Man-machine conversation system and method based on generative artificial intelligence

Abstract

The invention discloses a man-machine dialogue system and method based on generative artificial intelligence, and the system can achieve natural and smooth man-machine interaction, and has the functions of context understanding ability, dynamic response generation, multi-round dialogue control and personalized sentiment analysis. Through cooperative operation of all the modules, the system is suitable for various industry scenes, and the conversation intelligence and the user experience are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence and natural language processing, in particular to a human-computer conversation system and method based on generative artificial intelligence. BACKGROUND

[0002] With the rapid development of artificial intelligence, human-computer conversation technology is widely used in customer service, intelligent assistants and online consultation scenarios. However, the existing human-computer conversation system has certain deficiencies in processing multi-turn conversations, maintaining context consistency, response generation flexibility and emotion recognition. Traditional rule or template driven systems are difficult to achieve highly personalized and emotional conversation experience. Therefore, a human-computer conversation system and method based on generative artificial intelligence is proposed, which further improves user interaction experience and conversation quality through emotion analysis and personalized response module. SUMMARY

[0003] The present application provides a human-computer conversation system and method based on generative artificial intelligence, aiming to achieve efficient, flexible and natural human-computer interaction through modular design. The system includes data processing module, language understanding module, response generation module, dialogue management module, user feedback module, emotion analysis module and personalized customization module, each module cooperates through specific algorithms and technical means to realize semantic understanding, context management, emotion analysis and personalized interaction.

[0004] Data processing module The data processing module is mainly used for receiving and preprocessing user input natural language data. The module includes two sub-modules of text cleaning and feature extraction, as follows: Text cleaning algorithm: through regular expression cleaning of user input text, remove irrelevant characters, punctuation marks and redundant spaces and other noise information, to ensure the quality of input data.

[0005] Feature extraction: TF-IDF (Term Frequency-Inverse Document Frequency) algorithm and word embedding (such as Word2Vec, GloVe) technology are used to convert the processed text into high-dimensional vector representation. These feature vectors will be input to the language understanding module for further processing.

[0006] The text cleaning and feature extraction steps of this module ensure the effectiveness and uniformity of the input data, laying a foundation for subsequent intent analysis.

[0007] Language understanding module The language understanding module is responsible for semantic parsing of preprocessed user input data to identify user intent and extract key information. This module is based on deep learning models or pre-trained language models, such as BERT (Bidirectional Encoder Representations from Transformers) and GPT (Generative Pre-trained Transformer), and its specific functions are as follows: Intent Classification: Utilizing deep learning models, BERT, and GPT pre-trained models, and fine-tuning them for specific domains through transfer learning, high-precision intent recognition is achieved. For example, if a user inputs "Book a meeting for tomorrow," the system can recognize the specific intent of "book a meeting."

[0008] Slot filling: Employs intent classification and slot filling algorithms to extract specific information from user input, such as location and time, to ensure accurate information delivery in multi-turn conversations.

[0009] Through the above technical means, the language understanding module can comprehensively analyze the intent and key information in the user input, providing the necessary semantic data for the response generation module.

[0010] Response generation module The response generation module is used to generate the system's natural language response. Its specific functions and algorithms are as follows: Generative models: Generative models based on the Transformer architecture (GPT-3) or Encoder-Decoder structure models generate natural language responses. These models automatically generate logical and natural-sounding answers based on user intent and context.

[0011] Template matching and generation combination: Before generating a response, the response generation module first matches a preset template to ensure that the generated response conforms to the application scenario and business requirements. If a template is matched, the template response is used first; otherwise, a dynamic response is generated using the generative model.

[0012] This module combines template matching and generative models to ensure the diversity and accuracy of system responses, providing users with a flexible and natural interactive experience.

[0013] Dialogue Management Module The dialogue management module is mainly used to manage contextual information in multi-turn dialogues, specifically including: Context tracking mechanism: Using hash tables or graph structures to store dialogue history information enables the system to maintain a coherent understanding of semantics across multiple rounds of dialogue. For example, when a user mentions "yesterday's activities" in a multi-turn dialogue, the system can understand that this refers to the specific activities discussed in previous conversations.

[0014] Dialogue State Manager: Dynamically manages dialogue states using a Finite State Machine (FSM). The FSM dynamically adjusts the dialogue flow based on the user's input intent state and resets or reconstructs the context when necessary to ensure the coherence and rationality of multi-turn dialogues.

