Intelligent car book information processing method and system

Abstract

The invention provides an intelligent car book information processing method and system. The method comprises the following steps: acquiring car book data; the vehicle book data is cleaned, fragmented and structured, the structured vehicle book data is mapped into vectors in a high-dimensional vector space through a large language model, and then the vectors are imported into the vector database; the vehicle book data is cleaned, knowledge points are extracted according to a predefined concept tree, then a knowledge graph is constructed, and the knowledge graph is imported into the graph database; after a user inputs a question, searching and replying the question in the vector database by using the large language model and searching and replying the question in the graph database by using the knowledge graph; and performing evaluation arbitration on the replies of the large language model and the knowledge graph according to a preset arbitration standard, and determining and outputting an optimal reply. According to the method, the advantages of the large language model and the knowledge graph can be deeply fused, and the vehicle-mounted interaction experience and the information acquisition efficiency are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent vehicle-mounted scene technology, and in particular to an intelligent vehicle book information processing method and system based on the combination of a large language model and a knowledge graph in a vehicle-mounted scene. Background Art

[0002] With the rapid development of information technology, especially the deep integration of artificial intelligence, natural language processing and big data technology, the functions of intelligent in-vehicle systems are becoming increasingly rich, and user experience is constantly improving. Among them, the in-vehicle intelligent question-and-answer system, as a core component of human-computer interaction, plays a vital role in improving driving safety, convenience and entertainment. However, when dealing with complex, professional or specific field problems such as car book questions, existing in-vehicle question-and-answer systems often have problems such as insufficient response accuracy and low information retrieval efficiency, and cannot fully meet users' needs for real-time and accurate information acquisition. "Car books" usually refer to instructions, manuals or materials related to cars, covering information on the car's configuration, usage, maintenance, troubleshooting and other aspects.

[0003] Smart Car Book Q&A mainly focuses on the detailed description and specifications of the vehicle itself, such as the functional introduction of each component of the vehicle, the specific operating guidelines in the user manual, the maintenance cycle and requirements, etc. These contents usually come from the official documents provided by the vehicle manufacturer, that is, the car book.

[0004] At present, large language models have shown unique advantages in in-vehicle question-answering systems. Large language models (LLMs) such as GPT-3 and BERT have demonstrated excellent performance in many natural language processing tasks with their powerful language understanding and generation capabilities. These models learn massive amounts of text data through large-scale pre-training, and are able to understand and generate natural, fluent, and logical texts, providing a strong language understanding and generation foundation for in-vehicle question-answering systems. However, large language models may have limitations when dealing with professional fields or fine-grained knowledge such as car book questions, especially in the absence of relevant training data, and the accuracy of their answers may be affected.

[0005] As a structured form of knowledge representation, knowledge graphs organize and store various entities and their relationships in a graph data structure consisting of nodes, edges, and attributes. They can accurately and efficiently carry large-scale, multi-domain professional knowledge. The application of knowledge graphs in question-answering systems can provide accurate and authoritative knowledge support for the system, especially when dealing with professional and complex problems, and can make up for the shortcomings of large language models in understanding professional knowledge.

[0006] Although large language models and knowledge graphs each demonstrate unique advantages in in-vehicle question-answering systems, how to effectively combine the two to complement each other's strengths in order to accurately answer complex and professional questions in the in-vehicle environment is a technical challenge that needs to be solved urgently. Summary of the invention

[0007] The purpose of the present invention is to provide an intelligent vehicle book information processing method and system to solve the above-mentioned problems. It can deeply integrate the advantages of large language models and knowledge graphs, not only have extensive language understanding capabilities, but also can accurately call on professional knowledge to answer complex problems, significantly improving the in-vehicle interactive experience and information acquisition efficiency.

