Intelligent dialogue method and system based on generative model

Abstract

The invention relates to the technical field of artificial intelligence natural language processing, and discloses an intelligent dialogue system based on a generative model, and the system comprises a semantic feature extraction unit which is used for carrying out the online detection of the intention of the dialogue content of a user, and achieving the multi-round and context association detection; the service terminal carries a generative dialogue model, a model processing module and a data storage unit, the generative dialogue model analyzes scenes, frequencies and corresponding dialogue targets of interaction intentions on the basis of historical dialogue records, and the model processing module optimizes intention understanding through feature vectorization conversion, an attention mechanism and the like, so that the interaction intention is analyzed. The data storage unit is used for storing multi-dimensional data; and the interaction feedback unit displays the intention expression scene. The system realizes accurate analysis and interactive response adjustment of intentions through a dynamic adaptive sequence generation algorithm, an attention mechanism, scene coordinate system construction and the like. The intention understanding accuracy and the interaction service quality of the intelligent dialogue system are improved, and the method is suitable for scenes needing efficient intelligent dialogues.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence natural language processing, in particular to an intelligent dialogue method and system based on a generative model. BACKGROUND

[0002] In terms of intent detection, traditional intelligent dialogue systems often have difficulty in achieving accurate detection of multiple rounds and context association. Most systems can only identify intent for a single round of dialogue and cannot effectively combine historical dialogue content and context, resulting in large deviations in intent understanding in complex dialogue scenarios. For example, when a user switches topics or expresses implied intent in multiple rounds of dialogue, the system often cannot accurately capture it, affecting the interactive experience.

[0003] From the application of generative models, the existing technology does not sufficiently integrate and utilize historical dialogue records, scene information, and user portraits in the training and application process of generative dialogue models. The model has difficulty in analyzing the scene and frequency of interactive intent and understanding the specific content of intent expression in combination with the dialogue target in the corresponding scene, making the generated response content lack pertinence and adaptability and unable to meet the personalized needs of users.

[0004] In the data processing and intent understanding link, traditional methods have limited conversion and optimization processing capabilities for natural language signals. Through simple feature extraction and model processing, it is difficult to effectively realize noise filtering, context fusion, and intent intensity analysis, resulting in low accuracy and efficiency of intent understanding. At the same time, the data storage unit has a single function and cannot comprehensively store multi-dimensional information such as user dialogue basic data, interactive data received by the semantic feature extraction unit, and intent processing records, which is not conducive to the optimization of the model and the improvement of dialogue quality.

[0005] The existing intelligent dialogue system lacks intuitive display of specific scenes of intent expression in interactive feedback, and users have difficulty in clearly observing the needs associated with dialogue interaction. The system cannot dynamically adjust the intensity of interactive response according to the intensity and frequency of intent expression, making it difficult to achieve intelligent and precise interactive services. SUMMARY

[0006] The present application aims to provide an intelligent dialogue method and system based on a generative model to solve the problems raised in the background art.

[0007] To achieve the above-mentioned purpose, the present application provides the following technical solution: an intelligent dialogue system based on a generative model, the system comprising:

[0008] a semantic feature extraction unit for detecting intent in user dialogue content in an online state, achieving multi-round and context-associated detection;

[0009] The service terminal is internally provided with a generative dialogue model, a model processing module and a data storage unit, and receives interactive data of the semantic feature extraction unit; the generative dialogue model is used to obtain historical dialogue records as a basic reference basis for model training, analyze the scene and frequency of interactive intention, and analyze the specific content of intention expression in combination with the dialogue target under the corresponding scene; the model processing module is used to convert natural language signals into digital features through feature vectorization conversion, and optimize intention understanding through attention mechanism, sequence modeling, noise filtering and context fusion; the data storage unit is used to store various user dialogue basic data, interactive data received by the semantic feature extraction unit, intention processing records, user portrait information data and dialogue network topology data.

[0010] The interactive feedback unit is used to display the specific scene of intention expression, and intuitively observe the user demand associated with the corresponding dialogue interaction.

[0011] Preferably, the semantic feature extraction unit captures interactive intention signals, and then sends the intention signals to the model processing module for data processing, and the processed intention signal data is conducted to the generative dialogue model, which analyzes the intention scene and intention intensity in combination with the user dialogue basic data, dialogue network topology information and historical interactive data, and sends the analysis result to the interactive feedback unit for display.

[0012] Preferably, the data storage unit receives user dialogue basic data, copies and sends the intention signal data to the data storage unit for storage backup when the semantic feature extraction unit receives the intention signal, and sends the processing records and results to the data storage unit for storage backup after the generative dialogue model completes the processing of the intention signal data.

[0013] Preferably, the present application further comprises an intelligent dialogue method based on a generative model, according to the intelligent dialogue system based on the generative model, the method comprises the following specific steps:

[0014] S1, collecting basic data and constructing a generative dialogue model;

[0015] S2, performing intention signal screening verification, processing, analyzing, screening and determining the captured intention signal;

[0016] S3, performing interactive actual execution, analyzing the intention signal, and adjusting the interactive response strength according to the intention expression strength, intention occurrence frequency and user dialogue basic data.

[0017] Preferably, the step S1 comprises:

[0018] S1.1, selecting the deployment position of the semantic feature extraction unit according to the interaction habits of each user of the dialogue service platform;

[0019] S1.2, collecting basic data of the dialogue network within the processing range of the generative dialogue model, including: user geographic location, interaction preference, use duration, historical dialogue content and service feedback record, as well as dialogue network topology information and global view of the dialogue service platform, and inputting the collected basic data into the data storage unit for storage;

[0020] S1.3, constructing the generative dialogue model based on the dynamic self-adaptive sequence generation algorithm, combining user dialogue basic data, dialogue network topology information and user portrait information, and combining the attention mechanism into the generative dialogue model;

[0021] S1.4, obtaining the intention signal detected by the semantic feature extraction unit that has been put into use as the training set of the generative dialogue model, and training the generative dialogue model using the training set;

[0022] S1.5, the generative dialogue model receives user dialogue basic data, establishes a scene coordinate system with the user in the dialogue network, and the origin of the coordinate system is the position of the semantic feature extraction unit;

[0023] S1.6, proportionally adjusting the global view of the dialogue service platform and merging it into the coordinate system, and verifying the accuracy of the dialogue path in the dialogue network and the corresponding position in the view;

[0024] S1.7, when the intention signal is obtained, the signal is copied and sent to the model processing module and the data storage unit respectively, and when the intention signal analysis result is obtained by the generative dialogue model, the result is copied and sent to the interaction feedback unit and the data storage unit respectively.

