

[0105]

[0106] in, It is the weighted natural language confidence score; Weighted confidence level of dialect features; It is a weighted confidence level for industry-specific terms; , , It is a context adjustment factor; It is a factor that normalizes all scores, ensuring that the final probability sum is 1;

[0107] Let the initial structured semantic information be denoted as I; let the weighted structured semantic information be denoted as... ,in This includes confidence levels for diverse text content and weighted natural language, dialect features, and industry terminology.

[0108] 2) Determine whether the diverse text content is sensitive information based on the weighted confidence value to determine the retrieval path;

[0109] Set the sensitivity threshold to 0.8;

[0110] When the weighted confidence value is greater than or equal to the preset sensitivity threshold, it indicates that the text content contains a lot of sensitive information, is marked as sensitive information, and the core database is selected as the retrieval path.

[0111] When the weighted confidence value is lower than the set sensitivity threshold, it indicates that there is little sensitive information in the text content, and it is marked as non-sensitive information. The extended knowledge base is then selected as the retrieval path.

[0112] The retrieval process of the core database involves inputting the sensitive information into the core database, encrypting the input sensitive information using an encryption retrieval function, and then retrieving the information in the database using indexing technology to obtain the retrieval results.

[0113] The retrieval process of the extended knowledge base involves inputting the non-sensitive information into the extended knowledge base, retrieving it from the database using indexing technology, and obtaining the retrieval results.

[0114] M3, a closed-loop learning optimization module, captures in real time the dialect recognition error, emotion response deviation, and terminology database missing alarm signals of the hybrid neural network model and inputs them into the incremental model; the incremental model is used to optimize the hybrid neural network model; the optimized hybrid neural network model is synchronized to each business terminal through a secure channel; the incremental model is established in the enterprise intranet data sandbox;

[0115] The dialect recognition error refers to the errors that occur when the hybrid neural network model processes accents, vocabulary, or grammar from different regions, and the calculation formula is as follows:

[0116]

[0117] in, This indicates dialect recognition error; This represents the predicted value of the hybrid neural network system for the i-th dialect sample; This represents the actual label value, i.e., the correct dialect recognition result; The number of dialect samples;

[0118] The sentiment response deviation value represents the difference between the prediction result of the hybrid neural network model and the actual user sentiment when performing sentiment analysis. The calculation formula is as follows:

[0119]

[0120] in, Indicates the deviation value of the emotional response; This indicates that the hybrid neural network model is for the first... Predicted sentiment values ​​for each sentiment sample; Indicates the actual sentiment value; The number of emotion samples;

[0121] The terminology database missing alarm signal indicates that a missing industry term has been detected during the processing of the hybrid neural network system. The calculation formula is as follows:

[0122]

[0123] in, Ter indicates a missing term alarm signal; Indicates the first A term; This represents the current terminology database (containing all known industry terms); 1 is an indicator function, indicating when the term... Not in the terminology database If the condition is met, return 1; otherwise, return 0. For the number of terms;

[0124] Based on the calculation of dialect recognition error, emotional response deviation, and terminology database missing alarm signals, a security sandbox is further introduced to ensure that the incremental model can handle and store data more securely during the training phase. The security sandbox environment includes hardware configuration and data isolation.

[0125] The hardware configuration refers to an independent server cluster deployed on the enterprise intranet, equipped with a national cryptographic level 2 authentication encryption card;

[0126] The data isolation mentioned above employs dual-channel memory isolation technology to achieve physical memory partitioning, expressed by the following formula:

[0127]

[0128] in, This represents the data matrix ultimately used to train the incremental model; This refers to the original training data matrix within the security sandbox; The mask matrix is ​​dynamically generated and destroyed after each training session.

[0129] In a sandbox environment that ensures data security, the incremental model enters its training phase. The training steps of the incremental model include training data preprocessing, incremental model initialization, incremental training loop, real-time signal monitoring, and security auditing and rollback.

[0130] The input x of the incremental model includes the structured semantic information output in M1; the structured semantic information weighted by the confidence value and the retrieval path output in M2; the specific retrieval results output in M3; and the dialect recognition error, emotion response deviation value, and term database missing alarm signal of the hybrid neural network model captured in M4.