[0008] The present invention proposes an intelligent car book information processing method, comprising the following steps:

[0009] Acquire vehicle book data, wherein the vehicle book data includes text and pictures;

[0010] Use car book data to build a vector database: clean, segment and structure the car book data, map the structured car book data into vectors in a high-dimensional vector space through a large language model, and then import them into the vector database;

[0011] Use car book data to build a graph database: clean the car book data, extract knowledge points according to the predefined concept tree, build a knowledge graph and import it into the graph database;

[0012] After the user enters a question, the large language model is used to search the question in the vector database and then reply, and the knowledge graph is used to search the question in the graph database and then reply;

[0013] The responses of the large language model and knowledge graph are evaluated and arbitrated according to the preset arbitration criteria, and the optimal response is determined and output.

[0014] In one embodiment, after the user inputs a question, the user intention is first determined, that is, whether the user question is a car-book question is determined based on a predefined car-book intention question set;

[0015] If yes, the large language model and knowledge graph are input respectively to retrieve and reply to the user’s question;

[0016] If not, a catch-all response is directly output, and the catch-all response is a universal answer.

[0017] In one embodiment, constructing a vector database using vehicle book data specifically includes:

[0018] Use structured tools to clean the car book data: replace or delete garbled characters and special symbols, and perform ID processing on the image data;

[0019] Segment the cleaned car book data: segment the cleaned text data according to the predetermined length to form multiple Chunk modules;

[0020] Structural processing: Generate structured data in JSON format from the segmented text and the images it contains;

[0021] Vector mapping: Map structured data into vectors in a high-dimensional vector space through a large language model and then import them into the vector database.

[0022] In one embodiment, the atlas database is constructed using the vehicle book data, specifically including:

[0023] Use data extraction tools to clean car book data: remove redundant information, correct erroneous data, and unify data formats;

[0024] Knowledge point extraction: extract knowledge points according to the predefined concept tree;

[0025] Construct a knowledge graph: construct a triple knowledge graph after manually verifying or supplementing the extracted knowledge points, and import it into the graph database.

[0026] In one embodiment, the question is retrieved and answered in the graph database using the knowledge graph, specifically including:

[0027] Match the entities in the user's question with the corresponding entities in the knowledge graph through the entity linking mechanism;

[0028] According to the associated attributes between entities, logical inference is performed through the attribute reasoning mechanism to retrieve positioning related information in the graph database;

[0029] Output a triple that best matches the user's question as the knowledge graph's response to the user's question.

[0030] In one embodiment, a large language model is used to retrieve and reply to questions in a vector database, specifically including:

[0031] Use a large language model to map user questions into a high-dimensional vector space;

[0032] Perform similarity matching in the vector database to select the N vectors that are most similar to the user's question vector and their corresponding car book contents;

[0033] According to the screening results, combined with user questions and context information, based on the set prompt requirements, the retrieved relevant car book content is fine-tuned to generate a large language model to respond to user questions.

[0034] In one embodiment, the large language model also performs security checks on the responses to user questions before evaluation and arbitration, i.e., determining whether the responses involve sensitive content;

[0035] If yes, then output a catch-all response, which is a general answer;

[0036] If not, the large language model's response to the user's question is output.

[0037] The present invention also proposes an intelligent car book information processing system, which is used to implement the intelligent car book information processing method as described above, comprising:

[0038] Input module, input user questions;

[0039] The large language model module maps the car book data into vectors in a high-dimensional vector space and uses the large language model to answer user questions;

[0040] The knowledge graph module builds a knowledge graph based on the car book data and uses the knowledge graph to answer user questions;

[0041] The answer arbitration module evaluates and arbitrates the responses from the large language model and knowledge graph according to the preset arbitration criteria to determine the optimal response;

[0042] Output module, outputs the optimal response.

[0043] In one embodiment, the intelligent car book information processing system also includes an intention classification module, which is a model with billions of parameters and determines whether the user question is a car book question based on a pre-defined car book intention question set.

[0044] In one embodiment, the intelligent car book information processing system also includes a security detection module to detect whether the response of the large language model to the user's question involves sensitive content.