[0025] Preferably, the intention signal screening and verification specifically includes the following steps:

[0026] S2.1, capturing the intention signal by the semantic feature extraction unit;

[0027] S2.2, converting the natural language signal into digital features by the model processing module through feature vectorization conversion, optimizing the intention understanding through attention mechanism, sequence modeling, noise filtering and context fusion, and then sending the processed interaction intention signal to the generative dialogue model;

[0028] S2.3, the generative dialogue model analyzes the interaction intention signal, then extracts the intention signal scene information and user scene information, and compares them;

[0029] S2.4, the semantic feature extraction unit determines the relative relationship between the intent signal scene and the user scene according to the capture error of the scene information, and judges, and then sends service personnel to the dialogue link to check;

[0030] S2.5, after checking, upload the service log to the data storage unit, and distribute the service log as service feedback record data in the corresponding user dialogue basic data by the generative dialogue model.

[0031] Preferably, the interaction actually includes the following steps:

[0032] S3.1, the generative dialogue model receives the intent signal, processes the intent signal using a dynamic adaptive sequence generation algorithm, extracts the scene information of the intent signal, and marks the intent signal data in the coordinate system according to the intent signal scene;

[0033] S3.2, use attention mechanism to analyze the intent signal scene distribution data and user dialogue basic data;

[0034] S3.3, check the intent signal scene, when the signal error range of the semantic feature extraction unit is within the user scene, it is considered as the user intent signal, otherwise it is considered as the dialogue network facility intent signal;

[0035] S3.4, the generative dialogue model analyzes the user interaction intent for the user intent signal and the dialogue network facility intent signal respectively;

[0036] S3.5, the generative dialogue model combines the user interaction intent to analyze the dialogue network intent, and according to the intent expression intensity, the intent appearance frequency and the user dialogue basic data, the abnormal situation of local dialogue interaction appears is warned.

[0037] Preferably, in step S2.4, the specific way of judgment is:

[0038] 1) the intent signal scene error is within the user scene range, then the generative dialogue model calls the corresponding user basic data and exports, and sends service personnel to the dialogue link to check;

[0039] 2) the intent signal scene error is outside the user scene range, then the generative dialogue model matches the intent signal scene according to the dialogue network, and calls the corresponding user basic data in the network, analyzes the users directly affected by the corresponding dialogue network of the intent scene, analyzes the weight relationship of the corresponding users affected, and sends service personnel to check the users according to the weight descending order.

[0040] Preferably, in step S3.4, the specific processing method of user interaction intent analysis is:

[0041] 1) The generative dialogue model retrieves user intent signals, interaction rule data in dialogue network topology information, and user dialogue basic data to analyze user interaction intent, and then dispatches personnel to adjust the user dialogue link;

[0042] 2) The generative dialogue model retrieves dialogue network facility intent signals, dialogue network topology information, user geographic location, and user real-time interaction content to analyze other facility intent, and then dispatches personnel to investigate the intent scene.

[0043] Preferably, in step S3.5, the specific processing mode of dialogue network intent analysis is that the generative dialogue model sets the interaction response priority according to the intent expression intensity, adjusts the model training parameters according to the intent occurrence frequency, optimizes the dialogue generation strategy in combination with the service feedback records in the user dialogue basic data, and generates early warning prompt information for possible abnormal situations in local dialogue interaction.

[0044] Compared with the prior art, the present application has the following advantages:

[0045] In terms of semantic feature extraction and intent detection, the semantic feature extraction unit can detect the intent of the user dialogue content in an online state, realize multi-round and context-related detection, effectively solve the limitations of traditional single-round dialogue intent recognition, enable the system to accurately capture the intent changes of the user in multi-round dialogue, and improve the intent understanding ability in complex dialogue scenarios.

[0046] The application of the generative dialogue model greatly optimizes the intent analysis process. The model takes historical dialogue records as the basis for training reference, can deeply analyze the scene and frequency of interaction intent, and accurately understand the specific content of intent expression in combination with the dialogue target in the corresponding scene. In this way, the response content generated by the system is more targeted and can better meet the individual needs of users in different scenarios.

[0047] The model processing module converts natural language signals into digital features through feature vectorization conversion, and optimizes intent understanding by using attention mechanism, sequence modeling, noise filtering, and context fusion technologies, significantly improving the accuracy and efficiency of intent understanding. This enables the system to more accurately grasp the user's intent, reduce understanding bias, and improve the smoothness and effectiveness of interaction.

[0048] The design of the data storage unit is comprehensive and efficient, and can store user dialogue basic data, interaction data, intent processing records, user portrait information data, dialogue network topology data, and other multi-dimensional information. Not only does it achieve comprehensive data backup, but it also provides rich data support for model training and optimization, which helps to continuously improve the performance and dialogue quality of the system.

[0049] The interactive feedback unit can intuitively display the specific scene of the intention expression, so that the user and the service personnel can clearly observe the user demand associated with the dialogue interaction, and provide an intuitive basis for subsequent interaction adjustment and service optimization.

[0050] In the method steps, the dynamic adaptive construction and training of the model are realized by collecting basic data and constructing the generative dialogue model; the accuracy of the intention signal is ensured by comprehensively processing, analyzing, screening and determining the captured intention signal in the intention signal screening and verification process; and the intelligent and precise interactive service is realized by adjusting the interaction response strength according to the intention expression intensity, frequency and user dialogue basic data in the interactive actual execution link.

[0051] The application of the dynamic adaptive sequence generation algorithm enables the model to better process the intention signal, extract scene information and mark in the coordinate system, analyze the scene distribution and user data in combination with the attention mechanism, and improve the accuracy of scene analysis and the pertinence of interactive response. By determining the relative relationship between the intention signal scene and the user scene, the service personnel are dispatched for troubleshooting, and the fault troubleshooting and service optimization capability of the system is improved. The user intention signal and the dialogue network facility intention signal are analyzed respectively, and the warning prompt information is generated in combination with the intention expression intensity and frequency, the dialogue network abnormal condition is discovered and processed in time, and the stable operation of the dialogue network is ensured. BRIEF DESCRIPTION OF DRAWINGS

[0052] Figure 1 The working principle diagram of the intelligent dialogue system based on the generative model is described in the present application;

[0053] Figure 2 The flowchart of semantic feature extraction and processing is described in the present application;

[0054] Figure 3 The general flowchart of the intelligent dialogue method is described in the present application;

[0055] Figure 4 The flowchart of the intention signal screening and verification is described in the present application. DETAILED DESCRIPTION

[0056] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0057] Please refer to Figures 1-4The application relates to an intelligent dialogue system based on a generative model, which comprises a semantic feature extraction unit, a service terminal and an interaction feedback unit. The specific implementation steps are as follows:

[0058] The semantic feature extraction unit is used for intent detection on user dialogue content in an online state, realizing multi-round and context-related detection. Specifically, the semantic feature extraction unit can capture the intention signals of the user in the dialogue process in real time, which may include the user's demand, question, instruction and the like. Through analysis of multi-round dialogue and association of the context, the real intention of the user can be more accurately understood. For example, when the user asks questions in multiple rounds, the semantic feature extraction unit can combine the previous dialogue content to judge the background and purpose of the current question of the user, so that the intention can be more accurately detected.

