Text extraction method and system based on large language model
Through the text extraction method based on the large language model, combined with multi-level intent recognition and RAG method, the problem of flexibility and insufficient context understanding of the existing traffic information query system in handling complex queries is solved, and more efficient and accurate users are achieved Intent to understand and personalize service.
Patent Information
- Application Number
- CN202411107236.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-13
- Publication Date
- 2024-11-15
AI Technical Summary
The existing traffic information query system is not flexible enough when handling complex or fuzzy natural language queries, and cannot understand the user's historical query and dialogue context, resulting in low accuracy and poor interactivity.
Using text extraction method based on large language model, multi-level intent recognition and RAG method, combined with historical query records and real-time traffic information, comprehensive answers are generated, and behavioral models are dynamically optimized to provide personalized services, which can perform multiple verification of complex queries and confirm.
It improves the accuracy and practicality of user experience and answers, can better handle varied and vague user intentions, provide a personalized and conversational query experience, and maintains the timeliness and accuracy of the system.
Smart Images

Figure CN118964575A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of traffic information query systems, and in particular to a text extraction method and system based on a large language model. Background Art
[0002] With the development of intelligent transportation systems, traffic information query systems are becoming more and more important in daily life. These systems help users obtain real-time traffic conditions, plan travel routes, and provide information such as accidents and road conditions. However, there are some significant deficiencies in current traffic information query systems. Existing traffic information query systems mainly rely on keyword matching and predefined databases. When a user enters a query, the system matches it according to preset rules and keywords and returns relevant results. This approach performs well when processing simple queries, but often exhibits the following deficiencies when facing complex or ambiguous natural language queries: Insufficient flexibility: Existing systems mainly rely on fixed rules and keyword matching, which makes it difficult to handle complex or changing queries. Lack of context: It is unable to continuously understand the user's historical queries and conversation context, resulting in users having to repeatedly enter similar questions. Low accuracy: The approach that relies on keyword matching often returns results that do not meet the user's actual needs. Poor interactivity: Lack of interaction with users and inability to provide a personalized, conversational query experience Traditional traffic information query systems mainly rely on keyword matching and predefined databases, which cannot handle complex natural language queries and lack flexibility and context understanding capabilities.
[0003] For example, a "text extraction model training method, text extraction method and device" disclosed in a Chinese patent document, with the publication number: CN118132683A, discloses the following steps: obtaining traffic text data; inputting the traffic text data and a first prompt information into a large language model, wherein the first prompt information is used to prompt the acquisition of keywords corresponding to traffic elements; extracting keywords corresponding to traffic elements in the traffic text data through the large language model, and outputting text feature data corresponding to the traffic text data; the text feature data includes the correspondence between traffic elements and keywords; constructing a training data set, wherein the training data set includes multiple traffic text data and corresponding labels, wherein the labels corresponding to the traffic text data are text feature data; training the text extraction model based on the training data set until the text extraction model meets the preset convergence conditions, thereby obtaining a target text extraction model. However, this solution also lacks optimization for complex and changeable queries. Summary of the invention
[0004] In order to solve the problem that the prior art cannot process complex natural language queries, the present invention provides a text extraction method and system based on a large language model, which can perform multiple verifications and confirmations on complex queries through multi-level intent recognition, and can better handle changeable and ambiguous user intentions.
[0005] In order to achieve the above object, the present invention provides the following technical solutions: A text extraction method based on a large language model, comprising: obtaining historical query records and establishing a behavior model through machine learning; Determine the query mode and behavior characteristics of the behavior model through the model recognition algorithm, and dynamically optimize the behavior model in combination with the deep recognition model; Obtain real-time traffic information, generate comprehensive answers through the RAG method in combination with the behavior model, and modify the output of the behavior model based on the comprehensive answer to obtain an accurate answer. It can deeply understand the user's query habits, preferences and behavior patterns, so as to provide more personalized services. This personalization not only improves the user experience, but also makes the answers closer to the actual needs of users. The system can continuously learn and adapt to changes in query requests. This adaptive ability enables the system to quickly adjust and give more accurate answers when faced with changes in query requests or new query patterns, maintaining the timeliness and accuracy of the system. Combined with the behavior model, the comprehensive answer is generated through the RAG (Retriever-Augmenter-Generator) method, which can not only provide traffic information closely related to the user's query, but also give a more comprehensive and specific answer based on the user's historical behavior. This combination of real-time and comprehensiveness greatly improves the quality and practicality of the answer. It can perform multiple verifications and confirmations on complex queries, and can better handle changeable and ambiguous user intentions.