[0015] This module's context tracking and state management functions enable the system to continuously track and maintain the dialogue state, improving dialogue quality and user experience.

[0016] User feedback module The user feedback module is used to collect user interaction feedback and optimize the generated model. The specific implementation method is as follows: User feedback collection: The system collects user ratings and comments on the quality of the conversation through an interactive interface and stores them in a feedback database for later use.

[0017] Model update strategy: Employ a policy optimization algorithm (Q-learning or PPO) based on reinforcement learning to optimize the parameters of the generated model using user feedback data, thereby gradually improving the system's response accuracy and user satisfaction.

[0018] By continuously collecting user feedback and optimizing model parameters, the user feedback module has significantly improved the accuracy of system dialogue generation and user experience.

[0019] Sentiment Analysis Module The sentiment analysis module is used to identify emotional information in user input in order to optimize the emotional expression of the conversation, specifically including: Sentiment analysis models: Based on LSTM-CNN hybrid models or sentiment recognition algorithms such as RoBERTa, these models classify and analyze the sentiment in user-input text. The model combines emotional vocabulary and contextual information to determine whether the user's emotion is positive, neutral, or negative.

[0020] Emotional Tagging Classification: By using emotional categorization tags and weighted methods, the system determines the overall mood and adjusts the tone and content of its responses accordingly to achieve more emotionally resonant interactions. For example, when user dissatisfaction is detected, the system automatically generates a soothing response.

[0021] The sentiment analysis module ensures that the system can effectively identify and respond to users' emotions, thereby improving the system's user-friendliness.

[0022] Personalized customization module The personalization module generates personalized responses based on user preferences and historical interaction records, specifically including: User profile generation: The K-means clustering algorithm is used to classify user interaction behaviors and construct user profiles that include features such as interests, preferences, and frequently used words, so as to realize personalized content recommendations.

[0023] Personalized recommendations: Combining collaborative filtering and content recommendation algorithms to generate suggestions that match user interests and preferences, or directly adjusting the response strategy of the generation model based on user preferences to enhance the customization of the user experience.

[0024] The personalized customization module provides users with a personalized interactive experience through user profiling and recommendation algorithms, effectively improving user stickiness and satisfaction.

[0025] Beneficial effects 1. Improve the accuracy and naturalness of dialogue: Through intent classification and slot filling technology in the language understanding module, the system can accurately understand the core intent of user input, ensuring that the dialogue content is accurate and meets user needs. Meanwhile, the response generation module combines generative models and template matching to enhance the flexibility and naturalness of the dialogue.

[0026] 2. Enhanced User Experience: The sentiment analysis module can identify users' emotional states and adjust the tone and expression of the system's responses, making the system's replies more human-like. The personalization module provides personalized recommendations based on user profiles, enabling the system to offer responses that are closer to the needs of different users, significantly improving user satisfaction and interactive experience.

[0027] 3. Ensuring Coherence in Multi-Turn Dialogues: The dialogue management module ensures the continuity and logical consistency of the system in multi-turn dialogues through context tracking and dialogue state management. The system can dynamically adjust the dialogue state based on the context information, resulting in a smoother and more seamless user experience.

[0028] 4. Adaptive System Optimization: By collecting user feedback through a user feedback module and continuously optimizing the model using reinforcement learning algorithms, the system can adaptively adjust, improving its intelligence and learning capabilities. This closed-loop feedback mechanism allows the system to continuously learn and improve, enhancing the accuracy of dialogue and the quality of response.

[0029] 5. Modular design, flexible and scalable: Each functional module is implemented independently yet collaborates with each other, facilitating system customization and expansion in different application scenarios. For example, sentiment analysis and personalized customization modules can be enabled or disabled based on specific needs, making the system architecture more flexible and adaptable. Attached Figure Description

[0030] Figure 1 It is a data processing module; Figure 2 It is the language understanding module; Figure 3 It is a response generation module; Figure 4 It is the dialogue management module; Figure 5 It is the user feedback module; Figure 6 It is the sentiment analysis module; Figure 7 It is a personalized customization module. Detailed Implementation

[0031] The embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0032] Example: 1. Data Processing Module The main task of the data processing module is to receive and preprocess the natural language data input by the user, providing high-quality data input for the subsequent language understanding module. The specific implementation includes the following aspects: Text Cleaning Submodule: This submodule uses regular expressions to clean the data, removing noise characters, punctuation marks, and extra spaces from the text. For example, user input may contain redundant punctuation marks (such as "Help me! Check it out."). This module automatically removes irrelevant characters while maintaining the validity of the content, ensuring that the parsing of the input text is not interfered with by impurities.

