Intelligent voice dialogue scene script intervention method and system based on customer portrait

Abstract

The application provides a customer portrait-based intelligent voice dialogue scene language intervention method and system, which comprises: scene intention recognition, entity recognition and multi-round dialogue; emotion recognition based on voice and semantics; definition and recognition of potential information label system; through an interface connection mode, portrait information of a customer is obtained based on unique customer information and is stored and recorded; based on the recognized scene intention of the customer, the emotional information of the customer, the potential information label of the customer and the portrait information of the customer, a plurality of applicable experience language and recommendation values are generated through a preset language generation model, and an experience language is randomly returned each time the customer interacts with the intelligent voice; and the reply language is controlled through intervention range configuration and intervention mode configuration. The application can find or generate the most matched language from the constructed language library while processing the reply of the scene intention of the customer, and improves the interactive experience of the intelligent voice dialogue as a whole.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to a method and system for intervening in intelligent voice dialogue scenarios based on customer profiles. Background Technology

[0002] Intelligent dialogue relies on semantic understanding and processing and is currently applied across various industries. Human-computer dialogue reduces costs, improves communication efficiency, and enhances user experience. In the financial and banking sector, it has permeated multiple scenarios, such as intelligent voice, intelligent outbound calling, SMS notifications, and agent assistance. Among these, intelligent dialogue applications, particularly intelligent voice providing 24 / 7 service, have already seen significant development and widespread adoption in the financial and banking field.

[0003] However, the monotonous business response methods can no longer meet users' personalized experience needs. More and more users hope that when receiving customer service, they can be matched with the most suitable service according to their own situation and behavior, including product recommendations, service guidance, and service perception.

[0004] Patent document CN111797210A (application number: CN202010138636.8) discloses an information recommendation method based on user profiles, including the following steps: receiving human-computer dialogue chat records in a first scenario in real time through the log collection system Flume; de-identifying the chat records to obtain first data; removing stop words from the first data to obtain second data; extracting keywords from the second data using a preset word map; obtaining a first tag set based on the keywords; deduplicating the first tag set data to obtain a second tag set; generating a user interest profile based on the second tag set and storing the user interest profile in a database; receiving a recommendation instruction and obtaining the user interest profile based on the recommendation instruction; obtaining the information to be recommended corresponding to the recommendation instruction based on the user interest profile.

[0005] Currently, intelligent voice solutions for call centers primarily rely on intelligent dialogue platforms to provide chatbot services. Typically, task-oriented dialogue engines connect to business systems to acquire user profiles, transaction information, and other data. This data is then simply categorized into different dimensions, and different scripts are configured for each dimension. While this approach can meet user needs to some extent, it has several drawbacks. First, there's the issue of granularity. Simple maintenance and categorization gradually reduce user experience after a period of use, failing to achieve a personalized experience or allow for rapid adjustments and applications based on accumulated user behavior over time. Second, there's the issue of dialogue delivery scope. The solution mainly utilizes task-oriented dialogue to support business system integration for script optimization, lacking optimization capabilities for other dialogue engines such as FAQ single-turn knowledge-based questions and structured data question answering. Third, there's the issue of configuration and maintenance costs. To achieve script optimization across multiple dialogue engines, in addition to system support, operations personnel need to update and configure the scripts on the corresponding engines. Similarly, to continuously meet the needs of business changes and user experience, operations personnel must constantly adjust the configurations across multiple dialogue engines. Summary of the Invention

[0006] To address the shortcomings of existing technologies, the purpose of this invention is to provide a method and system for intelligent voice dialogue scenario intervention based on customer profiles.

[0007] The intelligent voice dialogue scenario intervention method based on customer profile provided by the present invention includes:

[0008] Step 1: Perform scene intent recognition, entity recognition, and multi-turn dialogue;

[0009] Step 2: Perform emotion recognition based on speech and semantics;

[0010] Step 3: Define and identify potential information labeling systems;

[0011] Step 4: Through interface integration, obtain customer profile information based on the customer's unique information and store and record it;

[0012] Step 5: Based on the customer's scenario intent, emotional information, potential information tags, and profile information, generate multiple applicable experience scripts and recommendation values ​​through a preset script generation model, and randomly return an experience script each time the customer interacts with the intelligent voice.

[0013] Step 6: Control the response script by configuring the scope and method of intervention.