[0045] Compared with the prior art, the intelligent car book information processing method and system of the present invention have the following beneficial effects:

[0046] 1) The present invention provides users with the best answers by combining knowledge graphs and large language models. The large language model, with its powerful language processing capabilities, can deeply analyze the semantics and contextual information of user questions, while the intelligent knowledge graph provides a rich knowledge base in the field of car books, helping the system to quickly locate knowledge points related to user questions. Through the synergy of the two, the system can accurately match the answer that best meets user needs and present it to the user, providing more accurate questions and answers for complex and professional questions in the in-vehicle environment, improving the in-vehicle interactive experience and information acquisition efficiency.

[0047] 2) The catch-all reply in the present invention can provide users with a reasonable and universal answer method, which can make users feel the responsiveness and care of the system even when no specific car book answer can be provided. In this way, accurate services in professional fields are guaranteed, while taking into account the user interaction needs in a wider range of scenarios, ensuring the consistency and integrity of the user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 A schematic diagram of a flow chart of an intelligent car book information processing method according to an embodiment of the present invention;

[0049] Figure 2 The figure is a schematic diagram of the operation steps of the intelligent car book information processing method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0050] In order to make the purpose, technical solution and advantages of the present invention more understandable, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments of the specification. It should be noted that many specific details are set forth in the following description to facilitate a full understanding of the present invention, but the present invention can also be implemented in other ways different from those described herein, and those skilled in the art can make similar generalizations without violating the connotation of the present invention, so the present invention is not limited by the specific embodiments disclosed below.

[0051] Secondly, "one embodiment" or "an embodiment" in this application refers to a specific feature, structure or characteristic that can be included in at least one implementation of the present invention. The "in one embodiment" and "an embodiment" that appear in different places in this specification do not all refer to the same embodiment, nor are they separate or selective embodiments that are mutually exclusive with other embodiments. The terms "including" and "comprising" indicate the presence of the claimed features, but do not exclude the presence of one or more other features. The term "and / or" used in this application includes any and all combinations of one or more of the relevant listed items.

[0052] The present invention proposes an intelligent car book information processing method, see Figure 1 , Figure 2 , including the following steps:

[0053] Get car book data, which includes text, pictures, etc. The data can be in text, PDF, etc.

[0054] Use car book data to build a vector database: clean, segment and structure the car book data, map the structured car book data into vectors in a high-dimensional vector space through a large language model, and then import them into the vector database;

[0055] Use car book data to build a graph database: clean the car book data, extract knowledge points according to the predefined concept tree, build a knowledge graph and import it into the graph database;

[0056] After the user enters a question, the large language model is used to search the question in the vector database and then reply, and the knowledge graph is used to search the question in the graph database and then reply;

[0057] The responses of the large language model and knowledge graph are evaluated and arbitrated according to the preset arbitration criteria, and the optimal response is determined and output.

[0058] The following is a detailed explanation of the intelligent car book information processing method.

[0059] The vector database is constructed by using the car book data through the large language model. It is rigorous and sophisticated, aiming to ensure the accuracy of the data and the efficiency of processing. The specific processing flow includes the following:

[0060] 1) First, use structured tools to perform preliminary cleaning of the car book data. At this stage, special attention should be paid to the garbled characters and special symbols that may appear after the document format conversion, and these unnecessary elements should be replaced or deleted to ensure the purity of the data. At the same time, the image data will be processed with ID identification so that it can be accurately tracked and referenced in subsequent data processing.

[0061] 2) Segment the data after preliminary cleaning. The key to this step is to segment the cleaned plain text data according to the predetermined length to form multiple Chunk modules. Such segmentation helps to use data more efficiently in subsequent processing and improve the model training effect.

[0062] 3) Structural processing: Generate structured data in JSON format from the segmented text paragraphs and the images they contain. This structured data is not only convenient for storage and transmission, but also provides a unified data format for subsequent model processing.

[0063] 4) Vector mapping: Map the structured Chunk module data into vectors in a high-dimensional vector space through a large language model (such as the BERT model, an advanced natural language processing model). In this way, the intrinsic association and semantic information of the data can be better captured, providing strong support for subsequent analysis and prediction.

[0064] 5) Finally, these high-dimensional vectors are imported into the vector database for subsequent retrieval and application.