[0131] 1) The training data preprocessing refers to the process by which the incremental model collects the input data and processes it through cleaning, denoising, and standardization steps to ensure that the data can meet the needs of subsequent incremental model training.

[0132] 2) The incremental model initialization uses the hybrid neural network as the basic architecture for incremental training, and sets the weights and bias parameters using the Xavier initialization method. The calculation formula is as follows:

[0133]

[0134] in, These are weight parameters, representing the weights from the first... layer to the first Layer weights; This indicates that random values ​​are drawn uniformly from the interval; It is the number of neurons in the input layer; It is the number of neurons in the output layer;

[0135] 3) The incremental training loop includes forward propagation, loss calculation, backpropagation, and gradient calculation;

[0136] ① The forward propagation mentioned above refers to the training process using an incremental model to perform forward propagation and calculate the prediction result. The calculation formula for the forward propagation is as follows:

[0137]

[0138] in, This is the predicted output of the incremental model; It is the input data; These are the parameters (weights and biases) of the incremental model;

[0139] ② The loss calculation uses a loss function to calculate the difference between the incremental model's predicted value and the actual label. In this embodiment, the cross-entropy loss function is used for calculation, and the specific calculation formula is as follows:

[0140]

[0141] in, It is a loss function; , , These are the weighting coefficients; This represents dialect recognition error; This represents the deviation value of the emotional response; A warning signal indicating a missing terminology database is displayed.

[0142] ③ The backpropagation and gradient calculation involve calculating the gradient of the loss function with respect to the incremental model parameters through backpropagation, and then updating the incremental model parameters using gradient descent. The formula for the update process is as follows:

[0143]

[0144] in, The parameters represent the updated incremental model; This represents the incremental model parameters before the update; It is the learning rate, which controls the step size for updating parameters; It is the gradient of the loss function with respect to the parameters of the incremental model;

[0145] 4) The real-time signal monitoring evaluates the accuracy of the incremental model by monitoring the dialect recognition error, emotion response deviation, and terminology database missing alarm signal of the hybrid neural network model, and continues the incremental training loop when the accuracy is lower than the set first threshold of 0.7.

[0146] 5) The security audit and rollback mechanism is activated when the accuracy of the incremental model is lower than the second threshold of 0.5, restoring the incremental model to its state before the last incremental training cycle.

[0147] The optimized hybrid neural network model refers to the hybrid neural network model trained incrementally and synchronized to various business terminals through a national cryptographic algorithm encrypted channel.

[0148] M4, the multi-terminal dialogue secure delivery module, generates and feeds back user response text through the optimized neural network model, integrates enterprise-level gateway and private API communication protocol, and uses national cryptographic algorithms to encrypt the channel, realizing end-to-end non-cloud data transmission.

[0149] The hybrid neural network model outputs an intermediate vector result of the response intent and converts it into natural language response text through word mapping, which is then fed back to the user.

[0150] The proprietary API communication protocol ensures that all data transmission does not rely on external public cloud services, but is completed within the enterprise's secure environment, providing encryption protection for communication, ensuring the confidentiality and integrity of data transmission, and preventing data from being leaked or tampered with during transmission.

[0151] The text data is transmitted in encrypted form through a national cryptographic algorithm encryption channel. The national cryptographic algorithms (such as SM2, SM3 and SM4) are encryption standards issued by the State Cryptography Administration of China to ensure that the text data in transmission is fully encrypted and protected, and to ensure that the text data is not accessed or tampered with by unauthorized parties during transmission, in accordance with national information security standards.

[0152] The aforementioned end-to-end non-cloud data transmission means that data is not transmitted through public cloud services, but is processed and transmitted through the enterprise's own network and infrastructure. This design reduces dependence on external cloud platforms, avoids potential data security risks, and ensures that the enterprise has complete control over its sensitive data.