[0014] Preferably, based on traditional machine learning models, including Naive Bayes models, support vector machines, or neural network models, including convolutional neural networks and recurrent neural networks, the labeled training data is classified and learned. During recognition, the text extracted from the user's newly input voice information is mapped to the intent through classification, and the corresponding response is returned based on the dialogue flow.

[0015] Preferably, a multimodal deep learning model integrating speech and semantics is used to classify and recognize the user's speech and text to obtain the user's emotion category, and the emotional features of the customer's speech information in the intelligent voice scenario are extracted through speech signal acquisition and semantic understanding.

[0016] Preferably, corresponding tags are created for potential information beyond the customer's intention during dialogue interactions in different intelligent voice scenarios;

[0017] The text generation model is based on an end-to-end text generation model to obtain response scripts. The text generation model is based on an encoder-decoder framework. First, the encoder encodes the input information through LSTM and Transformer neural networks. In the generation stage, the decoder combines the information from the encoder and the already generated text information to generate the next sentence.

[0018] Preferably, the scope of intervention: specifies the knowledge scope of the intervention, including task-based conversations, FAQs, casual conversations, and third-party engines;

[0019] Intervention conditions: Set conditions for nodes, with options including: Attitude: affirmative, negative, no attitude; Emotion: positive, negative, neutral; Profile tags: built-in city, gender, age, and occupation type, with support for custom tags via customer interface;

[0020] Intervention actions: Supports adding sentence prefixes and suffixes and outputting fixed responses;

[0021] Priority can be adjusted among multiple rules according to their order. If multiple rules are enabled, condition judgments will be performed sequentially. Rules can be moved and dragged to meet the specific needs of different businesses and scenarios.

[0022] The intelligent voice dialogue scenario intervention system based on customer profile provided by the present invention includes:

[0023] Scene Intent Dialogue Module: Performs scene intent recognition, entity recognition, and multi-turn dialogue;

[0024] Multimodal emotion recognition module: performs emotion recognition based on speech and semantics;

[0025] Potential information tagging module: Defines and identifies potential information tagging systems;

[0026] Customer profiling integration module: Through interface integration, it obtains and stores customer profile information based on unique customer information.

[0027] Experience script generation module: Based on recognizing the customer's scenario intent, emotional information, potential information tags, and customer profile information, it generates multiple applicable experience scripts and recommendation values ​​through a preset script generation model, and randomly returns an experience script each time the customer interacts with the intelligent voice.

[0028] The script intervention configuration module allows you to control the response script by configuring the scope and method of intervention.

[0029] Preferably, the scenario intent dialogue module includes: classifying and learning labeled training data based on traditional machine learning models, including Naive Bayes models, support vector machines, or neural network models, including convolutional neural networks and recurrent neural networks; mapping the extracted text of the user's newly input voice information to intents through classification during recognition; and returning the corresponding response script based on the dialogue flow.

[0030] Preferably, the multimodal emotion recognition module includes: classifying and recognizing the user's speech and text through a multimodal deep learning model that integrates speech and semantics to obtain the user's emotion category; and extracting the emotional features of the customer's speech information in the intelligent voice scenario through speech signal acquisition and semantic understanding.

[0031] Preferably, the potential information tagging module includes: creating corresponding tags for potential information beyond the customer's intention during dialogue interaction in different intelligent voice scenarios;

[0032] The experience dialogue generation module includes: obtaining response dialogue based on an end-to-end text generation model. The text generation model is based on an encoder-decoder framework. First, the encoder encodes the input information through LSTM and Transformer neural networks. In the generation stage, the decoder integrates the information from the encoder and the already generated text information to generate the next sentence.

[0033] Preferably, the script intervention configuration module includes:

[0034] Intervention Scope: Specify the knowledge scope of the intervention, including task-based conversations, FAQs, casual conversations, and third-party engines;

[0035] Intervention conditions: Set conditions for nodes, with options including: Attitude: affirmative, negative, no attitude; Emotion: positive, negative, neutral; Profile tags: built-in city, gender, age, and occupation type, with support for custom tags via customer interface;

[0036] Intervention actions: Supports adding sentence prefixes and suffixes and outputting fixed responses;

[0037] Priority can be adjusted among multiple rules according to their order. If multiple rules are enabled, condition judgments will be performed sequentially. Rules can be moved and dragged to meet the specific needs of different businesses and scenarios.