[0065] The data processing flow of building a knowledge graph using car book data and creating a graph database is as follows:

[0066] 1) First, we rely on data extraction tools, which can efficiently complete the cleaning of vehicle book data to ensure the quality and accuracy of the data. During the cleaning process, the tool will remove redundant information, correct erroneous data, and unify the data format, providing a solid foundation for subsequent processing.

[0067] 2) According to the predefined concept tree, the data extraction tool will accurately extract knowledge points. As the skeleton of the knowledge graph, the concept tree provides clear guidance for knowledge point extraction. Guided by the concept tree, the tool can accurately identify and extract knowledge points related to specific fields, providing rich materials for the construction of the knowledge graph.

[0068] 3) The extracted knowledge points are not directly used for graph construction, but need to be manually verified or supplemented. The importance of this link lies in that it can ensure the accuracy and completeness of the knowledge points, and at the same time, it can make necessary adjustments and supplements to the knowledge points according to actual needs. Through the collaborative work of humans and tools, an accurate and comprehensive triple knowledge graph is constructed.

[0069] 4) Finally, the verified and supplemented triple knowledge graph is imported into the graph database. This step not only realizes the storage and management of knowledge, but also provides strong support for subsequent knowledge retrieval, reasoning and application.

[0070] When answering questions on the car book, after the user enters the question, the user's intention will be determined first, that is, the user's question will be determined based on the pre-defined car book intention question set to determine whether it is a car book question. If so, the large language model and knowledge graph are input respectively to match the answer to the user's question; if not, for example, if the system determines that the question raised by the user is not a car book question and involves security and sensitive content, that is, it does not belong to the pre-defined car book intention question set, the system will not continue to execute the car book-related processing flow, but directly output a catch-all reply. The catch-all reply is a general answer, which is to allow users to feel the system's response and care when no specific car book answer can be provided. The key to this process lies in a model with a billion parameters, which is finely trained with a large amount of manually annotated data, so that it can accurately identify the user's intention. This model predefines a set of car book intention questions, which can efficiently determine whether the user's question belongs to this set. If the question clearly involves the specific configuration, function operation, maintenance and other contents of the vehicle that are clearly recorded in the vehicle manual, it can be identified as a car book question. Examples of car-related consultation include: asking what a certain indicator light on the vehicle's dashboard means, wanting to know the specific items and intervals for regular vehicle maintenance, and asking about the operation of specific vehicle functions, such as how to set the vehicle's cruise control.

[0071] The knowledge graph is used to retrieve and reply to questions in the graph database. Specifically, when the user's question is determined to be a question related to the car book field, the knowledge graph will start the entity linking and attribute reasoning mechanism. Entity linking is responsible for accurately matching the entities in the user's question with the corresponding entities in the knowledge graph, while attribute reasoning makes logical inferences based on the associated attributes between these entities, thereby accurately locating relevant information in the graph database. After this series of retrieval and processing processes, a triple that best matches the user's question will be output as the knowledge graph's reply to the user's question, becoming a candidate answer.

[0072] The large language model is used to retrieve and reply to questions in the vector database. Specifically, once the user question is determined to be a question related to the car book field, the large language model will quickly map the user question to a high-dimensional vector space without tedious structural processing. In this way, similarity matching can be efficiently performed in the vector database, and the N vectors that are most similar to the user question vector and their corresponding car book content can be quickly screened out. After completing the similarity matching, the next step is to fine-tune the large model reasoning and content summarization. In this stage, based on the TOP N results retrieved, combined with the user question and context information, specific prompt requirements are set for the large language model. In this way, the large language model can be fine-tuned according to the retrieved car book content to generate accurate and user-friendly speech. In the large language model ensemble, "prompt" refers to the input text or instruction provided to the model to guide the model to generate a specific type of response. The prompt can be a question, a description, a task description, or even a part of the conversation history. By designing and optimizing the prompt, the model can be guided to generate expected responses or complete specific tasks.