[0059] The service terminal is internally provided with a generative dialogue model, a model processing module and a data storage unit, and receives the interaction data of the semantic feature extraction unit. The generative dialogue model is used to obtain historical dialogue records as a basic reference basis for model training, analyze the scene and frequency of the occurrence of the interaction intention, and analyze the specific content of the intention expression in combination with the dialogue target in the corresponding scene. The model processing module is used for converting natural language signals into digital features through feature vectorization conversion, and optimizing intention understanding through attention mechanism, sequence modeling, noise filtering and context fusion. The data storage unit is used to store various user dialogue basic data, interaction data received by the semantic feature extraction unit, intention processing records, user portrait information data and dialogue network topology data. After receiving the interaction data of the semantic feature extraction unit, the model processing module of the service terminal will first process the data, and convert the natural language into digital features, so as to analyze the generative dialogue model. The generative dialogue model is trained by using historical dialogue records, and can better understand the intention expression in different scenes, for example, in an e-commerce scene, the user may inquire about the price, size and other information of the goods, and the generative dialogue model can analyze the intention of the current user according to the processing mode of the similar scene in the historical dialogue. The data storage unit stores all related data for subsequent query and use.

[0060] The interaction feedback unit is used to display the specific scene of the intention expression, and intuitively observe the user demand associated with the corresponding dialogue interaction. When the generative dialogue model analyzes the specific scene of the intention expression and the user demand, the interaction feedback unit displays these information in an intuitive way, for example, through a chart, a text or the like, so that the relevant personnel can quickly understand the user demand.

[0061] Embodiment 1:

[0062] In the system, the semantic feature extraction unit captures the interactive intent signal, and then sends the intent signal to the model processing module for data processing. The processed intent signal data is transmitted to the generative dialogue model. The generative dialogue model analyzes the intent scene and intent intensity in combination with user dialogue basic data, dialogue network topology information, and historical interaction data, and sends the analysis results to the interaction feedback unit for display. The specific implementation is as follows:

[0063] In the actual operation process of the system, the semantic feature extraction unit is in an online monitoring state, continuously scanning the user's dialogue content in real time. Its working mechanism is based on natural language processing technology to analyze the text, voice and other dialogue forms input by the user, identify the key information that can reflect the user's intent from them, and these key information constitutes the interactive intent signal. For example, when the user mentions "product price" and "preferential activities" in the dialogue on the e-commerce platform, the semantic feature extraction unit will capture the intent signal carried by these words and judge that the user may have the intention to purchase consultation.

[0064] After capturing the intent signal, the semantic feature extraction unit will immediately send the signal to the model processing module in a specific data format. The data transmission process here needs to ensure the integrity and real-time nature of the signal to avoid data loss or delay in the transmission process, thereby affecting the subsequent processing efficiency. After receiving the intent signal, the model processing module first performs a feature vectorization conversion operation. The essence of this operation is to convert the intent signal in natural language form into a digital feature vector that can be processed by a computer. Specifically, the model processing module will perform word segmentation, part-of-speech tagging and other preprocessing on the words and sentences in the intent signal, and then map each word to a vector in a high-dimensional space through word embedding technology, so that these vectors can retain the semantic information of the words.

[0065] After the feature vectorization conversion is completed, the model processing module will use various techniques to optimize the digital features to improve the accuracy of intent understanding. Among them, the application of attention mechanism can make the model pay more attention to the key part of the intent signal related to the current scene. For example, when the user asks "how is the battery life of this phone?", the attention mechanism will make the model focus on the key information "battery life" and ignore other irrelevant content. Sequence modeling technology is used to process the temporal relationship of intent signals, because the user's intent expression often has a certain order, and through sequence modeling, the development context of the intent can be better understood. Noise filtering technology can remove interference information that may exist in the intent signal, such as user input errors, meaningless mood words, etc., making the intent signal more pure. Context fusion technology combines the current intent signal with the previous dialogue history to better understand the user's intent. For example, the user mentioned "mobile phone brand" in the previous round of dialogue, and asked "price" in the current round, the context fusion technology can combine the information of the two rounds of dialogue to understand that the user is asking the price of a specific brand of mobile phone.

[0066] After a series of processing by the model processing module, the optimized intent signal data will be transmitted to the generative dialogue model. After receiving the data, the generative dialogue model will retrieve various data from the data storage unit for comprehensive analysis. Among them, the user dialogue basic data contains the user's historical dialogue content, interaction preferences and other information, which can help the generative dialogue model understand the user's habits and needs. The dialogue network topology information describes the connection relationship and interaction rules between nodes in the dialogue network, which helps the generative dialogue model determine the network environment of the intent signal. Historical interaction data records the processing method and results of similar intent signals in the past, providing a reference for current analysis.

[0067] The generative dialogue model first analyzes the intent scenario. It will judge the specific scenario of the intent according to the key information in the intent signal and the retrieved data. For example, in the customer service scenario, the user mentions "order inquiry", the generative dialogue model can combine the relevant rules of the customer service scenario in the dialogue network topology information to determine that it is an order inquiry scenario. At the same time, the generative dialogue model will also evaluate the intent intensity. The evaluation of intent intensity includes the urgency of the user's expression of intent, the frequency of use of keywords, etc. For example, the user repeatedly asks the same question or uses words such as "urgently needed" "as soon as possible", which usually indicates a high intent intensity.

[0068] After the generative dialogue model completes the analysis of the intent scenario and intent intensity, it sends the analysis results in a structured data form to the interaction feedback unit. After receiving the data, the interaction feedback unit converts it into intuitive display content. For example, it displays the distribution of intent scenarios in the form of a chart, using different colors or icons to represent the level of intent intensity. Relevant personnel can quickly understand the user's intent scenario and intent intensity through the display content of the interaction feedback unit, and make appropriate responses accordingly.