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] (1) This invention further digitizes and structures the existing conversation data in the industry, and provides a corresponding tag system to achieve more refined customer classification;

[0040] (2) This invention utilizes NLP technology to improve the user experience of bank customers using telephone banking to conduct business, realizes personalized interactive services for each customer, and maximizes customer conversion rate and customer satisfaction.

[0041] (3) This invention supports model-generated dialogue and rule configuration. Based on these two intervention strategies, the mode can better balance the effectiveness and controllability of intervention dialogue; it supports intervention strategy configuration for the entire customer service scenario, centralized and unified management, effectively reducing the difficulty of global strategy management and improving operational efficiency.

[0042] (4) This invention provides built-in data based on financial industry scenarios, and provides the ability to understand and identify scenario intent, emotional information, and potential labels on the basis of general scenarios in the financial industry, as well as corresponding intervention strategies, which can significantly reduce the configuration cost in the startup phase.

[0043] (5) After receiving the customer's voice information in the intelligent voice scenario, the present invention identifies the customer's scenario intent, the customer's emotional information, and the customer's potential information tags. At the same time, it combines the customer profile for comprehensive analysis. While processing the customer's scenario intent response, it finds or generates the most matching script from the constructed script library, thereby improving the overall interactive experience of intelligent voice. Attached Figure Description

[0044] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0045] Figure 1 This is a basic business process diagram of the present invention;

[0046] Figure 2 This is a system architecture diagram of the present invention;

[0047] Figure 3 This is a diagram illustrating a business scenario.

[0048] Figure 4Flowchart for training the intent recognition model;

[0049] Figure 5 Flowchart for training entity recognition model;

[0050] Figure 6 This is a flowchart of the emotion recognition process. Detailed Implementation

[0051] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the scope of protection of the present invention.

[0052] Example 1:

[0053] This invention provides a method for intervening in intelligent voice dialogue scenarios based on customer profiles, comprising: Step 1: performing scenario intent recognition, entity recognition, and multi-turn dialogue; Step 2: performing emotion recognition based on speech and semantics; Step 3: defining and recognizing a potential information tag system; Step 4: obtaining and storing customer profile information based on unique customer information through interface integration; Step 5: generating multiple applicable experience dialogues and recommendation values ​​based on the recognized customer scenario intent, customer emotion information, customer potential information tags, and customer profile information through a preset dialogue generation model, and randomly returning one experience dialogue each time the customer interacts with the intelligent voice; Step 6: controlling the response dialogue through intervention scope and intervention method configuration.

[0054] Based on traditional machine learning models, including Naive Bayes models and support vector machines, or neural network models, including convolutional neural networks and recurrent neural networks, the labeled training data is classified and learned. During recognition, the text extracted from the user's newly input voice information is mapped to the intent through classification, and the corresponding response is returned based on the dialogue flow.

[0055] By using a multimodal deep learning model that integrates speech and semantics, the system classifies and identifies users' speech and text to obtain their emotion categories. Through speech signal acquisition and semantic understanding, it extracts the emotional features of customers' speech information in intelligent voice scenarios.

[0056] The system creates corresponding tags for potential information beyond the customer's intent during dialogue interactions in different intelligent voice scenarios; it obtains response scripts based on an end-to-end text generation model. The text generation model is based on an encoder-decoder framework. First, the encoder encodes the input information through LSTM and Transformer neural networks. In the generation stage, the decoder combines the information from the encoder and the already generated text information to generate the next sentence.

[0057] Intervention Scope: Specify the knowledge scope of the intervention, including task-based conversations, FAQs, casual conversations, and third-party engines;

[0058] Intervention conditions: Set conditions for nodes, with options including: Attitude: affirmative, negative, no attitude; Emotion: positive, negative, neutral; Profile tags: built-in city, gender, age, and occupation type, with support for custom tags via customer interface;

[0059] Intervention actions: Supports adding sentence prefixes and suffixes and outputting fixed responses;

[0060] Priority can be adjusted among multiple rules according to their order. If multiple rules are enabled, condition judgments will be performed sequentially. Rules can be moved and dragged to meet the specific needs of different businesses and scenarios.

[0061] Example 2:

[0062] Example 2 is a preferred example of Example 1.