[0073] In order to ensure the security and relevance of the output content, a strict relevance and security detection mechanism is also set up. The large language model's response to the user's question will undergo a security test before evaluation and arbitration, that is, to determine whether the response content involves sensitive content. If so, a fallback response is output, which is a general answer; if not, the large language model's response to the user's question is output. A detailed security and sensitive information review of the output content can avoid any inappropriate content. Using a model with a 1B parameter level, through the preset question and answer pairs, it is judged whether the answer content of the large language model really meets the user's expectations. Any bad content involving pornography and tyranny is blocked in real time to ensure that users receive pure and valuable information.

[0074] The answer arbitration mechanism is to ensure the accuracy and rationality of the answers. This mechanism compares and analyzes multiple potential answers, such as comparing the responses of large language models and knowledge graphs to user questions, and combines preset arbitration rules and strategies to determine whether a normal response should be given directly, or a fallback response should be used to provide a more general or conservative answer. In the answer arbitration stage, the answers provided by the knowledge graph are given priority to ensure that the answers given are the most appropriate for the user's questions and the most accurate and reliable. In this way, we can not only provide users with efficient, accurate, and professional knowledge in the field of car books, but also ensure the user's basic interactive experience through fallback responses when the answer is uncertain or exceeds the coverage of the knowledge graph.

[0075] The following is a detailed explanation of the above-mentioned intelligent car book information processing method with reference to some specific question-and-answer scenario examples.

[0076] Scenario 1:

[0077] Q: What is adaptive cruise control?

[0078] A: Adaptive cruise control helps the vehicle maintain the set speed. When the vehicle approaches the vehicle ahead, adaptive navigation will automatically adjust the speed to keep the vehicle at the set distance.

[0079] Scene 2:

[0080] Q: What is the full-journey intelligent driving assistance system?

[0081] A: The driving assistance system (Travel Assist) combines adaptive cruise control (ACC) with adaptive lane guidance. The vehicle can therefore maintain a driver-preselected distance from the vehicle ahead within the system limits and maintain a priority position within the lane + corresponding picture.

[0082] Scenario 3:

[0083] Q: Where can I buy discounted accessories for this car? (Non-car book related question, please reply in the Q&A system)

[0084] A: I cannot answer your question for now. Please change the question.

[0085] The following uses the second scenario of the car book problem as an example to explain the process of the above intelligent car book information processing method:

[0086] For this problem, the user intention is determined first:

[0087] 1.0 Intent classification: The user's question is determined by the 1B parameter level intent classification model to determine whether it belongs to the predefined car book intent question set;

[0088] Input: User question text "What is the full journey intelligent driving assistance system?"

[0089] Processing: Use a 1B parameter level model and train it with manually annotated data. This model has predefined some sets of car book-related intent questions and determines whether the user's question belongs to these predefined sets.

[0090] Output: intent classification result (determine whether the user's question belongs to the car book question and answer intent) + user question text;

[0091] Then we will enter the large language model retrieval and knowledge graph retrieval process.

[0092] The knowledge graph retrieval process is as follows:

[0093] 2.1.1 Knowledge Graph - Entity Linking & Attribute Reasoning

[0094] Input: User question text “What is the full journey intelligent driving assistance system?”

[0095] Processing: Entity linking accurately matches the entities in the user's question with the corresponding entities in the knowledge graph, while attribute reasoning makes logical inferences based on the associated attributes between these entities;

[0096] Output: Entity information of the user’s question [Full-journey intelligent driving assistance system] and associated attributes [Usage method] [Precautions].

[0097] 2.1.2 Knowledge Graph-Graph Database

[0098] Input: Entity information of user question [Full-journey intelligent driving assistance system] and associated attributes [Usage method] [Precautions]

[0099] Processing: Accurately locate relevant information in the knowledge graph database;

[0100] Output: A triple that best matches the user's question as a candidate answer.