[0006] Preferably, when establishing a behavior model, query habits and historical data are obtained based on historical query records. After cluster analysis of the query habits and historical data, the behavior model is obtained by training based on a machine learning model. Similar query behaviors or data patterns are grouped to achieve refined classification of query behaviors. This classification helps to gain a deeper understanding of the characteristics and behavior patterns of user groups, and provides more accurate input for subsequent model training. The data after cluster analysis is purer and more concentrated, reducing the impact of noise and outliers on model training. Training a machine learning model based on such data can significantly improve the accuracy and predictive power of the model, making the generated behavior model closer to the user's real behavior.
[0007] As a preferred method, it includes deep intent recognition, semantic analysis based on deep learning models, intent evaluation based on semantic analysis results, and optimization of behavior models based on behavior characteristics. Through deep learning technology, the solution can more accurately identify the user's deep intent, not just the surface text or query, and can more accurately grasp the subtle differences between different query requests, improving the accuracy of intent evaluation.
[0008] As a preferred method, it includes pattern recognition, using K-means clustering algorithm to identify query patterns and behavior characteristics; detecting abnormal query behavior, and filtering query requests that are not in historical query records. It can achieve refined classification of query behavior and quickly identify query requests that do not conform to normal behavior, providing a basis for further security analysis or business optimization.
[0009] Preferably, when modifying the output of the behavior model, the comprehensive answer is traversed, and when the comprehensive answer is hit, the corresponding comprehensive answer is replaced with the output of the behavior model. It can more accurately reflect the user's current state and preferences, and can dynamically adjust the output according to real-time data. Therefore, replacing it with the output of the behavior model can enhance the flexibility and adaptability of the system, so that it can better adapt to different scenarios and query requests. By replacing the comprehensive answer with the output of the behavior model, it is actually using the actual scenario to provide feedback and verification to the model.
[0010] As a preferred method, it includes waking up, obtaining a query request, and identifying a wake-up word in the query request; after the wake-up word is detected, text recognition is performed to obtain the query text. This can effectively reduce false triggers and improve the accuracy and efficiency of system response. The query text can be continuously obtained without additional operations, which improves the fluency and coherence of the interaction.
[0011] Preferably, after waking up, preliminary intent recognition is performed through Rasa to determine the basic intent in the query request; the basic intent is identified and predicted through the behavior model. Rasa can quickly obtain the query purpose, and combined with the behavior model, the accuracy of intent recognition can be significantly improved, which is convenient for solving ambiguous and vague problems in different query requests.
[0012] A text extraction system based on a large language model, applicable to a text extraction method based on a large language model as described in any one of the embodiments, comprises: an input module, obtaining a query request; The input module is connected to an intent recognition model, which performs multi-layer intent recognition based on query requests; The intent recognition model is connected to the RAG model, which provides the intent recognition model with the scope of intent responses based on real-time information. It can perform multi-layer intent recognition, deeply understand the complexity and main requirements of query requests, and capture the nuances and potential intents in query requests. It can reduce the information loss between retrieval and generation in traditional methods, thereby improving the efficiency and accuracy of the entire system.
[0013] Preferably, the intent recognition model includes a behavior model and a deep learning model connected in sequence. The behavior model recognizes intent based on historical query records; the deep learning model recognizes intent on the behavior model. The phased and progressive recognition process helps reduce misjudgments and missed judgments and improves the accuracy of intent recognition.
[0014] Preferably, an output module connected to the intention recognition module is included, and the output module generates answer text and synthesizes it into speech output, so as to realize multi-modal information output, so as to adapt to a wider range of scenarios and query requirements.
[0015] The present invention has the following advantages: (1) Improving the accuracy and flexibility of intent recognition: Through a multi-level adaptive intent recognition mechanism, the present invention combines the query request analysis module and context memory to more accurately understand and respond to user query needs, ensuring efficient processing of complex and continuous queries; (2) Real-time data access and dynamic information integration: The present invention integrates multiple real-time traffic data sources, and through retrieval enhanced generation RAG technology, combines real-time data with generation models to provide the latest and comprehensive high-quality traffic information responses, significantly improving the efficiency and effectiveness of traffic management. BRIEF DESCRIPTION OF THE DRAWINGS
[0016] The drawings described below are merely exemplary. A person skilled in the art can derive other implementation drawings based on the provided drawings without inventive effort.