[0033] Feature extraction submodule: After text cleaning, the TF-IDF algorithm or word embedding models (such as Word2Vec and GloVe) are used to transform the text into vectorized feature representations. These vectors can better preserve the semantic information in the text, thus providing a more comprehensive data representation for the language understanding module. For example, for semantically similar phrases "check the weather" and "weather conditions," this submodule will map them to similar feature spaces through vectorization, thereby reducing understanding errors.

[0034] 2. Language Comprehension Module The language understanding module is primarily responsible for semantic parsing of preprocessed text data to identify user intent and extract key information. This module employs deep learning models, such as BERT and GPT, combined with intent classification and slot filling techniques to complete the semantic understanding task.

[0035] Intent classification: Using deep learning models such as BERT or GPT, fine-tuned in conjunction with domain-relevant corpora, to accurately identify users' specific needs. For example, if a user enters "book a meeting room for tomorrow," the system will identify the specific intent of "book a meeting" based on a pre-trained model, thereby understanding the user's actual needs.

[0036] Slot filling: To enhance the effectiveness of multi-turn dialogues, the language understanding module also extracts key information (such as time and location) through a slot filling mechanism. For example, in the sentence "Book a meeting room for 10:00 AM tomorrow", the system will automatically use "10:00 AM tomorrow" as the time slot and "meeting room" as the location slot, and pass this information to the dialogue management module for subsequent operations.

[0037] 3. Response Generation Module The response generation module is responsible for generating the system's natural language responses to users, aiming to provide answers that meet user needs and are semantically fluent. This module combines generative models and template matching to ensure the flexibility and accuracy of the response content.

[0038] Generative Model: Based on the Transformer architecture, this generative model (GPT-3) generates natural language responses that fit the context after receiving user intent and contextual information. For example, when a user asks, "Have you booked the meeting room?", the system will automatically generate a response based on the context: "Your meeting room booking has been completed, and the meeting time is tomorrow at 10:00 AM." Template matching and generative combination: To avoid randomness in the generated model, the system presets some commonly used dialogue templates. When a user's request matches a template, the system prioritizes using the template to generate a response; if no matching template is found, the generative model is invoked to generate a reply. For example, if the system detects that a user's inquiry pertains to "confirm booking," it directly matches the confirmation reply template, improving the efficiency and accuracy of the response.

[0039] 4. Dialogue Management Module The dialogue management module is used to maintain the context information and state of dialogues to ensure the coherence and effectiveness of multi-turn dialogues. The implementation of this module includes a context tracking mechanism and a dialogue state manager.

[0040] Context tracking mechanism: This mechanism stores dialogue history information through hash tables or graph structures and tracks the context of users in multi-turn conversations. For example, in a multi-turn conversation, users may use pronouns such as "that" or "this" to refer to previously mentioned items. Through context tracking, this mechanism associates the actual meaning of the pronouns with previous semantic information, ensuring an accurate understanding of the user's pronoun references.

[0041] Dialogue State Manager: Manages changes in dialogue state based on a Finite State Machine (FSM). The FSM controls the dialogue flow through different state transitions, ensuring the system dynamically adapts to changes in user needs. For example, when a user's need changes from "checking the weather" to "booking a meeting," the FSM automatically resets the current state and adjusts the system's interaction logic according to the new intent, ensuring a smooth dialogue flow.

[0042] 5. User Feedback Module The user feedback module continuously optimizes the system's generation model by collecting user feedback, enabling the system to better adapt to user needs. The specific implementation includes the following steps: User feedback collection: The system collects feedback from users through the interface, such as user ratings and opinions on the dialogue content, and stores this data in the feedback database.

[0043] Model update strategy: Based on reinforcement learning's Q-learning and PPO algorithms, the system's generative model is optimized. By using user feedback as a reward or penalty signal, the system continuously adjusts model parameters to improve response quality. For example, when a user gives a low rating to a response, the system will reduce the probability of generating such answers in subsequent learning processes, thus improving response accuracy.

[0044] 6. Sentiment Analysis Module The sentiment analysis module is used to identify emotional information in the user's input text and adjust the tone of the system's response accordingly to achieve emotional interaction.

[0045] Sentiment analysis model: This model uses an LSTM-CNN hybrid model or the RoBERTa algorithm to identify emotional information in text. It combines emotional vocabulary and contextual information to determine the user's emotional attribute (e.g., positive, negative, neutral), allowing the system to adjust its response accordingly.