[0101] #EntityFull-journey Intelligent Driving Assistance System

[0102] #Attribute usage

[0103] #Plain text attribute value

[0104] #Image and video attribute values

[0105] 2.1.3 Knowledge Graph-Triple Attribute Value Extraction

[0106] Input: A triple that best matches the user's question

[0107] Processing: Extract the corresponding triple attribute value information from the knowledge graph database;

[0108] Output: Generate answers based on the text and video image information corresponding to the triple.

[0109] Knowledge graph search answer A1: The Travel Assist system automatically controls the accelerator pedal, brakes and steering system. In addition, the Travel Assist system follows the vehicle ahead in the lane and decelerates within the system limit until it stops and starts again. You can take over the adjustment of the Travel Assist system at any time.

[0110] The retrieval response process of the large language model is as follows:

[0111] 2.2.1 Large Language Model - BERT Model

[0112] Input: User question text “What is the full journey intelligent driving assistance system?”

[0113] Processing: Use the BERT model to convert user questions into vector representations (Embedding);

[0114] Output: User question vector representation [0.57 0.51 0.25 0.31 0.36…].

[0115] 2.2.2 Large Language Model-Vector Database

[0116] Input: user question vector representation [0.57 0.51 0.25 0.31 0.36…];

[0117] Processing: Use vector similarity matching (such as cosine similarity) to retrieve the N questions and their contexts that are most similar to the user's question from the database;

[0118] cos([0.57 0.51 0.25 0.31 0.36…],[0.62 0.35 0.12 0.43 0.56…])=0.05

[0119] cos([0.57 0.51 0.25 0.31 0.36…],[0.83 0.52 0.32 0.35 0.46…])=0.65

[0120] Sort and select the N vectors that are most similar to the user's question and their corresponding car book content

[0121] Output: retrieval results (Top N similar questions and their context).

[0122] Chunk 2:{para1:Full-journey intelligent driving assistance system,para2:Topic introduction,para3:Driving assistance system will adapt to...,pictureId:①};

[0123] Chunk 3:{para1:Full-journey intelligent driving assistance system,para2:Topic introduction,para3:Full-journey intelligent driving assistance system (Travel Assist) uses…,pictureId:};

[0124] Q: What is the full-journey intelligent driving assistance system?

[0125] 2.2.3 Large Language Model - Fine-tuning Large Model Reasoning and Content Summarization

[0126] Input: search results (Top N similar questions and their context);

[0127] Processing: Given a prompt requirement, guide the large language model to generate corresponding words based on the retrieved car book content;

[0128] Output: Generated answer content;

[0129] Retrieval response of the large language model A2: The driving assistance system (Full Journey Intelligent Driving Assistance System (TravelAssist)) combines adaptive cruise control (ACC) with adaptive lane guidance. The vehicle can therefore maintain a driver-preselected distance from the vehicle in front within the system limits and maintain a priority position in the lane + corresponding picture.

[0130] 2.2.4 Large Language Model-Relevance and Security Testing

[0131] Input: Generated answer content;

[0132] Processing: Use the 1B parameter level model to detect the relevance and security of the answer content, and use the preset question and answer pairs to determine whether the answer content of the large model meets the user's expectations, and block the Huang Baozheng answer content;

[0133] Output: detection results (non-pornographic and tyrannical content) + the answer content A2 expected to be generated by the large model;

[0134] Finally, the responses from the large language model and knowledge graph are evaluated and arbitrated:

[0135] 3 Answers Arbitration

[0136] Input: Generated answer content;

[0137] Processing: By comparing and analyzing multiple potential answers and combining them with preset arbitration rules and strategies, it is determined whether a normal response should be given directly or a more general or conservative answer should be provided by using a fallback response;

[0138] Output: According to the arbitration result, A2 is determined to be the answer that better matches the user's question, and the arbitration result is output;

[0139] The best answer is A2: The driving assistance system (Travel Assist) combines adaptive cruise control (ACC) with adaptive lane guidance. The vehicle can therefore maintain a driver-preselected distance from the vehicle ahead within the system limits and maintain a priority position within the lane + corresponding picture.