[0017] Figure 1 Schematic diagram of the method steps in the embodiment.
[0018] Figure 2 It is a schematic diagram of the working logic in the embodiment. DETAILED DESCRIPTION
[0019] Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative work shall fall within the scope of protection of the present invention.
[0020] like Figure 1-Figure 2 As shown, in a preferred embodiment, the present invention discloses a text extraction method based on a large language model, including: obtaining historical query records, realizing feature analysis based on historical data; establishing a behavior model through machine learning; being able to deeply understand the user's query habits, preferences and behavior patterns, thereby providing more personalized services. This personalization not only improves the user experience, but also makes the answer closer to the user's actual needs.
[0021] The query mode and behavior characteristics of the behavior model are determined through the model recognition algorithm, and further data mining based on historical data characteristics is realized to achieve classification of different characteristics; the behavior model is dynamically optimized in combination with the deep recognition model; it can continuously learn and adapt to changes in query requests. This adaptive capability enables the system to quickly adjust and give more accurate answers when faced with changes in query requests or new query patterns, maintaining the timeliness and accuracy of the system.
[0022] Acquire real-time traffic information to reflect the real-time nature of the solution. Generate comprehensive answers through the RAG method in combination with the behavioral model to achieve the screening of real-time traffic information. According to the comprehensive answer, modify the output of the behavioral model to obtain an accurate answer. In combination with the behavioral model, the RAG method is used to generate comprehensive answers, which can not only provide traffic information closely related to the query, but also give more comprehensive and specific answers in combination with the user's historical behavior. This combination of real-time and comprehensiveness greatly improves the quality and practicality of the answers. It can perform multiple verifications and confirmations on complex queries and better handle the changing and ambiguous user intentions.
[0023] After generating a comprehensive answer, the method also modifies the output of the behavioral model based on the comprehensive answer to obtain a more accurate answer. This feedback loop mechanism helps reduce errors and improve the accuracy and reliability of the answer.
[0024] In other embodiments, when establishing a behavior model in the present solution, query habits and historical data are obtained based on historical query records. By clustering the query habits and historical data, similar query behaviors or data patterns can be grouped, thereby achieving refined classification of query requests. This classification helps to gain a deeper understanding of the characteristics and behavior patterns of user groups, and provides more accurate input for subsequent model training. After clustering the query habits and historical data, the data after clustering analysis is purer and more concentrated, reducing the impact of noise and outliers on model training. Training a machine learning model based on such data can significantly improve the accuracy and predictive power of the model, making the generated behavior model closer to the user's real behavior; training based on the machine learning model obtains a behavior model.
[0025] Cluster analysis not only helps identify the patterns of specific query requests, but also reveals the connections and differences between different patterns. This in-depth understanding of the data helps train models with stronger generalization capabilities, making reasonable predictions and responses even when faced with new and unseen queries.
[0026] Through cluster analysis, similar query habits can be identified, and query requests for specific types of information that are frequently searched can be responded to more accurately and quickly.
[0027] In other embodiments, the present solution includes deep intent recognition. Through deep learning technology, the solution can more accurately identify the deep intent of the query request, rather than just the surface text or query. This deep intent recognition capability enables the system to better understand the real needs of the query request issuer. Semantic analysis is performed based on the deep learning model. The semantic analysis based on the deep learning model can deeply understand the semantic content of the user input, so that the system can more accurately grasp the nuances of the query request. Intent evaluation is performed based on the results of the semantic analysis. And the behavioral model is optimized in combination with behavioral characteristics. This solution not only focuses on the immediate intent of the query request, but also optimizes the behavioral model in combination with the behavioral characteristics of historical query requests, which can more comprehensively understand the query request and achieve customized response to each query request.
[0028] In other embodiments, the present solution includes pattern recognition, which can determine the relevant features of different query patterns. Use the K-means clustering algorithm to identify query patterns and behavioral features; K-means can effectively identify patterns and features in query behavior through cluster analysis. Detect abnormal query behavior and filter out query requests that are not in historical query records. After K-means clustering, new query requests can be compared with existing clustering groups. If a query request has a low similarity with all existing groups, it is likely to be an abnormal query behavior. This method can quickly identify query requests that do not conform to conventional behavior, providing a basis for further security analysis or business optimization. Combined with appropriate monitoring and data processing mechanisms, K-means clustering can classify and detect anomalies of new query requests in real time.