[0063] This invention provides a customer profile-based intelligent voice dialogue scenario dialogue optimization system. After receiving customer voice information in an intelligent voice dialogue scenario, the system identifies the customer's scenario intent, emotional information, and potential information tags. Simultaneously, it performs comprehensive analysis in conjunction with the customer profile. While processing the customer's scenario intent response, the system finds or generates the most matching dialogue from the pre-built dialogue library, thereby improving the overall interactive experience of intelligent voice dialogue.

[0064] By modeling and analyzing the identified scene intent, customer emotional information, potential customer information tags, and customer profile after receiving customer voice information in an intelligent voice scenario, the system can respond with the optimal experience script while meeting the scenario requirements.

[0065] By employing a script generation model, which uses the current customer's contextual intent, emotional information, potential information tags, and customer profile as input, customized scripts are generated for different customers, thereby reducing the cost of manually configuring the script library required by the original model. The script generation model is based on a large industry model pre-trained with relevant banking industry data, resulting in better performance.

[0066] By adopting the UNIT dialogue intervention and interaction mechanism, the original scripts can be intervened and optimized based on the current customer scenario intent. For different scenarios and different knowledge, the intervention-enhanced experience scripts can be added without affecting the original response content. This enables the ability to provide customers with personalized guidance, reassurance, and product recommendation experience scripts in actual business scenarios.

[0067] like Figure 1 This invention is divided into three stages: data collection and processing, model generation, and dialogue intervention application. In the dialogue intervention stage, it can generate call prediction scripts, and it can also generate personalized scripts based on the user's emotions and profile information during the human-machine interaction process.

[0068] like Figure 2 This is the system architecture diagram of the present invention. It is developed and deployed using cloud-native technologies to ensure high availability, high performance, and easy operation and maintenance of the system.

[0069] like Figure 3 This is a business scenario example of the present invention. Based on the user's profile information and business processing information, customized services are provided to the user to accelerate business processing and improve user satisfaction.

[0070] This invention provides an intelligent voice scenario dialogue intervention system based on customer profiles, comprising:

[0071] The scene intent dialogue module includes scene intent recognition, entity recognition, and a multi-turn dialogue process. Based on pre-configured scene intents, entities, and dialogue processes, corresponding models are trained. Upon receiving customer voice information in an intelligent voice scenario, the model identifies the customer's intent and entities, and returns the corresponding response based on the dialogue process. Customer intent recognition is essentially a classification problem; the input to a classification problem is a sample, and the output is a category. A classification model is learned using labeled training data. During recognition, the text extracted from the user's newly input voice information is mapped to the intent through the classification model. The model can generally be a traditional machine learning model, such as Naive Bayes or Support Vector Machine (SVM), or a neural network model, such as Convolutional Neural Networks (CNNS) or Recurrent Neural Networks (RNN).

[0072] like Figure 4 The process of training the intent recognition model involves preprocessing the labeled samples and then training the model to produce a classification model. For new samples that need to be recognized, the same preprocessing is performed before intent recognition is carried out based on the trained model.

[0073] Entity recognition is essentially a sequence labeling problem. The input to a sequence labeling problem is an observation sequence, and the output is a label sequence or a state sequence. A model is learned using labeled training samples. During recognition, a new observation sequence is input, and the output is a label sequence. The model can generally be a traditional machine learning model, such as a Hidden Markov Model (HMM) or a Conditional Random Field (CRF), or a neural network model, such as a Long Short-Term Memory (LSTM) network.

[0074] like Figure 5 This is the entity recognition model training process. After preprocessing the labeled samples, the model is trained to produce a sequence labeling model. For new samples that need to be recognized, the same preprocessing is performed, and the entity is recognized based on the trained model.

[0075] The multimodal emotion recognition module includes speech and semantic-based emotion recognition. It extracts emotional features from customer speech information in intelligent voice scenarios through speech signal acquisition and semantic understanding. Multimodal emotion recognition can define emotional features for banking industry scenarios and has the ability to train and predict models based on these emotional features. Essentially, multimodal emotion recognition is a classification task. Based on the characteristics of dialogue systems, the overall strategy adopts a hierarchical model structure. Each model uses a multimodal deep learning model integrating speech and semantics to classify and identify the user's emotions, such as pleasure, gratitude, complaint, anger, and disgust, from the user's input speech and text.