[0140] The following uses the non-car book problem scenario three as an example to explain the process of the above-mentioned intelligent car book information processing method:

[0141] For this problem, the user intention is determined first:

[0142] 1.0 Intent classification: The user question is determined to be within the predefined set through the 1B parameter level intent classification model;

[0143] Input: User question text "Where can I buy discounted accessories for this car?"

[0144] Processing: Use a 1B parameter level model and train it with manually annotated data. This model has predefined some sets of car book-related intent questions and determines whether the user's question belongs to these predefined sets.

[0145] Output: intent classification result (determine whether the user's question belongs to the car book question and answer intent) + user question text;

[0146] 2.0 Back-up reply:

[0147] Input: The text of a user question that is not related to car books: "Where can I buy discounted accessories for this car?"

[0148] Processing: Determine the fallback response based on preset strategies and rules. These rules are usually based on a variety of factors, such as the type of question, field, common situations, etc. During the processing, some general principles may be considered to generate a relatively general or conservative answer.

[0149] Output: Give a pre-set catch-all response. This response is usually broad and general, intended to provide some basic information or guidance rather than a precise answer to a specific question.

[0150] Final reply A3: I cannot answer your question for now. Please change the question.

[0151] The present invention also proposes an intelligent vehicle book information processing system, which is used to implement the intelligent vehicle book information processing method as described above, including an input module, a large language model module, a knowledge graph module, an answer arbitration module, an output module, an intent classification module and a security detection module.

[0152] The input module is used to input user questions. The intent classification module is connected to the input module. It is a model with billions of parameters. It determines whether the user question is a car book question based on the pre-defined car book intent question set. The large language model module is connected to the intent classification module, maps the car book data into vectors in a high-dimensional vector space, and uses the large language model to reply to user questions. The security detection module is connected to the large language model module to detect whether the large language model's reply to user questions involves sensitive content. The knowledge graph module is connected to the intent classification module, builds a knowledge graph based on the car book data, and uses the knowledge graph to reply to user questions. The answer arbitration module is connected to the security detection module and the knowledge graph module respectively, and evaluates and arbitrates the replies of the large language model and the knowledge graph according to the preset arbitration criteria to determine the optimal reply. The output module is connected to the answer arbitration module to output the optimal reply.

[0153] The present invention has the following beneficial effects:

[0154] 1) The present invention provides users with the best answers by combining knowledge graphs and large language models. The large language model, with its powerful language processing capabilities, can deeply analyze the semantics and contextual information of user questions, while the intelligent knowledge graph provides a rich knowledge base in the field of car books, helping the system to quickly locate knowledge points related to user questions. Through the synergy of the two, the system can accurately match the answer that best meets user needs and present it to the user, providing more accurate questions and answers for complex and professional questions in the in-vehicle environment, improving the in-vehicle interactive experience and information acquisition efficiency.

[0155] 2) The catch-all reply in the present invention can provide users with a reasonable and universal answer method, which can make users feel the responsiveness and care of the system even when no specific car book answer can be provided. In this way, accurate services in professional fields are guaranteed, while taking into account the user interaction needs in a wider range of scenarios, ensuring the consistency and integrity of the user experience.

[0156] The construction and arrangement of the present application shown in a plurality of different exemplary embodiments are only exemplary. Although only a few embodiments are described in detail in this disclosure, it should be readily understood by those who refer to this disclosure that many modifications are possible (e.g., parameter values, mounting arrangements, changes in orientation, etc.) without substantially departing from the novel teachings and advantages of the subject matter described in the application. For example, the element shown as integrally formed can be composed of a plurality of parts or elements, the position of the element can be inverted or otherwise changed, and the nature or number or position of the discrete element can be changed or changed. Therefore, all such modifications are intended to be included within the scope of the present invention. The order or sequence of any process or method steps can be changed or reordered according to alternative embodiments. In the claims, any "device plus function" clause is intended to cover the structure of performing the function described herein, and is not only structurally equivalent but also equivalent structure. Without departing from the spirit and essence of the present invention, a person skilled in the art can make various corresponding modifications and changes according to the present invention, but these corresponding modifications and changes should all fall within the scope of protection of the present invention.