[0029] The K-means algorithm is an automated data processing technology that can automatically classify similar query requests into the same group. When processing large-scale data sets, the K-means algorithm shows good scalability and efficiency. It can quickly group data and maintain a high processing speed as the amount of data increases. The algorithm can adjust the number of clusters (K value) according to actual needs to adapt to different application scenarios and data characteristics. This flexibility allows K-means to be widely used in various complex data sets.
[0030] In other embodiments, when modifying the output of the behavior model, the solution includes traversing the comprehensive answers and performing customized screening based on the comprehensive answers, so that different query requests are optimized using their respective comprehensive answers. When a comprehensive answer is hit, the corresponding comprehensive answer is replaced with the output of the behavior model. The behavior model is usually trained based on the user's specific behavior data, so it can more accurately reflect the user's current status and preferences. By replacing the comprehensive answer with the output of the behavior model, the response can be made more personalized and accurate, better meeting the actual needs of the query request.
[0031] Traditional comprehensive answer sets may be limited by their predefined content and are difficult to cope with complex and changing user behaviors and needs. Behavioral models can dynamically adjust outputs based on real-time data, so replacing the outputs with behavioral models can enhance the flexibility and adaptability of the system, enabling it to better adapt to different scenarios and user groups.
[0032] By replacing the comprehensive answer with the output of the behavioral model, we are actually using the actual scenario to provide feedback and verification to the model. This real-time feedback helps to discover deficiencies and errors in the model, thereby promoting continuous optimization and improvement of the model. At the same time, this also helps the model better understand and adapt to the complexity and diversity of actual application scenarios.
[0033] In other embodiments, this solution includes wake-up, which quickly activates the device through the wake-up word, greatly improving the convenience and immediacy of use. Obtain a query request and identify the wake-up word in the query request; start the system by identifying a specific wake-up word, which can effectively reduce false triggers and improve the accuracy and efficiency of system response. When the wake-up word is detected, text recognition is performed to obtain the query text. After the wake-up word is detected, text recognition is performed immediately to improve the fluency and coherence of the interaction.
[0034] In other embodiments, the present solution includes performing preliminary intent recognition through Rasa after wake-up, and the basic intent can be quickly extracted from the user's query. Rasa is a powerful natural language processing (NLP) tool that can understand and parse the user's natural language input based on a machine learning model, and quickly identify the user's preliminary needs or query purpose. This preliminary screening greatly reduces the complexity and computational complexity of subsequent processing steps. Determine the basic intent in the query request; identify and predict the basic intent through a behavioral model. On the basis of preliminary intent recognition, further recognition and prediction through a behavioral model can significantly improve the accuracy of intent recognition. The behavioral model can infer the true intent represented by the query request based on historical behavior, contextual information, and more complex logical rules. This dual recognition mechanism helps to resolve ambiguity and fuzziness in natural language, allowing the system to more accurately understand the true needs of the query request. By accurately identifying the user's intent, a more personalized and expected response can be provided.
[0035] In other embodiments, the present solution discloses a text extraction system based on a large language model, which is applicable to a text extraction method based on a large language model as described in any embodiment, including: an input module to obtain a query request; obtaining a query request through the input module.
[0036] The input module is connected to the intent recognition model, which recognizes the query request. The intent recognition model performs multi-layer intent recognition based on the query request; it can deeply understand the complexity and multi-level nature of the user's query request. Multi-layer intent recognition not only identifies the user's main needs, but also captures the nuances and potential intent in the query, thereby providing a more accurate and personalized response.
[0037] The intent recognition model is connected to the RAG model. The introduction of the RAG model enables the system to provide the intent recognition model with the scope of intent answers based on real-time information. This real-time performance is particularly important for processing queries with strong timeliness, such as news events, market trends, etc. By combining real-time data, the system can generate more accurate and up-to-date answers to meet users' needs for the latest information. The RAG model provides the intent recognition model with the scope of intent answers based on real-time information. By integrating the retrieval and generation processes into a large model, the RAG model can reduce the information loss between retrieval and generation in traditional methods, thereby improving the efficiency and accuracy of the entire system. In addition, the powerful language processing capabilities of the large language model itself also provide a solid foundation for the system, enabling the system to handle more complex and diverse query requests.