[0069] Throughout the process, data transmission and processing between units are closely linked. The semantic feature extraction unit, model processing module, generative dialogue model, and interaction feedback unit form a complete processing chain, with each link playing an important role in ensuring accurate understanding and timely feedback of user intent. For example, in a financial customer service scenario, a user asks, "What is my account balance?" The semantic feature extraction unit captures this intent signal and sends it to the model processing module, which converts it into numerical features and optimizes the processing before transmitting it to the generative dialogue model. The generative dialogue model analyzes that this is an account inquiry scenario with general intent intensity based on the user's historical transaction data, account information, etc., and then sends the analysis results to the interaction feedback unit for display. Customer service personnel can provide account balance inquiry services for users based on the display content.

[0070] Embodiment 2:

[0071] The data storage unit receives user dialogue basic data, and when the semantic feature extraction unit receives the intent signal, it copies and sends the intent signal data to the data storage unit for storage backup. After the generative dialogue model completes the processing of the intent signal data, it sends the processing records and results to the data storage unit for storage backup. The specific implementation is as follows:

[0072] In the initial stage of system operation, the data storage unit begins to continuously receive user dialogue basic data. These data cover multiple dimensions, including user's basic identity information, historical dialogue content, interaction habits, device information, and timestamp of dialogue occurrence. For example, the personal information submitted by the user when registering for the dialogue service platform, as well as the text, voice, and other interactive content generated during each dialogue process, are transmitted to the data storage unit in real time. The data storage unit uses a distributed storage architecture to store these data according to different categories and user identifiers, forming a structured database for subsequent fast retrieval and calling.

[0073] When the semantic feature extraction unit monitors user conversation content and captures intent signals in online mode, the data replication mechanism is triggered immediately. Specifically, the semantic feature extraction unit will make a complete copy of the original data containing the intent signal, using incremental replication technology to ensure data integrity and consistency, and to avoid deviations in subsequent analysis due to data loss. After replication is complete, the intent signal data is sent to the data storage unit through a dedicated data transmission channel. After receiving the intent signal data, the data storage unit will first verify the data, including data format verification, integrity verification, and validity verification. After verification, it will store the data in the corresponding database table according to the type of intent signal and user identification, and generate a unique data storage identifier for subsequent tracking and management.

[0074] During the processing of intent signal data by the generative dialogue model, the model processing module performs a series of operations such as feature vectorization conversion, attention mechanism optimization, sequence modeling, noise filtering, and context fusion on the intent signal. The generative dialogue model analyzes the intent scenario and intent intensity in combination with user conversation basic data, dialogue network topology information, and historical interaction data. After the generative dialogue model completes the processing of the intent signal data, it forms a complete processing record and analysis result. The processing record includes the algorithms used, parameter settings, intermediate results, and other information during processing, and the analysis result includes intent scenario judgment, intent intensity evaluation, and recommended interaction response strategies.

[0075] The generative dialogue model packages the processing record and result in a standardized data format and sends it to the data storage unit for storage backup. After receiving the data package, the data storage unit performs data verification again to ensure that the data has not been corrupted or lost during transmission. After verification, the processing record and result are stored in association with the corresponding intent signal data and user conversation basic data according to their nature. For example, the processing record and result are stored in the database table associated with the intent signal data, and a fast retrieval channel is established through the indexing mechanism.

[0076] The data storage unit also regularly maintains and manages the stored data. For example, using data archiving technology, historical data that has not been accessed for a long time is migrated to archival storage media to release space on the main storage device while ensuring the accessibility of historical data. Using data backup strategies, critical data is regularly backed up off-site to prevent data loss due to hardware failure, natural disasters, and other reasons. In addition, the data storage unit also has data security protection mechanisms, including access control, data encryption, security auditing, etc., to ensure that the stored data cannot be accessed or tampered with by unauthorized personnel.

[0077] In practical application scenarios, for example, in an intelligent customer service system, when a user inquires about the use method of a product, the semantic feature extraction unit captures the user's intention signal and sends it to the data storage unit for storage backup. The generative dialogue model processes the intention signal, analyzes the user's specific needs and problem scenarios, and generates corresponding processing results and suggested reply content. After processing is completed, the generative dialogue model sends the processing record and result to the data storage unit, which stores it in association with the user's dialogue basic data, intention signal data, etc. When a subsequent user inquires about a similar problem again, the system can quickly retrieve relevant historical data and processing records from the data storage unit to provide a reference for generating more accurate and efficient replies.

[0078] For example, in an intelligent recommendation system on an e-commerce platform, a user mentions interest in a certain type of product in a conversation, and the semantic feature extraction unit captures and stores this intention signal. After processing, the generative dialogue model analyzes the user's shopping intention and preference, generates a recommendation strategy and a list of products. After storing the processing record and result, the data storage unit can provide data support for subsequent user portrait updating and recommendation algorithm optimization, enabling the system to better understand user needs and provide more personalized services.

[0079] This data storage and backup mechanism of the data storage unit ensures that all key data generated during the operation of the system is saved completely and reliably. These data not only support the current dialogue interaction, but also provide rich historical data resources for the long-term optimization and iteration of the system. Through analysis and mining of these data, the performance of the generative dialogue model can be continuously improved, the detection algorithm of the semantic feature extraction unit can be optimized, and the service quality and user experience of the entire intelligent dialogue system can be improved.

[0080] Embodiment 3:

[0081] Based on the above intelligent dialogue method of the intelligent dialogue system, three specific steps are included: collecting basic data and constructing a generative dialogue model, performing intention signal screening and verification, and performing interactive actual execution. The specific implementation is as follows:

[0082] In the stage of collecting basic data and constructing a generative dialogue model, the deployment position of the semantic feature extraction unit needs to be selected according to the interaction habits of each user of the dialogue service platform. For example, in an e-commerce platform, by analyzing common dialogue scenarios of users, such as product consultation, order inquiry, after-sales service, etc., the semantic feature extraction unit is deployed in the customer service dialogue box, order detail page, after-sales application portal, and other positions where users interact frequently, to ensure that it can efficiently capture intention signals.

[0083] Next, the basic data of the dialogue network within the processing range of the generative dialogue model is collected. Taking an education dialogue service platform as an example, the collected data includes user geographic location (such as distribution in different cities), interaction preferences (such as preference for text consultation or voice questioning), usage duration (average time of each consultation), historical dialogue content (such as inquiry about course arrangement, faculty strength, etc.), and service feedback records (satisfaction evaluation of customer service response), as well as dialogue network topology information (such as associated paths between different consultation boards) and global view of the dialogue service platform (overall architecture and function distribution of the platform). These data are transmitted in real time to the data storage unit through the data collection interface and stored in a structured table form for subsequent calling.