[0046] Emotional Tag Classification: By using emotional tags (such as positive, neutral, and negative) and weighting methods, the system can automatically adjust the tone of responses. For example, if the system detects that a user's emotion is negative, the generated response will be more soothing to improve the user experience.

[0047] 7. Personalized Customization Module The personalization module provides a personalized conversational experience based on user preferences and history. Specific implementations include the following functions: User profile generation: User behavior data is analyzed using a clustering algorithm (K-means) to generate user profiles, including user interests, preferences, and frequently used phrases.

[0048] Personalized recommendations: Based on user profiles, the system combines collaborative filtering and content recommendation algorithms to generate responses that match user preferences. For example, if a user prefers a particular style of expression in a specific context, the system will retrieve similar expressions to improve user satisfaction and the relevance of the system's interaction.

[0049] As the system's entry point, the data processing module first cleans and extracts features from the user-input text to ensure the standardization and validity of the input data, providing high-quality input for subsequent semantic analysis.

[0050] The language understanding module receives standardized text from the data processing module, understands the user's needs through intent classification and slot filling, and then passes the parsed intent and slot information to the dialogue management module. This module is highly coupled with the data processing module to ensure the complete transmission of semantic information from the text.

[0051] The dialogue management module acts as the "central hub" of the system, responsible for managing the context and state of the dialogue. Through a context tracking mechanism and a finite state machine (FSM), this module maintains the coherence of multi-turn dialogues and adjusts the interaction logic according to the user's dynamic needs, ensuring that each user input receives appropriate contextual processing.

[0052] The sentiment analysis and personalization modules provide intelligent and human-like support during the conversation. The sentiment analysis module identifies the user's emotional state and adjusts the tone of response, making the conversation more approachable. The personalization module utilizes the user's historical behavior and interests to generate personalized responses for the system, further enhancing the user experience.

[0053] The response generation module generates the final system response. Through a combination of generative models and template matching, it ensures both flexibility and accuracy and coherence of the content. The response generation module needs to comprehensively consider the status of the dialogue management module, the results of sentiment analysis, and the needs of personalized customization to provide the highest quality response content.

[0054] The user feedback module collects user feedback data at the final stage of the system process and uses reinforcement learning algorithms to continuously optimize the language understanding and response generation models. Through closed-loop adjustments based on feedback, the system continuously improves the accuracy and naturalness of the dialogue as users continue to use it.

[0055] The modules are interdependent and progressively layered. The data processing and language understanding module is responsible for receiving front-end information and semantic parsing; the dialogue management module ensures process coherence and logical consistency; the response generation module provides content output; and the sentiment analysis, personalization, and user feedback modules further enhance the emotional and human aspects of the dialogue. In the overall architecture, the user feedback module drives continuous system improvement to enhance the user's interactive experience, forming a closed-loop interaction process.

[0056] The human-computer interaction method of this system is carried out according to the following steps: Obtain natural language information from user input The system receives natural language input from the user, such as text information like questions and requests. This input serves as the initial data for the entire dialogue process.

[0057] Data preprocessing The data processing module preprocesses the user-input text data, including text cleaning, stop word removal, word segmentation, and feature extraction. It uses techniques such as TF-IDF or Word2Vec to convert the text into a feature vector format usable by the model.

[0058] Analyzing user intent The preprocessed text is fed into the language understanding module. This module uses deep learning models such as BERT and GPT for intent recognition and performs slot filling (e.g., extracting key information such as time and location). This step ensures that the system can identify the user's core needs or questions, preparing for response generation.

[0059] Generate natural language responses The response generation module generates the system's response content based on the parsed user intent.

[0060] If the input matches a preset template, the response generation module will prioritize template-based generation; otherwise, it will generate a natural language response using a generative model (such as GPT).

[0061] Update the context information of the conversation The dialogue management module updates the context information of the current dialogue. The module records key information from this dialogue and updates the dialogue state through a state management mechanism (such as a finite state machine). This step ensures that the system can adjust subsequent responses based on dialogue history information, thereby achieving coherence in multi-turn dialogues.

[0062] Analyzing user sentiment The sentiment analysis module performs emotion recognition on the user-input text, determines the emotion type (such as positive, negative, neutral, etc.), and analyzes the user's current emotional state. Based on the sentiment analysis results, the system can adjust the tone of its response to provide a response that matches the user's emotional state.