[0076] like Figure 6 This is the process of emotion recognition. For input speech and text, a multimodal deep learning model is used to first identify the positive or negative emotion, and then further subdivide it into the user's emotion. For example, "Thank you!" is first identified as a positive emotion, and then further subdivided into gratitude.

[0077] The potential information tagging module includes the definition and identification of the potential information tagging system. The so-called potential information mainly refers to some potential information beyond the customer's intention during the dialogue interaction process in different intelligent voice scenarios. For example, in the scenario of consulting financial products, if the customer repeatedly asks about financial knowledge, then the potential information tag "lack of financial knowledge" can be applied to the customer.

[0078] The customer profiling integration module mainly includes obtaining profile information; the interface integration method is based on the customer's unique information as input, returning the customer's profile information in real time, and storing and recording it.

[0079] The experience dialogue generation module primarily includes the ability to generate experience dialogue. Based on recognizing the customer's scenario intent, emotional information, potential information tags, and customer profile information, the dialogue generation model generates multiple applicable experience dialogues and recommended values. It can randomly return one experience dialogue each time the customer interacts with the intelligent voice system. The dialogue generation model is essentially an end-to-end text generation model. The input includes preceding information, the current scenario, and the customer's emotions; the output is the response dialogue. The model is based on an encoder-decoder framework. First, the encoder encodes the input information using LSTM, Transformer, etc. During the generation phase, the decoder combines the information from the encoder and the already generated text information to generate the content of the next sentence.

[0080] The dialogue intervention configuration module effectively controls response dialogue to avoid risks, including configuring the intervention scope and intervention method. During human-computer dialogue, the robot's response content is uniformly configured and indiscriminate. The answer generation intervention allows for dynamic adjustment of response content, intervening in standard response content based on user emotions, attitudes, and tags. This achieves personalized responses without altering the original meaning, improving user satisfaction. The module comprises three parts: Intervention Scope: Specifies the knowledge range that this rule can intervene in, including task-based conversations, FAQs, casual conversation, and third-party engines. Intervention Conditions: Similar to node condition settings, supporting both normal and advanced modes. Selectable content includes: Attitude – affirmative, negative, no attitude; Emotion – positive, negative, neutral; Profile Tags – built-in city, gender, age, and occupation type, with support for custom tags via the customer interface. Intervention Actions: Supports adding sentence prefixes and suffixes and outputting fixed responses. The intervention configuration module allows you to intervene in one or more specific scenario intents, or select a specific dialogue flow node to intervene in for that intent scenario. The intervention scope allows for convenient control over the impact of the experience-based dialogue strategy. The intervention method configuration allows you to choose between direct response and response before or after the original script. The intervention method provides better control over dialogue flow nodes in different scenario intents. Multiple rules can be prioritized according to their order; if multiple rules are enabled, they can be evaluated sequentially. Rules can be moved and dragged to meet specific needs in different business scenarios.

[0081] Those skilled in the art will understand that, in addition to implementing the system, apparatus, and their modules provided by this invention in purely computer-readable program code, the same program can be implemented in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers by logically programming the method steps. Therefore, the system, apparatus, and their modules provided by this invention can be considered a hardware component, and the modules included therein for implementing various programs can also be considered structures within the hardware component; alternatively, modules for implementing various functions can be considered both software programs implementing the method and structures within the hardware component.

[0082] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A customer portrait-based intelligent voice dialogue scene script intervention method, characterized in that, Comprise: Step 1: scene intent recognition, entity recognition and multi-turn dialogue are carried out; Step 2: emotion recognition is carried out based on voice and semantics; Step 3: potential information label system definition and identification are carried out; Step 4: through the interface docking mode, based on the unique information of the customer, the portrait information of the customer is obtained and stored; Step 5: based on the identified scene intent of the customer, the emotional information of the customer, the potential information label of the customer and the customer portrait information, a plurality of applicable experience dialogues and recommendation values are generated through a preset dialogue generation model, and an experience dialogue is randomly returned each time the customer and the intelligent voice interact; Step 6: the reply dialogue is controlled through intervention range configuration and intervention mode configuration; Intervention range: specify the knowledge range of intervention, including task-based conversation, FAQ question and answer, casual chat and third-party engine; Intervention condition: node condition setting, the selected content includes: attitude: positive, negative, no attitude; Emotion: positive, negative, neutral; Image label: system built-in city, gender, age, occupation type, supporting customer interface custom label; Intervention action: support adding prefixes and suffixes and outputting fixed replies; Adjust the priority among multiple rules according to the order, if multiple rules are enabled, make conditional judgment in turn, and the rules support moving and dragging to realize the specific needs in different businesses and different scenes.