[0153] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0154] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An enterprise-level intelligent customer service interaction control system for multi-turn dialogue scenarios, characterized in that: include: M1, a multimodal semantic parsing module, deploys a hybrid neural network model consisting of a Transformer encoder and a BiLSTM network; it calculates scores for diverse text content through the hybrid neural network model, and further calculates the score confidence; it outputs structured semantic information including diverse text content and corresponding score confidence. M2, Dynamic Decision Engine Output Module: This module determines whether the diverse text content is sensitive information based on the structured semantic information to select a retrieval path. If it is sensitive information, the retrieval path is the core database; otherwise, the retrieval path is the extended knowledge base. Based on the retrieval path, the core database or the extended knowledge base is selected for retrieval to generate retrieval results. M3, a closed-loop learning optimization module, captures in real time the dialect recognition error, emotion response deviation, and terminology database missing alarm signals of the hybrid neural network model and inputs them into the incremental model; the incremental model is used to optimize the hybrid neural network model; the optimized hybrid neural network model is synchronized to each business terminal through a secure channel; the incremental model is established in the enterprise intranet data sandbox; M4, the multi-terminal dialogue secure delivery module, generates and feeds back user response text through the optimized neural network model, integrates enterprise-level gateway and private API communication protocol, and uses national cryptographic algorithms to encrypt the channel, realizing end-to-end non-cloud data transmission.

2. The enterprise-level intelligent customer service interaction control system for multi-turn dialogue scenarios according to claim 1, characterized in that, The diverse text content includes natural language, dialect features, and industry terminology; The diverse text content input by the user is transformed into word vector representations for each word through word embedding. ; This represents the word vector of the i-th word; Then each of the word vectors is... Perform a linear transformation to extract the query vector. Key vector Sum value vector ; The Transformer encoder optimizes the representation of each word vector by calculating attention weights between the word vectors through a self-attention mechanism. The formula for calculating the attention weights is as follows: in, This represents the self-attention mechanism; The dimension of the key vector; Represents the Softmax function; The Transformer encoder generates an optimized word vector sequence through calculation. , ,in, It is the first in the text Optimized word vectors for each word.

3. The enterprise-level intelligent customer service interaction control system for multi-turn dialogue scenarios according to claim 2, characterized in that, The optimized word vector sequence H is passed to the BiLSTM network to further model the time series dependency in the word vector sequence. The calculation formula is as follows: in, It is the hidden state from the previous moment; It represents the hidden state at the current moment, indicating the contextual information for optimizing the word vector; It is the optimized word vector sequence at time t, i.e. From ; Represents the LSTM function; Hidden state of the Transformer encoder output Scores corresponding to the natural language, dialect features, and industry terms are generated through linear transformation; The formula for calculating the score of the natural language is as follows: in, Represents natural language score; This represents the weight matrix related to natural language. Represents bias terms related to natural language; The formula for calculating the score of the dialect feature is as follows: in, Indicates the score for dialect features; Represents the weight matrix related to dialect features; Indicates bias terms related to dialect features; The scoring formula for the industry terms is as follows: in, Indicates the score for industry terminology; This represents a weight matrix related to industry terminology; Indicates bias terms related to industry terminology.

4. The enterprise-level intelligent customer service interaction control system for multi-turn dialogue scenarios according to claim 3, characterized in that, The scores corresponding to natural language, dialect features, and industry terms are passed to the Softmax function to calculate the confidence score of each natural language, dialect feature, and industry term. The confidence level of the natural language score The calculation formula is: in, It is the normalization factor for all scores, ensuring that the final probability sum is 1; The confidence level of the dialect feature score The calculation formula is: ; Confidence level of the industry terminology score The calculation formula is: .

5. The enterprise-level intelligent customer service interaction control system for multi-turn dialogue scenarios according to claim 1, characterized in that, The confidence-weighted evaluation of the structured semantic information is specifically expanded into the following formula: in, It is the weighted natural language confidence score; Weighted confidence level of dialect features; It is a weighted confidence level for industry-specific terms; , , It is a context adjustment factor; It is a factor that normalizes all scores, ensuring that the final probability sum is 1; Let the initial structured semantic information be denoted as I; let the weighted structured semantic information be denoted as... ,in This includes confidence levels for diverse text content and weighted natural language, dialect features, and industry terminology. The search path is determined by judging whether diverse text content is sensitive information based on a weighted confidence value: First, the sensitivity threshold is set to 0.8; when the weighted confidence value is greater than or equal to the preset sensitivity threshold, it is marked as sensitive information, and the core database is selected as the search path; when the weighted confidence value is lower than the set sensitivity threshold, it is marked as non-sensitive information, and the extended knowledge base is selected as the search path.