[0084] Then, the generative dialogue model is constructed based on the dynamic adaptive sequence generation algorithm. The historical dialogue sequence in the user dialogue basic data, the node connection rule in the dialogue network topology information, and the user portrait information (such as education level, age layer) are input into the model, and the attention mechanism is also integrated. For example, when processing the user's dialogue about "postgraduate course recommendation", the attention mechanism will focus on the professional direction, target university, etc. mentioned in the user's historical dialogue, improving the understanding accuracy of the current intent.

[0085] After that, the intent signals detected by the semantic feature extraction unit that has been put into use are obtained as the training set. For example, during the trial operation of the platform, the "course price" and "class hour arrangement" intent signals captured by the semantic feature extraction unit are labeled by artificial marking to form a training set, which is used to iteratively train the generative dialogue model, adjust the model parameters, and make the model more accurately generate responses that meet the scene.

[0086] After the generative dialogue model receives the user dialogue basic data, it establishes a scene coordinate system based on the user's dialogue network. For example, when the user is in the "English four and six level course" consultation board, the deployment position of the semantic feature extraction unit corresponding to the board is taken as the origin of the coordinate system, the x-axis represents the time sequence of the consultation process, the y-axis represents the question type (such as course content, registration process, etc.), and the z-axis represents the user portrait dimension (such as grade, English level), to construct a three-dimensional scene coordinate system.

[0087] Then, the global view of the dialogue service platform is scaled by 1:100 and merged into the coordinate system, and the accuracy of the dialogue path in the dialogue network and the corresponding position in the view is verified. For example, it is checked whether the mapping of the dialogue path from the "course detail page" to the "registration entrance" in the coordinate system is consistent with the actual layout of the platform, to ensure that the model can accurately locate the intent scene.

[0088] When an intent signal is obtained, such as when the user sends "Is this course live?", the system immediately copies the signal and sends it to the model processing module and data storage unit respectively; after the generative dialogue model analyzes and concludes that the intent belongs to the "course format consultation" scenario and is of medium intensity, the result is copied and sent to the interactive feedback unit and data storage unit respectively to achieve synchronous backup and processing of data.

[0089] During the intent signal screening and verification phase, for example, a user inquires about the "course refund policy." After the semantic feature extraction unit captures this intent signal, the model processing module performs feature vectorization, converting "refund policy" into a digital feature vector containing dimensions such as "refund conditions" and "delivery time." The attention mechanism then focuses on the keyword "refund," combining it with sequence patterns of similar questions in previous conversations to build a model. This filter eliminates noise from interjections like "please ask," and incorporates contextual information about the user's previous inquiries about "course validity period." This optimizes intent understanding and sends the processed signal to the generative dialogue model. The model extracts the contextual information for the intent (inquiring about the refund policy) and compares it with the user context (within 7 days of purchasing the course). If the semantic feature extraction unit's capture error is within the user context range (e.g., if the user's actual purchase time is 5 days, the error is within ±2 days), the user's basic data is retrieved and customer service personnel are dispatched to review the conversation. If the error exceeds this range, the conversation network matches the "user inquiring about a refund for a course they haven't purchased" scenario, retrieves the user basic data within the corresponding network, analyzes the weights of the affected users, and arranges customer service personnel for investigation in descending order of weight. After troubleshooting, the service log is uploaded to the data storage unit and assigned by the model as a service feedback record in the user conversation basic data.

[0090] In the actual execution phase of interaction, the generative dialogue model receives the user's intention signal of "course video cannot be played", processes it using a dynamic adaptive sequence generation algorithm, extracts the "technical failure" scene information, and marks it in the coordinate system as (play page, 3rd lesson, sophomore). Through attention mechanism analysis of the scene distribution data and user dialogue basic data (the user has appeared many times recently), the signal error range is checked. If it belongs to the user's intention signal, the technical support interaction rule in the dialogue network topology information is called, combined with the user's historical playback record analysis as "cache problem", and technical personnel are dispatched to remotely assist in cleaning up the cache; if it belongs to the dialogue network facility intention signal (such as server node failure), the network topology information, user geographic location (such as Beijing node) and real-time interaction content are analyzed as server overload, and personnel are dispatched to check the computer room equipment. At the same time, the model sets a high response priority according to the intention expression intensity (the user has fed back for 3 times in a row), adjusts the model training parameters according to the intention frequency (the problem has occurred 100 times in the past week), increases the training samples of the "video playback failure" scene, optimizes the dialogue generation strategy combined with the user service feedback record (such as the previous similar problem was solved by restarting), generates a warning prompt information of "suggesting to restart the device", and pushes it to the user who encounters the same problem in the future.

[0091] The whole method forms a closed-loop processing flow through the collection and model construction of basic data, the accurate screening and verification of intention signals, and the dynamic execution of interaction strategies. For example, in the medical consultation scene, the user describes "headache with fever", and the system constructs a model by collecting its geographic location (an area with a high incidence of epidemic situation), interaction preferences (text description), etc. In the screening and verification stage, the historical case data is combined to distinguish between common cold and infectious disease warning intention, and in the interaction execution stage, the video consultation is arranged in priority according to the intention intensity, and the disease judgment parameters of the model are updated according to the frequency of similar symptoms, so as to realize the accurate processing of the whole process of intelligent dialogue.

[0092] Embodiment 4:

[0093] The intent signal screening and checking specifically includes capturing the intent signal by the semantic feature extraction unit, converting the natural language signal into digital features by the model processing module through feature vectorization, optimizing the intent understanding through attention mechanism, sequence modeling, noise filtering, and context fusion, and then sending the processed interactive intent signal to the generative dialogue model; the generative dialogue model analyzes the interactive intent signal, extracts the intent signal scene information and user scene information, and compares them; the relative relationship between the intent signal scene and the user scene is determined according to the capture error of the semantic feature extraction unit for the scene information, and a judgment is made to dispatch service personnel to the dialogue link for investigation; after the investigation, the service log is uploaded to the data storage unit, and the service log is distributed to the service feedback record data in the corresponding user dialogue basic data by the generative dialogue model. When judging, if the intent signal scene error is within the user scene range, the generative dialogue model retrieves the corresponding user basic data and exports it, and dispatches service personnel to the dialogue link for inspection; if the intent signal scene error is outside the user scene range, the generative dialogue model matches the intent signal scene according to the dialogue network, retrieves the user basic data in the corresponding network, analyzes the users directly affected by the dialogue network corresponding to the intent scene, analyzes the weight relationship of the corresponding users affected, and dispatches service personnel to check the users in descending order of weight. The specific implementation is as follows:

[0094] Taking the user consultation scene of an e-commerce platform as an example, when the user sends "How come the clothes I bought haven't arrived yet?", the semantic feature extraction unit first monitors the dialogue content in real time. Through natural language processing technology, the semantic feature extraction unit identifies keywords such as "clothes" and "haven't arrived", determines that it is an intent signal about logistics inquiry, and immediately captures the signal.

