Automobile information security multi-granularity knowledge recall method based on problem understanding

Abstract

The application discloses a kind of based on problem understanding's automobile information security multi-granularity knowledge recall method, based on the technical document of automobile information security field generates document level knowledge unit, chapter level knowledge unit and content block level knowledge unit, shape multi-granularity knowledge index system, while constructing theme tree structure;The semantic analysis of user input question is carried out, and core object, constraint condition, answering intention and question type are extracted, based on core object retrieval in document level index obtains candidate document set, utilizes constraint condition to the screening of candidate document obtains candidate chapter set, then again to content block screening obtains initial evidence set;Based on theme tree, the theme expansion of initial evidence obtains expansion evidence set, and the union of initial evidence set and expansion evidence set is screened to obtain the initial candidate evidence set, to generate knowledge recall result.The application can effectively improve the relevance, integrity and adaptability of candidate evidence, to improve knowledge recall performance.

Description

Technical Field

[0001] This invention belongs to the field of automotive information security technology, and more specifically, relates to a multi-granularity knowledge retrieval method for automotive information security based on problem understanding. Background Technology

[0002] With the development of intelligent connected vehicle technology, automotive systems are gradually evolving from traditional closed electronic and electrical systems into open and complex systems that integrate functions such as in-vehicle terminals, in-vehicle networks, remote communication, cloud platforms, mobile applications, and online upgrades. Related automotive information security knowledge also exhibits characteristics such as dispersed sources, diverse types, complex structures, and inconsistent semantic granularity.

[0003] Existing automotive cybersecurity knowledge is typically distributed across various sources, including national standards, industry specifications, corporate technical documents, regulatory guidelines, vulnerability intelligence, security test reports, and threat analysis case studies. Traditional retrieval methods for addressing user-defined questions usually employ keyword retrieval, full-text retrieval, or vector retrieval to return several text fragments from a document collection. However, these methods have the following shortcomings in complex automotive cybersecurity question-and-answer scenarios:

[0004] First, existing retrieval methods typically employ a uniform retrieval strategy, failing to adequately differentiate between question types. For terminology explanation questions, rule comprehension questions, chart / graph explanation questions, and cross-document comprehensive questions, the required knowledge granularity and evidence forms differ. Using the same retrieval strategy can easily lead to a mismatch between the retrieved results and the question's requirements.

[0005] Second, existing retrieval methods mostly focus on single text blocks as the primary retrieval object, lacking collaborative retrieval mechanisms between document-level, chapter-level, content block-level, and topic-level searches. When the question involves standard clauses, chapter context, or knowledge relating multiple documents, single-granularity retrieval can easily lead to a lack of context.

[0006] Third, much of the content in automotive information security knowledge has a hierarchical structure and thematic relationships. For example, a certain security requirement may be distributed across multiple chapters, and a certain technical object may involve multiple topics such as communication security, software upgrades, vulnerability management, and access control. Existing methods struggle to supplement and recall topics based on the semantics of the problem.

[0007] Fourth, existing search results lack an evidentiary framework for generating subsequent questions and answers, which can easily lead to problems such as overly broad recall, omission of key constraints, lack of chart-based knowledge, and insufficient cross-document evidence. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a multi-granularity knowledge retrieval method for automotive information security based on problem understanding. By performing structured semantic parsing on the input question and integrating document-level, chapter-level, content block-level, and topic-level knowledge indexes, a candidate evidence set for complex automotive information security question answering is formed. This improves the relevance, completeness, and adaptability of the candidate evidence, thereby solving the problems of single granularity of retrieval, insufficient adaptation of question types, omission of cross-document evidence, and weak topic supplementation ability of existing retrieval methods, and improving knowledge retrieval performance.

[0009] To achieve the above-mentioned objectives, the present invention provides a multi-granularity knowledge retrieval method for automotive information security based on problem understanding, comprising the following steps:

[0010] S1: Obtain several technical documents in the field of automotive information security. For each technical document, perform structured parsing to extract the document title, chapter path, main text content, chart information, and metadata. Generate document-level knowledge units, chapter-level knowledge units, and content block-level knowledge units according to the document structure. Establish keyword indexes, semantic vector indexes, and metadata indexes for each type of knowledge unit to form a knowledge index system that supports multi-granularity retrieval. At the same time, extract topic information from the technical documents to obtain topic-level knowledge units, construct a topic tree structure, and mark the mapping between topic-level knowledge units and content block-level knowledge units.