[0038] In other embodiments, the intent recognition model in this solution includes a behavioral model and a deep learning model connected in sequence. The behavioral model performs intent recognition based on historical query records, which can capture patterns and trends in user behavior and is very effective in understanding users' long-term preferences and habits. The deep learning model is good at processing complex language structures and semantic relationships, and can more deeply understand the nuances and potential intentions of query requests. The behavioral model performs intent recognition based on historical query records; the behavioral model provides valuable prior information for the deep learning model. Preliminary intent recognition is performed by first using the behavioral model. The deep learning model performs intent recognition on the behavioral model, and the deep learning model can perform more refined analysis and processing on this basis. This phased and progressive recognition process helps to reduce misjudgments and missed judgments and improve the accuracy of intent recognition.
[0039] The behavior model can be continuously optimized and improved as the user's historical query records accumulate, thereby adapting to changes in user behavior. The deep learning model can adapt to new language phenomena and query patterns through continuous training and learning. This combination makes the entire intent recognition model highly adaptable and flexible, and can cope with ever-changing user needs and query environments.
[0040] In some cases, deep learning models may require a lot of computing resources to perform complex calculations and analysis. The behavior model, as a preliminary identification stage, can filter out some query requests that are obviously inconsistent with user intent, thereby reducing the amount of data that the deep learning model needs to process. This helps reduce the computing resource consumption of the entire system and improve the system's response speed and efficiency.
[0041] In another embodiment, the present solution discloses an intelligent traffic information query method based on a large language model, which provides intelligent and accurate traffic information query and answer by combining a large language model and retrieval enhancement generation technology. Specifically, the following steps are included: first, adaptive intent recognition is performed, and the intent recognition model is dynamically adjusted through a machine learning algorithm in combination with the user's historical query data and context information. By analyzing the user's past query records and conversation content, the system can adaptively adjust the understanding and processing methods of different queries. It can improve the ability to understand complex queries and the accuracy of intent recognition.
[0042] Then, multi-level intent recognition is performed, and a multi-level intent recognition architecture is introduced. The first layer uses Rasa for preliminary intent recognition, and the second layer uses a large language model for in-depth understanding and confirmation. Through the multi-level architecture, the system can perform multiple verifications and confirmations on complex queries, improving recognition accuracy. Dynamically update traffic information. This includes accessing and processing real-time traffic data, integrating multiple real-time traffic data sources, including traffic monitoring systems and traffic data APIs, and obtaining the latest traffic conditions, road conditions, and accident reports in real time. Through data stream processing, clean, integrate, and analyze real-time data. Through dynamic access and processing of real-time traffic data, ensure that the latest and most accurate traffic information is obtained.
[0043] Including integrating and generating dynamic information, using the RAG method to combine the traffic information retrieved in real time with the generative model to generate comprehensive, high-quality responses. When inquiring about the current traffic conditions in the city center, the RAG method is used to retrieve the latest data in real time and provide a comprehensive answer in combination with the generative model. By integrating and generating dynamic information, high-quality and comprehensive traffic information responses based on real-time data are provided.
[0044] In another specific embodiment, the solution includes the following steps: obtaining a query request: obtaining a voice query request inputted by voice, which is the starting point of the whole process, and a query related to traffic information is raised through natural language.
[0045] Wake-up activation: activate the system through the wake-up word. The wake-up module detects the predefined wake-up word to activate the system and prepare to receive the user's voice input.
[0046] Voice capture, capturing the user's voice input. Through a microphone array or other audio capture device, the user's voice input is captured and converted into a digital signal for subsequent processing.
[0047] Speech recognition, converting speech into text. Using automatic speech recognition ASR technology, the captured speech signal is converted into text format for subsequent natural language processing.
[0048] Preliminary intent recognition, performed through Rasa. This module analyzes the text entered by the user and identifies the user's basic intent, such as checking traffic conditions, finding the best route, etc.
[0049] User behavior analysis combines user historical data to analyze user query habits and behavior patterns. Through machine learning algorithms, a user behavior model is formed and applied to preliminary intent recognition to enhance the ability to understand and predict user intent.
[0050] Context memory saves the user's historical query and conversation context information. This module can remember the user's past query records and use this information in subsequent conversations to more accurately understand and respond to the user's current query.
[0051] Deep intent recognition uses the deep learning model GLM4 to deeply understand and confirm the initial recognition results. This module conducts a more detailed analysis of the user's intentions to further improve the accuracy of intent recognition.