[0095] After capturing the intent signal, the signal is transmitted to the model processing module. The model processing module performs feature vectorization conversion, decomposes the sentence "How come the clothes I bought haven't arrived yet?" into a word sequence, and then converts each word into a corresponding digital feature vector through word embedding technology. For example, "clothes" may be mapped to a vector containing material, style, etc. dimensions, and "haven't arrived" may correspond to dimensions related to the logistics status.

[0096] The model processing module uses attention mechanism to focus on the key information "haven't arrived", as it directly reflects the user's core demand. At the same time, sequence modeling technology is used to analyze the grammatical structure and word order of the sentence, to understand the user's expression logic. In addition, through noise filtering technology, the interference of mood words such as "how" is removed, making the intent signal more pure. Then, using context fusion technology, the user's previous dialogue history is combined, such as the user previously asking "What is the order number?", to more comprehensively understand the user's current logistics inquiry intent.

[0097] The processed interaction intention signal is sent to the generative dialogue model. After receiving the signal, the generative dialogue model analyzes it and extracts the scene information of the intention signal, i.e., the logistics query scene. At the same time, the generative dialogue model retrieves user scene information from the data storage unit, including the user's order information, purchase time, delivery address, etc.

[0098] The generative dialogue model compares the intention signal scene information with the user scene information. Assuming that the clothes purchased by the user are expected to arrive within 3 days, and the user inquires about the logistics on the 2nd day, the scene information captured by the semantic feature extraction unit has a small error with the user's actual scene and is within the user's scene range. In this case, the generative dialogue model will retrieve the user's basic data such as user ID, contact information, etc., and export related information, and dispatch customer service personnel to the dialogue link for inspection to check whether the logistics information is updated, whether there is system delay, etc.

[0099] In another scenario, the user sends "What is the warranty policy for this product?", the semantic feature extraction unit captures the intention signal and transmits it to the model processing module. After a series of feature vectorization conversion and optimization processing, the model processing module sends the signal to the generative dialogue model. The generative dialogue model extracts the scene information of the intention signal as warranty policy consultation and finds that the user has purchased the product for more than half a year after the warranty period when retrieving the user scene information. At this time, the scene error of the intention signal is outside the user's scene range, the generative dialogue model matches to the "over-warranty product consultation" scene according to the dialogue network, retrieves the user's basic data in the corresponding network, and analyzes the user groups directly affected by the dialogue network corresponding to this warranty policy consultation scene, such as all over-warranty product users. Then, analyze the weight relationship of these users affected, for example, determine the weight according to the length of time the user has purchased the product, the amount of consumption, etc., and dispatch service personnel to check these users in descending order of weight, giving priority to high-weight user consultations.

[0100] After the service personnel go to the dialogue link for inspection, the inspection process and results form a service log, which is uploaded to the data storage unit. The generative dialogue model receives the service log and assigns it as service feedback record data in the corresponding user dialogue basic data. For example, in the logistics query scene, the service log records that the customer service personnel found that the logistics information update was delayed and has urged the express delivery for the user. This log will be assigned to the user's service feedback record for subsequent viewing and analysis.

[0101] In the financial service scenario, the user asks "how was my credit card frozen", and the semantic feature extraction unit captures this intent signal. After feature vectorization conversion and optimization processing by the model processing module, the generative dialogue model extracts the intent signal scenario information as a credit card freezing consultation. When retrieving user scenario information, it is found that the user has suspicious transaction records recently. At this time, the intent signal scenario error is within the user scenario range, the generative dialogue model retrieves the user's basic data and exports it, and dispatches customer service personnel to contact the user to verify the transaction. After investigation, the service log records that the user's card was frozen due to a large amount of transactions in a different place, and has been unfrozen for the user. This log is assigned to the user's service feedback record.

[0102] Through such an intent signal screening and verification process, the system can accurately determine the relationship between the intent signal and the user scenario, promptly dispatch service personnel for investigation and processing, and accurately archive service logs, providing reliable data support for subsequent service optimization and model training. The entire process is closely linked, ensuring accurate understanding and effective response of the intelligent dialogue system to user intent, improving user experience and service quality.

[0103] Embodiment 5:

[0104] The interaction actually includes that the generative dialogue model receives the intent signal, processes the intent signal using a dynamic adaptive sequence generation algorithm, extracts the scenario information of the intent signal, and marks the intent signal data in the coordinate system according to the intent signal scenario; uses the attention mechanism to analyze the intent signal scenario distribution data and the user dialogue basic data; checks the intent signal scenario, and when the signal error range of the semantic feature extraction unit is within the user scenario, it is considered as the user intent signal, otherwise it is considered as the dialogue network facility intent signal; the generative dialogue model analyzes the user interaction intent for the user intent signal and the dialogue network facility intent signal respectively; the generative dialogue model analyzes the dialogue network intent combining the user interaction intent, sets the interaction response priority according to the intent expression intensity, adjusts the model training parameters according to the intent frequency, optimizes the dialogue generation strategy combining the service feedback record in the user dialogue basic data, and generates warning prompt information for possible abnormal situations in local dialogue interaction. When analyzing the user interaction intent, for the user intent signal, the generative dialogue model retrieves the user intent signal, the interaction rule data in the dialogue network topology information, and the user dialogue basic data for user interaction intent analysis, and then dispatches personnel to adjust the user dialogue link; for the dialogue network facility intent signal, the generative dialogue model retrieves the dialogue network facility intent signal, dialogue network topology information, user geographic location, and user real-time interaction content for other facility intent analysis, and then dispatches personnel to investigate the intent scenario. The specific implementation is as follows:

[0105] Taking the processing of user inquiries by an intelligent customer service system as an example, when the user sends “the air conditioner I purchased cannot cool”, the generative dialogue model receives the intention signal. The model uses a dynamic adaptive sequence generation algorithm to analyze the sequence of keywords such as “air conditioner” and “cannot cool”, extracts the scene information of “device fault repair” by combining the patterns of similar problems in historical dialogues. Then, the system takes the customer service dialogue node where the semantic feature extraction unit is located as the origin, and marks the intention signal data as (home appliance repair scene, cooling failure type, user purchase record dimension) in the previously established scene coordinate system, where the x-axis corresponds to the process stage of fault repair, the y-axis corresponds to the device type, and the z-axis corresponds to the user purchase time and other information.