[0011] S2: Use a large language model to perform semantic analysis on user input questions, extract four components: core object O, constraint C, answer intent I, and question type T, to form a structured question representation Q=(O,C,I,T);

[0012] S3: Retrieve from the document-level index based on the core object O to filter out a set of candidate documents related to the user's input question;

[0013] By using constraint C to filter the chapters of all documents in the candidate document set, a candidate chapter set is obtained;

[0014] For each content block in the candidate chapter set, calculate the content similarity between the user input question and the content block, and filter to obtain the initial evidence set. ;

[0015] S4: Obtaining the initial set of evidence Next, the topics associated with each content block are obtained, the frequency of each topic is counted, and the topics with higher frequencies are selected to form a set of topics to be expanded. Then, for each topic in the set of topics to be expanded, the topic nodes with superior, subordinate, and related relationships to that topic are obtained in the topic tree, and the content blocks corresponding to these topics are obtained to form an expanded set of evidence. ;

[0016] S5: Initial set of evidence With extended evidence set The evidence set is obtained by merging. After screening, the final set of candidate evidence is obtained. ;

[0017] S6: Based on the final candidate evidence set Generate knowledge retrieval results, with each result including the source document, chapter path, topic tags, and relevance information of the content block corresponding to the evidence.

[0018] This invention presents a multi-granularity knowledge retrieval method for automotive information security based on problem understanding. It generates document-level, chapter-level, and content block-level knowledge units from technical documents in the automotive information security field, forming a multi-granularity knowledge index system and constructing a topic tree structure. Semantic analysis is performed on user-input questions to extract core objects, constraints, answer intent, and question type. A candidate document set is obtained by retrieving core objects from the document-level index. Constraints are used to filter candidate documents to obtain a candidate chapter set, and then content blocks are filtered to obtain an initial evidence set. The initial evidence is expanded using the topic tree to obtain an extended evidence set. Finally, the union of the initial and extended evidence sets is filtered to obtain the initial candidate evidence set, thereby generating the knowledge retrieval result.

[0019] The present invention has the following beneficial effects:

[0020] 1) This invention performs structured understanding of the question before retrieval, which can identify the core objects, constraints, answer intentions and question types in the question, thereby avoiding semantic deviation caused by direct retrieval;

[0021] 2) This invention combines document-level, chapter-level, content block-level, and topic-level knowledge indexes for multi-granularity retrieval, which can simultaneously take into account document source, chapter context, specific evidence fragments, and cross-document topic associations;

[0022] 3) This invention, through a topic supplementation and recall mechanism, can discover related topics, superior topics, subordinate topics, and cross-document topic knowledge related to candidate content blocks, thereby reducing the omission of key evidence in complex question answering.

[0023] 4) This invention can be effectively applied to professional Q&A scenarios in the field of automotive information security, where standards and specifications are complex, technical objects are diverse, knowledge is scattered, and question types are varied. Attached Figure Description

[0024] Figure 1 This is a flowchart illustrating a specific implementation of the multi-granularity knowledge retrieval method for automotive information security based on problem understanding, as described in this invention.

[0025] Figure 2 This is an example diagram of the topic tree structure in this embodiment;

[0026] Figure 3 This is a distribution diagram of technical document types in this embodiment;

[0027] Figure 4 This is a distribution diagram of the problem data types in this embodiment;

[0028] Figure 5 This is a comparison chart of the experimental results of the present invention and the comparative method in this embodiment. Detailed Implementation

[0029] The specific embodiments of the present invention will now be described with reference to the accompanying drawings to enable those skilled in the art to better understand the invention. It should be particularly noted that in the following description, detailed descriptions of known functions and designs that might obscure the main content of the invention will be omitted here.

[0030] Example

[0031] Figure 1 This is a flowchart illustrating a specific implementation of the multi-granularity knowledge retrieval method for automotive information security based on problem understanding, as described in this invention. Figure 1 As shown, this invention presents a multi-granularity knowledge retrieval method for automotive information security based on problem understanding:

[0032] S101: Multi-granularity retrieval index construction:

[0033] To achieve a balance between recall scope, location accuracy, and evidence integrity, this invention constructs a three-tiered retrieval index—document-level, chapter-level, and content block-level—for the automotive information security knowledge base. A unified metadata association mechanism connects these different granularity indexes, forming a hierarchical knowledge access structure. The specific method for constructing this multi-granularity retrieval index is as follows:

[0034] This embodiment acquires several technical documents in the field of automotive information security. These documents include standards, regulations, guidelines, technical documents, vulnerability intelligence, and security case studies related to automotive information security. Each document undergoes structured parsing to extract its title, chapter path, main text content, charts, and metadata. Based on the document structure, document-level knowledge units, chapter-level knowledge units, and content block-level knowledge units are generated. For each type of knowledge unit, a keyword index, a semantic vector index, and a metadata index are established, forming a knowledge index system that supports multi-granularity retrieval. Generally, keywords are generated using a keyword extraction algorithm, and semantic vectors are obtained through encoding using a pre-trained text vector model.

