Intelligent dialogue system trained with massive multi-structured dialogue data

Abstract

The present invention provides an intelligent dialogue system trained with massive multi-structured dialogue data, belonging to the technical field of intelligent dialogue systems. The intelligent dialogue system trained with massive multi-structured dialogue data characterized in multiple languages ​​uses dialogue data with multiple structural features and forms for model training, automatically extracts the logical relationship between dialogues and the use of words, phrases and sentences through a large model, and forms knowledge stored inside the large model. Through natural language processing, speech recognition and speech synthesis technology, a real-time and intelligent dialogue experience is provided to users based on the large module in the above multi-structured dialogue data training. Data is collected according to the characteristics of multi-round intelligent dialogues, and the dialogues are optimized and trained through a large model to realize a more logical intelligent dialogue question-and-answer system based on a large model.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent dialogue systems, and in particular to an intelligent dialogue system trained with massive multi-structure dialogue data. Background Art

[0002] Intelligent dialogue modularization is a concept that combines artificial intelligence technology and modular design concepts. It aims to divide complex intelligent dialogue systems into multiple manageable and replaceable modules to improve the scalability, maintainability and flexibility of the system. With the rapid development of artificial intelligence technology, intelligent dialogue systems have been widely used in various fields, such as smart homes, smart customer service, and educational tutoring. However, these systems often have complex structures and functions, which brings challenges to development and maintenance. Modular design is widely used in software development as an effective solution to improve the manageability and scalability of the system.

[0003] The intelligent dialogue system in the existing technology lacks multi-structure dialogue data training, and the logical thinking of the intelligent dialogue system in answering questions is not rigorous enough. Summary of the invention

[0004] In order to make up for the above shortcomings, the present invention provides an intelligent dialogue system trained with massive multi-structured dialogue data, aiming to improve the problems in the prior art that the intelligent dialogue system lacks multi-structured dialogue data training and the intelligent dialogue system's logical thinking in answering questions is not rigorous enough.

[0005] The present invention is achieved in that: The present invention provides an intelligent dialogue system trained with massive multi-structured dialogue data represented in multiple languages, including Multi-structured dialogue data training uses dialogue data with multiple structural features and forms to train models, and automatically extracts the logical relationship between dialogues and the use of words, phrases and sentences through a large model to form knowledge stored inside the large model; intelligent dialogue module uses natural language processing, speech recognition and speech synthesis technology to provide users with real-time and intelligent dialogue experience based on the large module in the multi-structured dialogue data training; training optimization collects data based on the characteristics of multi-round intelligent dialogues, and optimizes the dialogue training through a large model. The intelligent dialogue system trained with massive multi-structured dialogue data represented in multiple languages ​​is an advanced dialogue system that combines multi-language support technology and training with massive multi-structured dialogue data to achieve cross-language intelligent dialogue. That is, the logical relationship between dialogues and the use of words, phrases and sentences are automatically extracted through a large model to form knowledge stored inside the large model. Through natural language processing, speech recognition and speech synthesis technology, based on the large module in the multi-structured dialogue data training, it provides users with real-time and intelligent dialogue experience. Finally, data is collected based on the characteristics of multi-round intelligent dialogues, and the dialogue is optimized and trained through a large model to achieve a more logical intelligent dialogue question-and-answer system based on a large model. Massive multi-structured dialogue data systems are used for training large amounts of diverse dialogue data. These data may come from different fields, scenarios, and languages, and have different structures and formats. The data may include text dialogues, voice dialogues, image descriptions, etc. These diverse data provide a rich training resource for the system, which helps to improve the system's dialogue understanding and generation capabilities. An intelligent dialogue system is an interactive system based on natural language processing technology that can have coherent and natural conversations with humans. The system usually includes modules such as speech recognition, natural language understanding, dialogue management, and speech synthesis.

[0006] Preferably, the multi-structure dialogue data training also includes model architecture selection data, hyperparameter tuning data, and training process stability and efficiency data.

[0007] Preferably, the architecture selection data includes data diversity, data balance and data annotation quality.

[0008] Preferably, the training optimization utilizes contextual information to enhance the model's understanding capabilities and continuously optimizes model performance through methods such as incremental training or transfer learning.

[0009] Preferably, the training optimization uses attention mechanism and visualization technology to improve the interpretability of the model.

[0010] Preferably, the natural language processing technology is responsible for understanding and analyzing the user's voice instructions and extracting key information.