[0052] Dynamic adjustment: dynamically adjust the intent recognition model based on the user's contextual memory and current query. By adjusting the model's parameters and strategies in real time, the system can better adapt to the user's query habits and changes.
[0053] Accurate answers generate accurate answer content. Through the deep learning model GLM4, the system generates detailed answers to user queries, ensuring the accuracy and relevance of the answer content.
[0054] Real-time data access integrates multiple real-time traffic data sources to obtain the latest traffic conditions, road conditions and accident reports. This module ensures that the system can update traffic information in real time and provide the latest data to users.
[0055] Dynamic information integration, using retrieval-augmented generation (RAG) technology, combines real-time retrieved traffic information with the generation model to generate comprehensive, high-quality responses. This module integrates multiple data sources to generate comprehensive and high-quality traffic information responses.
[0056] Personalized recommendations: By analyzing the user's historical query and behavior data, we provide personalized traffic query information recommendations and navigation suggestions. Based on the user's preferences and habits, we provide personalized information that better meets their needs.
[0057] Generate a response and generate the final response content. Combine the information and processing results of the above steps to generate the final answer text of the user's query.
[0058] Text return. Return the generated answer text to the user. This step transmits the text answer generated by the system to the speech synthesis module for speech output.
[0059] Speech synthesis, converting text to speech. Using text-to-speech (TTS) technology such as Vits, the text responses generated by the system are converted into speech signals.
[0060] Voice output: The voice response generated by the voice output module is played. The user finally receives the system's answer through hearing.
[0061] In another preferred embodiment, the solution includes the following modules: a user behavior analysis module, in the process of preliminary intention identification, combined with the user's historical data, to perform habit identification and enhancement through the user behavior analysis module. This module uses machine learning algorithms to analyze the user's query habits and historical data to form a user behavior model.
[0062] Data collection and storage: The system will collect and store each user's query record, including query content, time, geographic location and other information.
[0063] Behavior model construction: Using these data, through cluster analysis and pattern recognition algorithms, we build a user behavior model. The model can identify common query patterns and habits of users, such as when users tend to query traffic conditions in a certain area, and the roads they frequently query.
[0064] Habit recognition and enhancement: When performing preliminary intent recognition, the query is refined based on the user's behavior model. For example, when a user queries "current traffic conditions", the system can identify the roads that the user frequently queries and automatically associate them, improving recognition accuracy.
[0065] Deep intent recognition module: Use the deep learning model GLM4 to deeply understand and confirm the preliminary recognition results.
[0066] Multi-level processing: After the initial intent recognition, the deep intent recognition module further refines and confirms the results. The deep learning model re-evaluates the initial intent recognition to ensure the accuracy and completeness of the recognition results.
[0067] Semantic analysis: Through deep semantic analysis technology, the system can understand complex sentence structures and implicit intentions. For example, for multi-level queries or complex sentences containing multiple intentions, the deep learning model can accurately split and understand each intention.
[0068] In another specific embodiment, multiple real-time traffic data sources are integrated to obtain the latest traffic conditions, road condition information and accident reports.
[0069] Data source integration: The system establishes connections with various traffic data sources (such as traffic monitoring systems, traffic data APIs, public transportation data, etc.) to receive and process traffic information in real time.
[0070] Data stream processing: Use data stream processing technology to clean, filter and integrate the received data in real time to ensure the accuracy and timeliness of the data. For example, the system can receive road condition information from the traffic monitoring system in real time and integrate it into the user's query answer.
[0071] Dynamic information integration and generation,using retrieval-augmented generation (RAG) technology,combining the traffic information retrieved in real time with the,generative model to generate comprehensive and high-quality responses.
[0072] Information retrieval: When a user queries, the system will retrieve relevant traffic information in real time, including current road conditions, accident reports, weather impacts, etc.
[0073] Generative model integration: The retrieved information is combined with the generative model to generate a comprehensive answer through RAG technology. The generative model will generate detailed answer content based on the retrieval results and ensure the accuracy and coherence of the information.
[0074] High-quality responses: Through dynamic information integration and generation, the system can provide high-quality responses based on real-time data. For example, when a user queries "current traffic conditions in the city center", the system can retrieve the latest data in real time and generate a detailed answer.
[0075] By analyzing users' historical query and behavior data, we provide personalized traffic information recommendations and management suggestions.
[0076] User data analysis: The system analyzes the user's historical query and behavior data to understand their preferences and habits, such as the areas that users frequently query and the commonly used traffic management methods.