[0035] Simultaneously, this invention also requires extracting topic information from technical documents to obtain topic-level knowledge units, constructing a topic tree structure, thereby hierarchically organizing the semantic attribution of knowledge, and annotating the mapping between topic-level knowledge units and content block-level knowledge units, providing an organizational foundation for subsequent topic expansion and retrieval. The specific method for constructing the topic tree is as follows:

[0036] The process involves collecting keywords from all technical documents as candidate topics. These topics, while semantically consistent, are merged into a unified representation, forming a standardized candidate topic set. Topics with frequencies below a threshold or generalizations above a threshold are then removed, leaving a leaf topic set. Subsequently, for each content block-level knowledge unit, the most relevant topic from the leaf topic set is selected as its primary topic, establishing a mapping from content blocks to topics. Based on the semantic similarity between leaf topic names and their co-occurrence within chapters, leaf topics with similar semantics and strong shared structural features are merged to form parent topics. This results in a two-layer topic structure: "parent topic – leaf topic," leading to a topic tree. For example, "software upgrade security," "OTA upgrade security," and "remote upgrade security" are merged into a unified topic node, while "software integrity protection" and "upgrade package integrity verification" are merged into their subordinate topics. Figure 2 This is an example diagram of the topic tree structure in this embodiment.

[0037] As can be seen, the present invention ultimately forms a multi-granularity knowledge index system consisting of documents, chapters, content blocks, and topic nodes.

[0038] S102: Structured Analysis of the Problem:

[0039] A large language model is used to perform semantic analysis on user input questions, extracting four components: core object O, constraints C, answer intent I, and question type T, forming a structured question representation Q=(O,C,I,T). Here, core object O represents the technical object of interest in the question, constraints C limits the search scope and constraint coverage, answer intent I determines the desired form of knowledge content, and question type T determines the subsequent knowledge retrieval path and search focus. In this embodiment, question types include terminology explanation, rule understanding, diagram explanation, and cross-document synthesis, thus providing a technical foundation for multi-granularity retrieval of different question types.

[0040] In practical applications, the core objects and constraints can be encoded separately to obtain corresponding keywords and vector representations, which can be used for multi-granularity recall in subsequent retrieval.

[0041] In this embodiment, taking the question "Based on relevant domestic intelligent connected vehicle specifications, what measures should manufacturers take to ensure the reliability of the upgrade process for in-vehicle systems with online upgrade capabilities? If upgrade-related vulnerabilities are discovered, how should they be handled subsequently?" as an example, the core object obtained through structured parsing is "in-vehicle systems with online upgrade capabilities"; the constraints include "reliability of the upgrade process" and "subsequent handling of vulnerabilities"; the intended answer is "explanation of measures"; and the question type is "cross-document comprehensive".

[0042] S103: Multi-granularity collaborative retrieval:

[0043] After obtaining the structured question representation, a coarse-grained recall is first performed in the documents using the core object as the retrieval condition. This involves searching the document-level index based on the core object O to filter out a set of candidate documents related to the user's input question. The goal of this operation is to exclude documents irrelevant to the question's topic as much as possible, reducing cross-document noise. In this embodiment, the core object retrieval is performed by fusing keyword matching scores and semantic similarity scores to calculate a matching score, and documents with higher matching scores are selected to form the candidate document set. (Matching score) The calculation formula is:

[0044] ,

[0045] in, This represents the semantic similarity between the core object O and the document. This indicates the keyword matching degree between the core object O and the document. Indicates the weight.

[0046] Subsequently, relevant chapters are located in the chapter-level index using constraint C. Specifically, constraint C is used to filter the chapters of all documents in the candidate document set, resulting in a candidate chapter set. Each chapter node simultaneously records the chapter title, hierarchical path, keyword set, and chapter semantic representation, thus enabling joint matching based on title semantics, structural position, and local content theme. Chapter-level filtering avoids unconstrained searching across all knowledge blocks, improving the targeting of subsequent block-level retrieval.