[0011] Preferably, the speech recognition technology converts the user's speech input into text, providing a basis for subsequent dialogue processing.

[0012] Preferably, the speech synthesis technology converts the computer-generated text responses into speech, allowing the user to hear fluent and natural speech output.

[0013] The beneficial effects of the present invention are as follows: the intelligent dialogue system trained with massive multi-structured dialogue data obtained by the above-mentioned design and the intelligent dialogue system trained with massive multi-structured dialogue data represented in multiple languages ​​are advanced dialogue systems, which combine multi-language support technology and training with massive multi-structured dialogue data to achieve cross-language intelligent dialogue. That is, the logical relationship between dialogues and the use of words, phrases and sentences are automatically extracted through a large model to form knowledge stored inside the large model. Through natural language processing, speech recognition and speech synthesis technology, a real-time and intelligent dialogue experience is provided to users based on the large module in the above-mentioned multi-structured dialogue data training. Finally, data is collected according to the characteristics of multi-round intelligent dialogues, and the dialogues are optimized and trained through a large model to achieve a more logical intelligent dialogue question-and-answer system based on a large model. DETAILED DESCRIPTION

[0014] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention are described clearly and completely below. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention. Example

[0015] The present invention provides an intelligent dialogue system trained with massive multi-structured dialogue data represented in multiple languages, including multi-structured dialogue data training, using dialogue data with multiple structural features and forms to train models, automatically extracting logical relationships between dialogues and the use of words, phrases and sentences through a large model, and forming knowledge stored inside the large model. The intelligent dialogue module provides users with a real-time, intelligent dialogue experience based on the large module in the multi-structured dialogue data training through natural language processing, speech recognition and speech synthesis technology. Training optimization collects data based on the characteristics of multi-round intelligent dialogues, and optimizes dialogue training through a large model.

[0016] The intelligent dialogue system trained with massive multi-structured dialogue data represented in multiple languages ​​is an advanced dialogue system that combines multi-language support technology and training with massive multi-structured dialogue data to achieve cross-language intelligent dialogue. That is, the logical relationship between dialogues and the use of words, phrases and sentences are automatically extracted through the big model to form knowledge stored inside the big model. Through natural language processing, speech recognition and speech synthesis technology, the big module based on the above multi-structured dialogue data training provides users with a real-time and intelligent dialogue experience. Finally, according to the characteristics of multi-round intelligent dialogue, data is collected and the dialogue is optimized and trained through the big model to achieve a more logical intelligent dialogue question-answering system based on the big model.

[0017] Massive multi-structured dialogue data systems are used for training large amounts of diverse dialogue data. These data may come from different fields, scenarios, and languages, and have different structures and formats. The data may include text dialogues, voice dialogues, image descriptions, etc. These diverse data provide a rich training resource for the system, which helps to improve the system's dialogue understanding and generation capabilities. An intelligent dialogue system is an interactive system based on natural language processing technology that can have coherent and natural conversations with humans. The system usually includes modules such as speech recognition, natural language understanding, dialogue management, and speech synthesis.

[0018] In specific settings, the multi-structure conversation data training also includes the model's architecture selection data, hyperparameter tuning data, and the stability and efficiency data of the training process. The architecture selection data includes data diversity, data balance, and data annotation quality. Data diversity: Ensure that the conversation data covers various scenarios and topics, including daily conversations, professional consultations, emotional exchanges, etc., so that the model can adapt to various complex conversation environments. This diversity helps the model to show better generalization capabilities when dealing with different topics and situations. Data balance: Maintain a balance of information such as positive and negative emotions and various answer types in the conversation data to avoid model bias. Balanced data helps the model form a more comprehensive and objective understanding during the learning process. Data annotation quality: Accurately annotate the data to ensure that the model can learn the correct information. Annotation may include intent recognition, sentiment analysis, entity recognition, and other aspects. These annotation information will serve as supervisory signals during model training.

[0019] Furthermore, the training optimization uses context information to enhance the model's ability to understand, and continuously optimizes the model's performance through incremental training or transfer learning. According to the characteristics of multi-round dialogue, a suitable training method is adopted. The model's ability to understand can be enhanced by using context information; the model's performance can be continuously optimized through incremental training or transfer learning; and the interpretability of the model can also be improved by using technical means such as attention mechanism visualization.

[0020] In the above specific implementation, the natural language processing technology is responsible for understanding and analyzing the user's voice instructions and extracting key information. The speech recognition technology converts the user's voice input into text, providing a basis for subsequent dialogue processing.