Claims

1. An intelligent car book information processing method, It is characterized in that The steps include: Acquire vehicle book data, wherein the vehicle book data includes text and pictures; Use car book data to build a vector database: clean, segment and structure the car book data, map the structured car book data into vectors in a high-dimensional vector space through a large language model, and then import them into the vector database; Use car book data to build a graph database: clean the car book data, extract knowledge points according to the predefined concept tree, build a knowledge graph and import it into the graph database; After the user enters a question, the large language model is used to search the question in the vector database and then reply, and the knowledge graph is used to search the question in the graph database and then reply; The responses of the large language model and knowledge graph are evaluated and arbitrated according to the preset arbitration criteria, and the optimal response is determined and output.

2. The intelligent car book information processing method according to claim 1, It is characterized in that After the user inputs a question, the user intent is first determined, that is, whether the user question is a car-book question is determined based on the pre-defined car-book intent question set; If yes, the large language model and knowledge graph are input respectively to retrieve and reply to the user’s question; If not, a catch-all response is directly output, and the catch-all response is a universal answer.

3. The intelligent car book information processing method according to claim 1, It is characterized in that Use the car book data to build a vector database, including: Use structured tools to clean the car book data: replace or delete garbled characters and special symbols, and perform ID processing on the image data; Segment the cleaned car book data: segment the cleaned text data according to the predetermined length to form multiple Chunk modules; Structural processing: Generate structured data in JSON format from the segmented text and the images it contains; Vector mapping: Map structured data into vectors in a high-dimensional vector space through a large language model and then import them into the vector database.

4. The intelligent car book information processing method according to claim 1, It is characterized in that Use car book data to build a graph database, including: Use data extraction tools to clean car book data: remove redundant information, correct erroneous data, and unify data formats; Knowledge point extraction: extract knowledge points according to the predefined concept tree; Construct a knowledge graph: construct a triple knowledge graph after manually verifying or supplementing the extracted knowledge points, and import it into the graph database.

5. The intelligent car book information processing method according to claim 4, It is characterized in that Use the knowledge graph to retrieve and reply to questions in the graph database, including: Match the entities in the user's question with the corresponding entities in the knowledge graph through the entity linking mechanism; According to the associated attributes between entities, logical inference is performed through the attribute reasoning mechanism to retrieve positioning related information in the graph database; Output a triple that best matches the user's question as the knowledge graph's response to the user's question.

6. The intelligent car book information processing method according to claim 1, It is characterized in that Use a large language model to retrieve and reply to questions in a vector database, including: Use a large language model to map user questions into a high-dimensional vector space; Perform similarity matching in the vector database to select the N vectors that are most similar to the user's question vector and their corresponding car book contents; According to the screening results, combined with user questions and context information, based on the set prompt requirements, the retrieved relevant car book content is fine-tuned to generate a large language model to respond to user questions.

7. The intelligent car book information processing method according to claim 6, It is characterized in that The large language model also conducts security checks on responses to user questions before evaluation and arbitration, that is, to determine whether the responses involve sensitive content; If yes, then output a catch-all response, which is a general answer; If not, the large language model's response to the user's question is output.

8. An intelligent car book information processing system, It is characterized in that The method for implementing the intelligent car book information processing method according to any one of claims 1 to 7 comprises: Input module, input user questions; The large language model module maps the car book data into vectors in a high-dimensional vector space and uses the large language model to answer user questions; The knowledge graph module builds a knowledge graph based on the car book data and uses the knowledge graph to answer user questions; The answer arbitration module evaluates and arbitrates the responses from the large language model and knowledge graph according to the preset arbitration criteria to determine the optimal response; Output module, outputs the optimal response.

9. The intelligent car book information processing system according to claim 8, It is characterized in that It also includes an intention classification module, which is a model with billions of parameters and determines whether the user's question is a car-book question based on a pre-defined car-book intention question set.

10. The intelligent car book information processing system according to claim 8, It is characterized in that It also includes a security detection module to detect whether the large language model's responses to user questions involve sensitive content.

Images