[0047] After chapter selection, the system proceeds to the content block-level recall stage. For each content block of all candidate chapters in the candidate chapter set, the content similarity between the user input question and the content block is calculated, and an initial evidence set is obtained. Content similarity in this embodiment The calculation is based on a combination of semantic similarity, keyword coverage, and constraint matching degree. The specific calculation formula is as follows:

[0048] ,

[0049] in, This indicates the semantic similarity between the user's input question and the content block. This indicates the keyword coverage of the content block for the user's input question. This indicates the degree of match between the constraints of the user's input question and the content block. , , Indicates the weight.

[0050] As can be seen from the above process, this multi-granularity collaborative retrieval process fully combines the semantics of the question, the constraints, and the types of knowledge units to achieve collaborative retrieval of knowledge at the document level, chapter level, and content block level, ensuring that the initial evidence set covers the core objects and constraints of the question while taking into account the continuity of the context.

[0051] S104: Topic Expansion:

[0052] In obtaining the initial set of evidence Next, the topics associated with each content block are obtained, the frequency of each topic is counted, and the topics with higher frequencies are selected to form a set of topics to be expanded. Then, for each topic in the set of topics to be expanded, the topic nodes with superior, subordinate, and related relationships to that topic are obtained in the topic tree, and the content blocks corresponding to these topics are obtained to form an expanded set of evidence. By expanding the topic-based recall, the system can supplement the knowledge content that was not directly hit in the initial multi-granularity retrieval stage, such as upgrade failure recovery, security log auditing, and vulnerability lifecycle management, thereby improving the coverage of the candidate evidence set with the problem constraints and enhancing the relevance of cross-document and cross-topic knowledge.

[0053] In this embodiment, the topics "software upgrade security" and "vulnerability management" appear most frequently, so these topics are used as central nodes for expansion. First, the parent topic is retrieved, such as "vehicle network security management"; then, its child topics are retrieved, such as "upgrade package signature verification," "upgrade failure recovery," and "patch management"; finally, related topics at the same level are retrieved, such as "security audit" and "incident response." Compared to the initial recall, the newly added evidence includes upgrade rollback mechanisms, security log retention requirements, vulnerability lifecycle management requirements, and vulnerability remediation verification requirements. Through the topic expansion mechanism, related knowledge that was not directly retrieved is supplemented, improving evidence coverage.

[0054] S105: Evidence Screening

[0055] For the initial set of evidence With extended evidence set The evidence set is obtained by merging. After screening, the final set of candidate evidence is obtained. The scoring method used in the evidence screening in this embodiment. The calculation formula is as follows:

[0056] ,

[0057] in, Present evidence The semantic relevance of the content block to the question representation Q. Present evidence The coverage of the content block to constraint C, Present evidence The degree of matching between the content block and the question type T. , , Indicates the weight.

[0058] The evidence is ranked using the above scoring system, and the evidence with the highest scores is selected to generate the final candidate evidence set. For example, evidence involving upgrade package signature verification covers the "upgrade trustworthiness" constraint, while evidence involving vulnerability response procedures covers the "vulnerability handling" constraint, thus earning a higher score. This approach ensures that the output simultaneously meets the requirements of relevance, completeness, and diversity, providing high-quality, traceable evidence for subsequent question-and-answer generation.

[0059] S106: Generate knowledge retrieval results:

[0060] Based on the final candidate evidence set Generate knowledge retrieval results, with each result including the source document, chapter path, topic tags, and relevance information of the content block corresponding to the evidence.

[0061] In this embodiment, the final output covers multiple aspects, including upgrade package integrity verification, identity authentication mechanism, security log recording, vulnerability monitoring, vulnerability response, and vulnerability remediation verification. This comprehensively supports the user's two objectives: "upgrade credibility assurance" and "follow-up vulnerability handling." The resulting candidate evidence set not only reflects the direct knowledge related to the question but also supplements related knowledge through a topic expansion mechanism, improving the completeness and accuracy of knowledge retrieval in complex automotive information security question-and-answer scenarios. It also provides a traceable evidentiary basis for subsequent answer generation. The generated knowledge retrieval results can be directly used by the knowledge question-and-answer module to generate answers, perform knowledge analysis, or assist in intelligent reasoning tasks, realizing a complete technical process from question understanding and retrieval to evidence output.