6. The enterprise-level intelligent customer service interaction control system for multi-turn dialogue scenarios according to claim 1, characterized in that, The retrieval process of the core database involves inputting the sensitive information into the core database, encrypting the input sensitive information using an encryption retrieval function, and then retrieving the information from the database using indexing technology to obtain the retrieval results.

7. The enterprise-level intelligent customer service interaction control system for multi-turn dialogue scenarios according to claim 1, characterized in that, The dialect recognition error refers to the errors that occur when the hybrid neural network model processes accents, vocabulary, or grammar from different regions, and the calculation formula is as follows: in, This indicates dialect recognition error; This represents the prediction value of the hybrid neural network system for the i-th dialect sample; This represents the actual label value, i.e., the correct dialect recognition result; The number of dialect samples; The sentiment response deviation value represents the difference between the prediction result of the hybrid neural network model and the actual user sentiment when performing sentiment analysis. The calculation formula is as follows: in, Indicates the deviation value of the emotional response; This indicates that the hybrid neural network model is for the first... Predicted sentiment values ​​for each sentiment sample; Indicates the actual sentiment value; The number of emotion samples; The terminology database missing alarm signal indicates that a missing industry term has been detected during the processing of the hybrid neural network system. The calculation formula is as follows: in, Ter indicates a missing term alarm signal; Indicates the first A term; Indicates the current terminology database; 1 is an indicator function, when the term... Not in the terminology database If the condition is met, return 1; otherwise, return 0. For the number of terms.

8. The enterprise-level intelligent customer service interaction control system for multi-turn dialogue scenarios according to claim 1, characterized in that, The training steps of the incremental model include training data preprocessing, incremental model initialization, incremental training loop, real-time signal monitoring, and security auditing and rollback. The incremental model initialization uses the hybrid neural network as the basic architecture for incremental training, and sets the weights and bias parameters using the Xavier initialization method. The calculation formula is as follows: in, These are weight parameters, representing the weights from the first... layer to the first Layer weights; This indicates that random values ​​are drawn uniformly from the interval; It is the number of neurons in the input layer; It is the number of neurons in the output layer; The incremental training loop includes forward propagation, loss calculation, back propagation, and gradient calculation. The real-time signal monitoring evaluates the accuracy of the incremental model by monitoring the dialect recognition error, emotional response deviation, and terminology database missing alarm signals of the hybrid neural network model, and continues the incremental training loop when the accuracy is lower than a set first threshold. The security audit and rollback mechanism is activated when the accuracy of the incremental model falls below the set second threshold, restoring the incremental model to its state before the last incremental training cycle.

9. The enterprise-level intelligent customer service interaction control system for multi-turn dialogue scenarios according to claim 8, characterized in that, The forward propagation refers to the training process using an incremental model to perform forward propagation and calculate the prediction results. The calculation formula for the forward propagation is as follows: in, It is the predicted output of the incremental model; It is a word vector; These are the parameters of the incremental model; The loss calculation uses a loss function to calculate the difference between the incremental model's predicted value and the actual label, and employs the cross-entropy loss function. The specific calculation formula is as follows: in, It is a loss function; , , These are the weighting coefficients; This represents dialect recognition error; This represents the deviation value of the emotional response; A warning signal indicating a missing terminology database is displayed. The backpropagation and gradient calculation involve calculating the gradient of the loss function with respect to the incremental model parameters through backpropagation, and then updating the incremental model parameters using gradient descent. The formula for the update process is as follows: in, The parameters represent the updated incremental model; This represents the incremental model parameters before the update; It is the learning rate, which controls the step size for updating parameters; It is the gradient of the loss function with respect to the parameters of the incremental model.

Images