2. The method of claim 1, wherein the method further comprises: Based on traditional machine learning models, including Naive Bayes model, support vector machine, or neural network model, including convolutional neural network, recurrent neural network, the labeled training data is classified and learned, and the text extracted from the new input voice information of the user is mapped to the intent through classification, and the corresponding reply dialogue is returned based on the dialogue process.

3. The method of claim 1, wherein the method further comprises: Through the multi-modal deep learning model of voice and semantic integration, the voice and text of the user are classified and recognized to obtain the emotion category of the user, and the emotional features of the customer in the voice information in the intelligent voice scene are extracted through voice signal collection and semantic understanding.

4. The method of claim 1, wherein the method further comprises: Potential information corresponding to the intent existing in the dialogue interaction process of the customer in the intelligent voice in different scenes is created; The reply dialogue is obtained based on an end-to-end text generation model, and the text generation model is based on an encoder-decoder framework. First, the encoder encodes the input information through an LSTM, a Transformer neural network, and in the generation stage, the decoder integrates the information of the encoder and the generated text information to generate the next sentence content.

5. A customer portrait-based intelligent voice dialogue scene script intervention system, characterized in that, Comprise: Scene intent dialogue module: scene intent recognition, entity recognition and multi-turn dialogue are carried out; Multi-modal emotion recognition module: emotion recognition is carried out based on voice and semantics; Potential information label module: potential information label system definition and identification are carried out; Customer portrait docking module: through the interface docking mode, based on the unique information of the customer, the portrait information of the customer is obtained and stored; The experience dialogue generation module generates multiple applicable experience dialogues and recommendation values based on the identified scene intention of the customer, the emotional information of the customer, the potential information label of the customer and the customer portrait information, and randomly returns an experience dialogue each time the customer interacts with the intelligent voice. The dialogue intervention configuration module controls the reply dialogue through intervention range configuration and intervention mode configuration. The dialogue intervention configuration module includes: Intervention range: specifies the knowledge range of intervention, including task-based conversation, FAQ question and answer, casual conversation and third-party engine; Intervention condition: node condition setting, content selection including: attitude: positive, negative, no attitude; emotion: positive, negative, neutral; portrait label: system built-in city, gender, age, occupation type, supporting customer interface custom label; Intervention action: supports adding prefixes and suffixes before and after sentences and outputting fixed replies; The priority is adjusted according to the sequence among multiple rules. If multiple rules are enabled, the condition is judged in sequence. The rule supports moving and dragging to meet the specific needs in different businesses and different scenarios.

6. The customer persona based intelligent voice conversation scenario scripting intervention system as claimed in claim 5 wherein, The scene intention dialogue module includes: based on traditional machine learning models, including Naive Bayes model, support vector machine, or neural network model, including convolutional neural network, recurrent neural network, classifying learning on annotated training data, mapping the text extracted from the new input voice information of the user to the intention through classification in the identification, and returning the corresponding reply dialogue based on the dialogue process.

7. The customer persona based intelligent voice conversation scenario scripting intervention system as claimed in claim 5 wherein, The multi-modal emotion recognition module includes: through the multi-modal deep learning model of voice and semantic integration, classifying and recognizing the voice and text of the user to obtain the emotional category of the user, and extracting the emotional features of the customer in the voice information in the intelligent voice scene through voice signal collection and semantic understanding.

8. The customer persona based intelligent voice conversation scenario scripting intervention system as claimed in claim 5 wherein, The potential information label module includes: creating corresponding labels for potential information other than the intention of the customer in the dialogue interaction process in different intelligent voice scenarios. The experience dialogue generation module includes: obtaining the reply dialogue based on the end-to-end text generation model, the text generation model is based on the encoder-decoder framework, first encoder encodes the input information through LSTM, Transformer neural network, in the generation stage, decoder integrates the information of encoder and the generated text information to generate the next sentence content.

Images