[0106] The generative dialogue model uses attention mechanism to associate the scene distribution data of the intention signal (such as the frequency and regional distribution of similar cooling failure inquiries in recent times) with the user dialogue basic data (the user's purchase time, historical repair record, whether to purchase an extended warranty, etc.). For example, if it is found that the user has purchased the air conditioner for only 3 months and has no historical repair record, the attention mechanism will focus on the possibility of “new machine failure”, improving the understanding accuracy of the user's specific needs.

[0107] When checking the intention signal scene, the system will evaluate the capture error of the semantic feature extraction unit. Assuming that the user actually feeds back “the air conditioner cooling effect is poor”, while the semantic feature extraction unit initially judges it as “cannot cool”, there is an error in the fault degree, but it still belongs to the user scene range of “cooling related problems”, at this time the signal is regarded as the user intention signal. On the contrary, if the user is actually asking “the installation cost of the air conditioner”, while the semantic feature extraction unit misjudges it as “fault repair”, the error exceeds the user scene range, and it will be regarded as a dialogue network facility intention signal (such as parameter abnormalities of the semantic recognition module).

[0108] For the user intention signal, the generative dialogue model retrieves the user intention signal, interaction rule data (such as the standard process of home appliance fault repair, priority rules) in the dialogue network topology information, and user dialogue basic data. For example, by combining the user's purchase credentials and geographic location (to determine whether it is within the scope of after-sales coverage), it analyzes the specific interaction intention as “emergency repair” and dispatches after-sales personnel to contact the user, arranges on-site repair, and automatically pushes the troubleshooting guide in the dialogue link for the user to temporarily refer to.

[0109] If it is determined that the dialogue network facility intention signal, such as the system finds that multiple users' "installation cost consultation" in the same period are misjudged as "fault repair", the generative dialogue model will call the dialogue network facility intention signal, dialogue network topology information (such as the deployment node of the semantic recognition module), user geographic location (such as the semantic recognition server in a certain area is overloaded) and user real-time interaction content, analyze that it is possible that the semantic processing server in a certain area is abnormal, causing recognition errors. At this time, the system will send technical personnel to troubleshoot the server node, and temporarily adjust the load distribution of other nodes to ensure the stability of the dialogue service.

[0110] In the dialogue network intention analysis link, the generative dialogue model will set the interaction response priority according to the intention expression intensity. For example, the user sends "air conditioner completely not cooling" three times in a row, the intention expression intensity is high, and the system will raise its response priority to the highest, and arrange the after-sales personnel to handle it in priority. According to the intention frequency, adjust the model training parameters, if the consultation quantity of "air conditioner not cooling" in a certain period suddenly increases by 20%, the model will automatically increase the training sample weight of this scene, and optimize the related semantic recognition parameters. Combined with the service feedback records in the user dialogue basic data, such as the user's satisfaction evaluation on "on-site speed" in the past similar faults, optimize the dialogue generation strategy, and actively inform the user in the reply that "the maintenance personnel will contact you within 2 hours", and improve the user experience.

[0111] For abnormal situations that may occur in local dialogue interaction, the system will generate warning prompt information. For example, when analyzing the dialogue data in a certain area, it is found that the description frequency of "compressor abnormal sound" in the "air conditioner fault repair" intention of the users in this area is 30% higher than usual, and multiple users feedback that "after-sales response is slow", the generative dialogue model will judge that there may be batch equipment quality problems or insufficient after-sales resources in this area, and then generate a warning prompt to suggest that maintenance personnel from surrounding areas be deployed to support, and actively contact the users in this area to explain the situation and make service arrangements in advance.

[0112] In the intelligent recommendation scenario of the e-commerce platform, the user sends "want to buy thermal underwear for parents, do you have recommendations?", and the generative dialogue model receives the intention signal, extracts the "recommendation of thermal underwear for the middle-aged and elderly" scene information, and marks it as (clothing recommendation scene, middle-aged and elderly group dimension, seasonal factor dimension) in the coordinate system. By analyzing the user's historical purchase records (once purchased health products for the middle-aged and elderly) and the current dialogue scene through the attention mechanism, it is determined that the user has high demand for warmth and comfort. After checking the scene error, it is considered as the user's intention signal, and the recommendation rules (such as high-score goods first, brand preference rules) in the dialogue network topology information are called, combined with the parent age and regional climate (such as cold northern regions) in the user's basic data, to generate a personalized recommendation list, and a customer service personnel is dispatched to introduce the details of the goods. If the system detects that multiple users frequently ask "washing method" in the recommendation link, and the current recommendation reply does not contain this information, the generative dialogue model will adjust the training parameters of the recommendation dialogue according to the frequency of the intention, and automatically add the washing instructions in the subsequent recommendations, optimizing the dialogue generation strategy.

[0113] The entire interactive actual execution process realizes the closed-loop management of the intelligent dialogue system through dynamically processing the intention signal, accurately distinguishing the user and facility intentions, and optimizing the response strategy combined with multi-dimensional data. From the capture of user needs to the scene-based response, and to the optimization iteration of the system itself, each link closely relies on data support and model analysis to ensure the efficiency and accuracy of the dialogue interaction. For example, in the medical consultation scenario, the user describes "cough with chest tightness", and the system extracts the "respiratory system inquiry" scene through the interactive execution process, combines the user's geographical location (epidemic risk area) and historical health data, and prioritizes video consultation, and updates the disease judgment model according to the consultation frequency of similar symptoms, to realize accurate medical consultation service response.

[0114] It should be noted that, in this document, the terms such as first and second are used merely to distinguish one entity or action from another, and do not necessarily require or imply that these entities or actions occur in any particular temporal or spatial order. Also, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0115] Although embodiments of the present application have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions, and alterations can be made hereto without departing from the principles and spirit of the application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. An intelligent dialogue system based on a generative model, characterized in that: Specifically include: Semantic feature extraction unit: used to detect the intent of user conversation content in an online state, and realize multi-round and context-related detection; Service terminal: The service terminal is equipped with a generative dialogue model, a model processing module, and a data storage unit, and receives the interaction data from the semantic feature extraction unit; The generative dialogue model is used to obtain historical dialogue records as a basic reference for model training, analyze the scenarios and frequencies of interaction intentions, and analyze the specific content of the intention expression in combination with the dialogue goals in the corresponding scenarios; the model processing module is used to convert natural language signals into digital features through feature vectorization conversion, and optimize intention understanding through attention mechanism, sequence modeling, noise filtering and context fusion; the data storage unit is used to store basic data of each user dialogue, interaction data received by the semantic feature extraction unit, intention processing records, user portrait information data, and dialogue network topology data; Interaction feedback unit: The interaction feedback unit is used to display the specific scenario of the intention expression and intuitively observe the user needs associated with the corresponding dialogue interaction.