[0062] To better illustrate the technical effects of the present invention, specific examples are used to experimentally verify the invention. In this embodiment, a total of 645 technical documents related to automotive information security were collected. Figure 3This is a distribution diagram of technical document types in this embodiment. For example... Figure 3 As shown, the technical documents in this embodiment cover various types, including standard documents, technical documents, specifications and guidelines, research reports, and laws and regulations. These documents contain both normative descriptions of automotive information security concepts, processes, and control requirements, as well as technical explanations of in-vehicle system architecture, communication mechanisms, and security issues, providing document-level knowledge support for complex question-and-answer sessions.

[0063] This embodiment constructs an evaluation dataset of 500 questions, reference answers, and corresponding evidence blocks, categorizing the questions into four types: terminology definition, rule understanding, chart interpretation, and cross-document questions. Different construction strategies are employed for different question types during data construction: For relatively well-structured questions such as terminology definition and rule understanding, candidate questions are first generated using a large language model based on domain documents, and corresponding evidence fragments are automatically located based on the document content, followed by manual review and correction. For chart interpretation questions, manual supplementation and verification are performed using chart titles, table titles, adjacent explanatory text, and related textual knowledge blocks to ensure accurate expression of chart semantics. For cross-document questions, the design is primarily undertaken by researchers with relevant backgrounds, involving joint annotation of multi-document evidence to reflect the question's requirements for cross-document knowledge integration and multi-hop reasoning capabilities. Figure 4 This is a distribution diagram of the problem data types in this embodiment.

[0064] Table 1 is the experimental environment setup table for this embodiment.

[0065]

[0066] Table 1

[0067] This embodiment sets up three types of methods for comparison with the method presented in this paper. The first type is BM25 retrieval generation, which uses keyword matching to return relevant document fragments and then directly inputs the results into the generation model to output the answer. The second type is vector retrieval RAG, which uses semantic vector retrieval to return relevant knowledge blocks and directly concatenates the candidate results before inputting them into the generation model, representing conventional semantic retrieval question answering methods. The third type is multi-granularity retrieval generation, which employs a document-level, chapter-level, and knowledge block-level multi-granularity hybrid recall mechanism consistent with this paper during the retrieval stage, but does not construct question-related subgraphs or linearize evidence; instead, it directly concatenates the returned results before inputting them into the generation model, used to examine the contribution of multi-granularity recall itself. The evaluation metrics used in this embodiment include:

[0068] Evidence recall: measures the extent to which the set of evidence returned by the system covers the reference evidence.

[0069] Precision: This measures the proportion of relevant evidence in the returned evidence.

[0070] F1-score: Used to comprehensively measure the balance between evidence recall and evidence precision.

[0071] Figure 5 This is a comparison chart of the experimental results of the present invention and the comparative method in this embodiment. For example... Figure 5 As shown, different methods show a trend of gradual improvement in various indicators from traditional retrieval methods to structure enhancement methods. However, there are certain differences between the methods in terms of evidence precision and recall, reflecting the characteristics of different retrieval and evidence organization strategies.

[0072] First, regarding the BM25 retrieval generation method, it did not lag significantly behind in Precision, but its Recall was at the lowest level. This indicates that while keyword-based methods can return some relevant evidence, their lack of semantic modeling capabilities makes it difficult to cover information with distorted expressions or scattered across documents, leading to missing key evidence and affecting the completeness of the answer. Therefore, this method is limited in complex question-answering scenarios.

[0073] Compared to BM25, the vector retrieval RAG showed significant improvements across all three metrics, particularly in Recall. This indicates that semantic vector retrieval can effectively mitigate the expression bias caused by literal matching and recall more semantically relevant evidence. However, the Precision metric shows a relatively limited improvement, suggesting that while expanding the recall scope, it also introduces a certain proportion of weakly relevant evidence, thus affecting the quality of the evidence to some extent.

[0074] Further introducing a multi-granularity retrieval mechanism reveals simultaneous improvements in both precision and recall, resulting in superior performance on the F1-score. This demonstrates that hierarchical recall at the document, chapter, and knowledge block levels can improve the quality of initial candidate evidence while controlling the search space. Specifically, document-level and chapter-level indexes effectively filter irrelevant content and reduce cross-document noise, while knowledge block-level indexes ensure fine-grained semantic matching capabilities, thereby alleviating the trade-off between recall and precision to some extent.

[0075] In comparison, the method of this invention achieves superior performance across all metrics. Looking at the relationships between the metrics, the method of this invention maintains high precision while improving recall, resulting in the highest F1-score. This indicates that its performance improvement does not solely rely on expanding the scale of evidence, but also stems from a more effective evidence screening and organization mechanism, fully validating the effectiveness and superiority of the technical solution of this invention.