[0063] Generate personalized responses The personalization module provides personalized services based on user profile data and sentiment analysis results. Collaborative filtering or clustering algorithms (such as K-means) are used to analyze user preferences, ensuring that the generated content matches user interests and personalized needs.

[0064] Receive user feedback and update the generated model The system collects user feedback through a user feedback module, including satisfaction levels or specific evaluations of the system's response. This collected feedback is used for model optimization; the system employs reinforcement learning algorithms (such as Q-learning or PPO) to adaptively optimize the generative model, thereby improving response quality.

Claims

1. A human-computer dialogue system based on generative artificial intelligence, characterized in that, include: The data processing module is used to receive and preprocess natural language data input by the user; The language understanding module is used to perform semantic parsing on preprocessed natural language data; The response generation module is used to generate corresponding responses based on the parsed semantic content; The dialogue management module is used to manage and maintain the context information of dialogues; The user feedback module is used to receive user feedback information and iteratively optimize the generative model. The sentiment analysis module is used to identify emotional information in user input; The personalization module is used to generate personalized responses based on user preferences and historical interaction records.

2. The system according to claim 1, characterized in that, The data processing module includes text cleaning and feature extraction sub-modules for handling noise and unstructured features in natural language input, specifically including: Text cleaning algorithms: including regular expression cleaning algorithms to remove irrelevant characters, punctuation marks, and spaces. Feature extraction: using techniques such as TF-IDF and word embeddings (Word2Vec, GloVe), the text is transformed into high-dimensional vectors for use as input to the language understanding module.

3. The system according to claim 1, characterized in that, include, The language understanding module, based on deep learning models or pre-trained language models, is used to accurately identify user intent, specifically including: BERT and GPT pre-trained language models are fine-tuned for specific domains through transfer learning, improving the accuracy of understanding domain-specific terminology or special expressions. Intent classification and slot filling algorithms are used to identify user intent and extract important information (location, time, to support effective communication in multi-turn conversations).

4. The system according to claim 1, characterized in that, This includes a response generation module that generates natural language responses based on a generative model, and can be optimized through template matching or language model generation. Specifically, it includes: Transformer generative models or encoder-decoder-based sequence generation models are used to generate natural language responses in real time. A template matching and generation combination algorithm is employed: before the system generates a response, a specific template is preferentially used if matched to ensure the accuracy and consistency of the generated response.

5. The system according to claim 1, characterized in that, include, The dialogue management module includes a context tracking mechanism for dynamic updates of information during multi-turn dialogues, specifically including: Context tracking mechanism: Stores dialogue history using hash tables or graph structures to maintain semantic coherence across multiple rounds of dialogue. Dialogue state manager: Manages dialogue state based on a finite state machine (FSM) and performs context reset or reconstruction when needed to ensure smooth multi-round dialogue.

6. The system according to claim 1, characterized in that, This includes a user feedback module containing data update and model retraining sub-modules to improve the model based on user feedback, specifically including: User feedback collection: User ratings and comments on the quality of the dialogue are collected through the interactive interface and stored in the feedback database. Model update strategy: Based on a Reinforcement Learning-based strategy optimization algorithm, the parameters of the generated model are adjusted according to user feedback data.

7. The system according to claim 1, characterized in that, This includes a sentiment analysis module that extracts emotional features from user input based on sentiment recognition algorithms, thereby optimizing the dialogue system's response. Specifically, this includes: a sentiment analysis model based on an LSTM-CNN hybrid model or RoBERTa, analyzing emotions through dialogue content and interjections; and sentiment label classification using sentiment classification labels (e.g., positive, neutral, negative) combined with a weighted approach to determine the overall sentiment and adjust the tone and content of the system's responses accordingly.

8. The system according to claim 1, characterized in that, This includes a personalized customization module that adjusts response generation strategies based on user history and conversation records to provide a personalized interactive experience. Specifically, this includes: User profile generation: Classifying user interactions using a K-means clustering algorithm and constructing user profiles that include features such as interests, preferences, and frequently used phrases. Personalized recommendation: Combining collaborative filtering and content recommendation algorithms to generate responses and suggestions based on user preferences to improve user satisfaction.

9. A human-computer dialogue method based on generative artificial intelligence, comprising: Obtain natural language information from user input; Preprocess user input; The language understanding module is used to analyze user intent. The response generation module generates the corresponding natural language response. Update the context information of the dialogue using the dialogue management module; The sentiment analysis module is used to analyze user emotions; The personalization module generates personalized responses that match user preferences. Receive user feedback and update the generated model.

Images