[0021] Speech recognition technology involves technologies in signal processing, data analysis, pattern recognition, machine learning, and other aspects. Its principle is to convert a speech signal into corresponding text information. The system mainly includes four parts: feature extraction, acoustic model, language model, and dictionary and decoding. In order to extract features more effectively, it is often necessary to perform preprocessing such as filtering and framing on the collected sound signal to extract the signal to be analyzed from the original signal. Feature extraction: Convert the sound signal from the time domain to the frequency domain to provide a suitable feature vector for the acoustic model. The specific process includes: Speech signal acquisition: Speech is usually input into the computer through a microphone, and the microphone converts the sound wave into a voltage signal, which is then sampled through an A / D device to convert the continuous voltage signal into a digital signal that can be processed by the computer. Preprocessing: Including filtering, A / D conversion, pre-emphasis, and endpoint detection. Filtering is intended to suppress components in the input signal that exceed a specific frequency and power supply frequency interference. A / D conversion is to convert the speech analog signal into a digital signal. The purpose of pre-emphasis processing is to enhance the high-frequency part and make the signal spectrum flat. Endpoint detection is to determine the starting and ending points of speech from a signal containing speech. Feature extraction: Linear predictive coding technology is commonly used to extract speech features to uniquely determine the speech linear prediction coefficient as a characteristic parameter of the speech signal by minimizing the mean square error between the predicted signal and the actual signal. Acoustic model: The spectrogram is passed to the deep learning-based acoustic model to predict the character probability at each time step. During training, the acoustic model is trained using a dataset containing hundreds of hours of audio and transcriptions of the target language. The acoustic model outputs based on how the words are pronounced, so it may contain repeated characters. Language model: Based on linguistics-related theories, the probability that the sound signal corresponds to a possible phrase sequence is calculated. Decoding: According to the existing dictionary, the phrase sequence is decoded to obtain the final possible text representation. The decoder and language model convert characters into word sequences based on the context, which can be further buffered into phrases and sentences and appropriately segmented.

[0022] The core principles of speech-to-text technology include steps such as audio acquisition, signal processing, feature extraction, pattern matching, and text generation. First, the sound signal is captured by a microphone and converted into a digital signal. Then, the digital signal is preprocessed, including operations such as denoising, enhancement, and framing. Next, the features of the audio signal, such as Mel-frequency cepstral coefficients, are extracted so that the computer can understand and process it. After that, the computer compares the extracted features with the known speech model to find the best matching speech pattern. Finally, the recognized speech pattern is converted into readable text.

[0023] Natural language understanding refers to computer technology that enables computers to understand and analyze human natural language. Its goal is to enable computers to understand natural language text like humans and extract useful information from it for further processing. This requires computers to be able to handle complex factors such as grammar, semantics, context, etc. in natural language, and to combine knowledge bases and language models for reasoning and analysis. The working principle of natural language understanding is based on a series of core technologies and steps, which together form the basis for computers to parse and understand human language. Text preprocessing: Tokenization: Split continuous text strings into meaningful units, such as words, phrases, or symbols. Text cleaning: Remove irrelevant characters, such as punctuation marks, special characters, etc. Normalization: Convert text to a standard format, such as converting all characters to lowercase, or converting abbreviations to full forms. Syntactic analysis: Part-of-speech tagging: Assign a part of speech to each vocabulary, such as noun, verb, etc. Dependency grammar analysis: Establish dependency relationships between words, such as the connection between subject and predicate. Semantic analysis: Entity recognition: Identify specific entities in the text, such as names, places, organizations, etc. Intent recognition: Determine the purpose or intent of user input, which is particularly important in chatbots and intelligent assistants. Context and dialogue understanding: Dialogue state management: Track the progress of the conversation and remember previous queries and responses to maintain the coherence of the conversation. Context understanding: Analyze the impact of previous input on the meaning of the current input.