[0077] Personalized recommendation generation: Based on user data, the system makes personalized recommendations when providing traffic information. For example, when a user inquires about road conditions, the system will give priority to providing real-time information on areas that the user frequently supervises, and recommend the best management solution based on their preferences.
[0078] Dynamic adjustment: Dynamically adjust recommendation content and strategies based on real-time user feedback and usage to ensure personalized and accurate recommendations.
[0079] In another specific embodiment, user A checks the traffic conditions in the area he is responsible for every day before going to work. By analyzing the query habits of user A, the system recognizes that he checks the traffic conditions in the same area between 7 and 8 am every day.
[0080] Behavior model construction: The system constructs a behavior model of user A based on user A’s historical query records.
[0081] Habit recognition and enhancement: When user A inquires about traffic conditions between 7 and 8 in the morning, the system automatically recognizes his intention and provides the latest traffic information in the area based on real-time data.
[0082] Contextual memory and dynamic adjustment: If the query content of user A changes, such as adding a frequently queried alternative road, the system will dynamically adjust the behavior model and provide real-time information about this alternative road in future queries.
[0083] In other embodiments, this solution combines user behavior models and pattern recognition results to identify users' common query habits and preferences when implementing habit recognition and enhancement.
[0084] Intelligent enhancement: During the initial intent recognition, the system intelligently enhances the accuracy of intent recognition based on the user's habits and preferences. For example, when a user queries "traffic conditions", the system can give priority to providing real-time traffic information in areas that the user frequently queries.
[0085] These modifications or improvements made on the basis of the present invention without departing from the spirit of the present invention all fall within the scope of protection required by the present invention.
Claims
1. A text extraction method based on a large language model, characterized in that: include: Obtain historical query records and build behavior models through machine learning; Determine the query mode and behavior characteristics of the behavior model through the model recognition algorithm, and dynamically optimize the behavior model in combination with the deep recognition model; Real-time traffic information is obtained and combined with the behavior model to generate a comprehensive answer through the RAG method. Based on the comprehensive answer, the output of the behavior model is modified to obtain an accurate answer.
2. The text extraction method based on a large language model according to claim 1, characterized in that: When establishing a behavior model, query habits and historical data are obtained based on historical query records. After clustering analysis of the query habits and historical data, the behavior model is obtained by training based on a machine learning model.
3. A text extraction method based on a large language model according to claim 1 or 2, characterized in that: It includes deep intent recognition, semantic analysis based on deep learning models, intent evaluation based on the results of semantic analysis, and optimization of behavior models based on behavioral characteristics.
4. A text extraction method based on a large language model according to claim 1 or 2, characterized in that: This includes pattern recognition, using the K-means clustering algorithm to identify query patterns and behavioral features; Detect abnormal query behavior and filter out query requests that are not in historical query records.
5. A text extraction method based on a large language model according to claim 4, characterized in that: When modifying the output of the behavior model, the comprehensive answers are traversed, and when the comprehensive answer is hit, the corresponding comprehensive answer is replaced with the output of the behavior model.
6. The text extraction method based on a large language model according to claim 1, characterized in that: It includes waking up, obtaining query requests, and identifying the wake-up words in the query requests; when the wake-up words are detected, text recognition is performed to obtain the query text.
7. A text extraction method based on a large language model according to claim 6, characterized in that: After being awakened, preliminary intent recognition is performed through Rasa to determine the basic intent in the query request; Identify and predict basic intentions through behavioral models.
8. A text extraction system based on a large language model, applicable to a text extraction method based on a large language model as claimed in any one of claims 1 to 7, characterized in that: include: Input module, obtains query request; The input module is connected to an intent recognition model, which performs multi-layer intent recognition based on query requests; The intent recognition model is connected to the RAG model, and the RAG model provides the intent recognition model with an intent answer range based on real-time information.
9. A text extraction system based on a large language model according to claim 8, characterized in that: The intent recognition model includes a behavior model and a deep learning model that are connected in sequence. The behavior model performs intent recognition based on historical query records; the deep learning model performs intent recognition on the behavior model.
10. A text extraction system based on a large language model according to claim 8, characterized in that: It includes an output module connected to the intention recognition module, which generates answer text and synthesizes it into speech output.
Citation Information
Cited By
Automobile intelligent question-answering system and method based on large model and retrieval enhancement
CN119202206A
Digital human real-time interaction method based on large model and RAG
CN119990309A