[0076] Although the illustrative specific embodiments of the present invention have been described above to enable those skilled in the art to understand the invention, it should be understood that the invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the invention as defined and determined by the appended claims, and all inventions utilizing the concept of the present invention are protected.

Claims

1. A multi-granularity knowledge retrieval method for automotive information security based on problem understanding, characterized in that... include: S1: Obtain several technical documents in the field of automotive information security. For each technical document, perform structured parsing to extract the document title, chapter path, main text content, chart information, and metadata. Generate document-level knowledge units, chapter-level knowledge units, and content block-level knowledge units according to the document structure. Establish keyword indexes, semantic vector indexes, and metadata indexes for each type of knowledge unit to form a knowledge index system that supports multi-granularity retrieval. At the same time, extract topic information from the technical documents to obtain topic-level knowledge units, construct a topic tree structure, and mark the mapping between topic-level knowledge units and content block-level knowledge units. S2: Use a large language model to perform semantic analysis on user input questions, extract four components: core object O, constraint C, answer intent I, and question type T, to form a structured question representation Q=(O,C,I,T); S3: Retrieve from the document-level index based on the core object O to filter out a set of candidate documents related to the user's input question; By using constraint C to filter the chapters of all documents in the candidate document set, a candidate chapter set is obtained; For each content block in the candidate chapter set, calculate the content similarity between the user input question and the content block, and filter to obtain the initial evidence set. ; S4: Obtaining the initial set of evidence Next, the topics associated with each content block are obtained, the frequency of each topic is counted, and the topics with higher frequencies are selected to form a set of topics to be expanded. Then, for each topic in the set of topics to be expanded, the topic nodes with superior, subordinate, and related relationships to that topic are obtained in the topic tree, and the content blocks corresponding to these topics are obtained to form an expanded set of evidence. ; S5: Initial set of evidence With extended evidence set The evidence set is obtained by merging. After screening, the final set of candidate evidence is obtained. ; S6: Based on the final candidate evidence set Generate knowledge retrieval results, with each result including the source document, chapter path, topic tags, and relevance information of the content block corresponding to the evidence.

2. The multi-granularity knowledge recall method for automotive information security according to claim 1, characterized in that, The technical documentation includes standards, specifications, regulations, guidelines, technical documents, vulnerability intelligence, and security case studies related to automotive information security.

3. The multi-granularity knowledge recall method for automotive information security according to claim 1, characterized in that, The specific method for constructing the topic tree is as follows: The process involves collecting keywords from all technical documents as candidate topics, merging semantically consistent but differently expressed candidate topics into a unified representation to form a standardized candidate topic set, and then deleting topics with a frequency below a threshold or a generalization level above a threshold. The remaining topics are then used to construct a leaf topic set. Subsequently, for each content block-level knowledge unit, the most relevant topic is selected from the leaf topic set as the primary topic to which the content block belongs, thus establishing a mapping relationship from content blocks to topics. Based on the semantic similarity between leaf topic names and their co-occurrence in chapters, leaf topics that are semantically close and have strong common occurrence characteristics in document structure are merged to form upper-level parent topics. This results in a two-layer topic structure of "parent topic - leaf topic", thus creating a topic tree.

4. The multi-granularity knowledge recall method for automotive information security according to claim 1, characterized in that, The core object retrieval process calculates a matching score by fusing keyword matching scores and semantic similarity scores, and then selects documents with higher matching scores to form a candidate document set; where the matching score... The calculation formula is: ， in, This represents the semantic similarity between the core object O and the document. This indicates the keyword matching degree between the core object O and the document. Indicates the weight.

5. The multi-granularity knowledge recall method for automotive information security according to claim 1, characterized in that, Content similarity The calculation formula is as follows: ， in, This indicates the semantic similarity between the user's input question and the content block. This indicates the keyword coverage of the content block for the user's input question. This indicates the degree of match between the constraints of the user's input question and the content block. , , Indicates the weight.

6. The multi-granularity knowledge recall method for automotive information security according to claim 1, characterized in that, The scoring method used in the evidence screening process The calculation formula is as follows: ， in, Present evidence The semantic relevance of the content block to the question representation Q. Present evidence The coverage of the content block to constraint C, Present evidence The degree of matching between the content block and the question type T. , , Indicates the weight.

Images