[0024] Dialogue management controls the process of human-computer dialogue. It determines the next system behavior and the updated dialogue state based on the semantic expression of user input and the current dialogue state. This is a process that goes back and forth and keeps flowing until the task is completed. The main functions of dialogue management include: Dialogue state tracking: recording all information generated during the dialogue process, including user input, system output, and dialogue context, so as to accurately understand the user's current intention. Dialogue strategy formulation: based on the current dialogue state and user input, determine what action the system should take next, involving the design of dialogue process, the formulation of dialogue strategy, and the selection of dialogue behavior. Dialogue process control: ensuring the coherence and effectiveness of the dialogue, including handling noise and ambiguity in user input, guiding users to conduct dialogue according to the preset process, etc. Its pipe fitting technology includes key technologies Dialogue state tracking: one of the key technologies in multi-round dialogue, responsible for tracking all information generated during the dialogue process. By maintaining the dialogue state, the system can accurately understand the user's current intention and generate appropriate responses accordingly. Dialogue strategy: based on the current dialogue state and user input, determine what action the system should take next. Involving the design of dialogue process, the formulation of dialogue strategy, and the selection of dialogue behavior, etc. Frame-based dialogue management: A commonly used multi-round dialogue implementation technology. By defining one or more frames, each frame contains a series of slots for storing information that needs to be collected during the dialogue. The system fills these slots based on the user's input and performs corresponding operations after all slots are filled.

[0025] Specifically, the speech synthesis technology converts computer-generated text responses into speech, allowing users to hear fluent and natural speech output.

[0026] The basic principle of speech synthesis technology is to convert text into speech signals. First, the input text is analyzed to extract linguistic information from the text, such as pronunciation, character shape, part of speech, grammatical structure, etc. Then, these linguistic information are converted into a format suitable for speech synthesis, such as phoneme sequence, prosodic features, etc. Next, the acoustic model converts these linguistic information into acoustic features of speech signals, such as spectrum, fundamental frequency, duration, etc. Finally, the vocoder converts these acoustic features into audible speech signals. The core technology includes the acoustic model: the acoustic model is the core part of speech synthesis technology, which is responsible for converting the linguistic information obtained from text analysis into the acoustic features of speech signals. At present, there are two main types of acoustic models based on deep learning: acoustic models based on recurrent neural networks and acoustic models based on convolutional neural networks. The acoustic model based on recurrent neural networks has a strong sequence modeling ability and can better capture the time domain characteristics of speech signals; while the acoustic model based on convolutional neural networks has a strong spatial modeling ability and can better capture the frequency domain characteristics of speech signals. Vocoder: A vocoder is a device that converts the acoustic features output by an acoustic model into an audible speech signal. A vocoder based on waveform synthesis directly synthesizes the waveform of a speech signal, which has a higher sound quality but a higher computational complexity. A vocoder based on parameter synthesis synthesizes the parameters of a speech signal and then converts these parameters into a waveform of a speech signal, which has a relatively lower sound quality but a lower computational complexity.

[0027] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. An intelligent dialogue system trained with massive multi-structure dialogue data, characterized by: include Multi-structure conversation data training uses conversation data with multiple structural features and forms to train models. The big model automatically extracts the logical relationship between conversations and the use of words and sentences to form knowledge stored in the big model. The intelligent conversation module uses natural language processing, speech recognition and speech synthesis technologies to provide users with a real-time, intelligent conversation experience based on the large modules trained in the above multi-structure conversation data; Training optimization: collect data based on the characteristics of multi-round intelligent conversations, and optimize conversation training through a large model.

2. The intelligent dialogue system trained with massive multi-structure dialogue data according to claim 1, characterized in that: The multi-structure dialogue data training also includes model architecture selection data, hyperparameter tuning data, and training process stability and efficiency data.

3. The intelligent dialogue system trained with massive multi-structure dialogue data according to claim 2, characterized in that: The architecture selection data includes data diversity, data balance and data annotation quality.

4. The intelligent dialogue system trained with massive multi-structure dialogue data according to claim 1, characterized in that: The training optimization utilizes contextual information to enhance the model's understanding ability and continuously optimizes the model performance through methods such as incremental training or transfer learning.

5. The intelligent dialogue system trained with massive multi-structure dialogue data according to claim 4, characterized in that: The training optimization uses attention mechanism and visualization technology to improve the interpretability of the model.

6. The intelligent dialogue system trained with massive multi-structure dialogue data according to claim 1, characterized in that: The natural language processing technology is responsible for understanding and analyzing the user's voice commands and extracting key information.

7. The intelligent dialogue system trained with massive multi-structure dialogue data according to claim 1, characterized in that: The speech recognition technology converts the user's speech input into text, providing a basis for subsequent dialogue processing.

8. The intelligent dialogue system trained with massive multi-structure dialogue data according to claim 1, characterized in that: The speech synthesis technology converts computer-generated text responses into speech, allowing users to hear fluent and natural speech output.