2. The generative model-based intelligent dialogue system according to claim 1, characterized in that: The semantic feature extraction unit captures the interaction intention signal and then sends the intention signal to the model processing module for data processing. The processed intention signal data is transmitted to the generative dialogue model. The generative dialogue model analyzes the intention scenario and intention intensity based on the user dialogue basic data, dialogue network topology information, and historical interaction data, and sends the analysis results to the interaction feedback unit for display.

3. The generative model-based intelligent dialogue system according to claim 1, characterized in that: The data storage unit receives basic user conversation data, and when the semantic feature extraction unit receives the intention signal, copies the intention signal data and sends it to the data storage unit for storage backup. After the generative conversation model completes processing the intention signal data, it sends the processing records and results to the data storage unit for storage backup.

4. An intelligent dialogue method based on a generative model, according to the intelligent dialogue system based on a generative model according to any one of claims 1-3, characterized in that: The specific steps include: S1. Collect basic data and build a generative dialogue model; S2. Perform intent signal screening and verification, and process, analyze, screen, and determine the captured intent signals; S3. Perform actual interaction execution, analyze intent signals, and adjust the interaction response strength based on the intensity of intent expression, frequency of intent occurrence, and basic user conversation data.

5. The generative model-based intelligent dialogue method according to claim 1, characterized in that: The step S1 comprises: S1.

1. Determine the deployment location of the semantic feature extraction unit based on the interaction habits of each user of the dialogue service platform; S1.

2. Collect basic data about the conversation network within the scope of the generative conversation model, including user location, interaction preferences, usage duration, historical conversation content, and service feedback records, as well as conversation network topology information and a global view of the conversation service platform. Store the collected basic data in the data storage unit. S1.

3. Build a generative dialogue model based on a dynamic adaptive sequence generation algorithm. Combine user dialogue basic data, dialogue network topology information, and user profile information, and incorporate the attention mechanism into the generative dialogue model. S1.

4. Obtain the intent signals detected by the operational semantic feature extraction unit as a training set for the generative dialogue model, and use the training set to train the generative dialogue model. S1.

5. The generative dialogue model receives basic user dialogue data and establishes a scene coordinate system based on the user's dialogue network. The origin of the coordinate system is the location of the semantic feature extraction unit. S1.

6. Scale the global view of the conversation service platform and merge it into the coordinate system. Verify the accuracy of the conversation paths in the conversation network and the corresponding positions in the view. S1.

7. When the intention signal is obtained, the signal is copied and sent to the model processing module and the data storage unit respectively. When the generative dialogue model obtains the intention signal analysis result, the result is copied and sent to the interactive feedback unit and the data storage unit respectively.

6. The generative model-based intelligent dialogue method according to claim 4, characterized in that: The intention signal screening and verification specifically includes the following steps: S2.1, the semantic feature extraction unit captures the intention signal; S2.

2. The model processing module then converts the natural language signal into digital features through feature vectorization. It optimizes intent understanding through attention mechanisms, sequence modeling, noise filtering, and context fusion. The processed interaction intent signal is then sent to the generative dialogue model. S2.

3. The generative dialogue model analyzes the interaction intent signal, extracts the intent signal scenario information and the user scenario information, and compares them. S2.

4. Determine the relative relationship between the intent signal scenario and the user scenario based on the semantic feature extraction unit's capture error of the scenario information, make a judgment, and then dispatch service personnel to the conversation link for investigation. S2.

5. After troubleshooting, the service log is uploaded to the data storage unit, and the generative dialogue model assigns the service log as service feedback record data in the corresponding user dialogue basic data.

7. The generative model-based intelligent dialogue method according to claim 4, characterized in that: The actual execution of the interaction specifically includes the following steps: S3.

1. The generative dialogue model receives the intent signal, processes it using a dynamic adaptive sequence generation algorithm, extracts the contextual information of the intent signal, and labels the intent signal data in a coordinate system according to the context of the intent signal. S3.

2. Use the attention mechanism to analyze the intent signal scenario distribution data and user conversation basic data; S3.

3. Verify the intent signal scenario. When the signal error range of the semantic feature extraction unit is within the user scenario, it is considered a user intent signal. Otherwise, it is considered a dialogue network facility intent signal. S3.4, the generative dialogue model analyzes user interaction intentions based on user intention signals and dialogue network facility intention signals respectively; S3.

5. The generative dialogue model combines user interaction intentions to perform dialogue network intention analysis, and issues early warnings for abnormal situations in local dialogue interactions based on the intensity of intent expression, frequency of intent occurrence, and basic user dialogue data.

8. The generative model-based intelligent dialogue method according to claim 6, characterized in that: In step S2.4, the specific method of determination is: 1) If the intent signal scenario error is within the user scenario range, the generative dialogue model retrieves and exports the corresponding user's basic data, and dispatches service personnel to the dialogue link for inspection; 2) If the intent signal scenario error is outside the user scenario range, the generative dialogue model matches the intent signal scenario based on the dialogue network, retrieves the basic user data in the corresponding network, analyzes the users directly affected by the dialogue network corresponding to the intent scenario, analyzes the weight relationship of the corresponding users affected, and dispatches service personnel to check the users in descending order of weight.

9. The generative model-based intelligent dialogue method according to claim 7, characterized in that: In step S3.4, the specific processing method of user interaction intention analysis is: 1) The generative dialogue model retrieves user intent signals, interaction rule data from the dialogue network topology information, and basic user dialogue data to analyze user interaction intent, and then dispatches personnel to adjust the user dialogue process; 2) The generative dialogue model retrieves the dialogue network facility intent signal, dialogue network topology information, user geographic location and user real-time interaction content to perform other facility intent analysis, and then dispatches personnel to investigate the intent scenario.

10. The generative model-based intelligent dialogue method according to claim 7, characterized in that: In step S3.5, the specific processing method of the dialogue network intent analysis is as follows: the generative dialogue model sets the interaction response priority according to the intensity of the intent expression, adjusts the model training parameters according to the frequency of the intent, optimizes the dialogue generation strategy based on the service feedback records in the user dialogue basic data, and generates early warning prompt information for abnormal situations that may occur in local dialogue interactions.

Images