Knowledge graph establishment method, device and product

Abstract

The invention discloses a knowledge graph establishment method, device and product. Knowledge in an original document is split into a plurality of knowledge points corresponding to vehicle functions by using a large language model, so that each node contains a complete function, the information integrity is improved, and the knowledge fragmentation problem is avoided. According to the method, knowledge points are extracted from an initial knowledge graph, the knowledge points are inserted into the initial knowledge graph, when newly added knowledge points appear in the initial knowledge graph, a large language model is used for recognizing the incidence relation between the newly added knowledge points and existing knowledge points in the initial knowledge graph, the limitation that a traditional knowledge graph only supports triples is broken through, and system and function barriers are broken through. The association relationship has different types and strengths, and can flexibly adapt to complex scenes of various function applications in the vehicle. And constructing the association relationship as an edge between the newly added knowledge point and the existing knowledge point, and establishing a knowledge graph. The knowledge graph comprises a comprehensive and accurate knowledge network, the association relationship of various vehicle functions is synchronously considered, and the risk of information loss is low.

Description

Technical Field

[0001] This application relates to the field of knowledge graph technology, and in particular to a method, apparatus and product for establishing a knowledge graph. Background Technology

[0002] In practical applications, large language models based on artificial intelligence (AI) often rely on external business knowledge. For example, Retrieval-Augmented Generation (RAG) technology obtains business knowledge by querying vector databases and knowledge graphs, providing a reference for model reasoning.

[0003] In existing technologies, there are complex interaction scenarios between various functions within a vehicle. For example, in-vehicle infotainment (IVI) systems need to consider multiple connection methods such as Bluetooth, Wi-Fi, and USB. However, models querying vector databases and knowledge graphs contain fragmented knowledge, which can only be used for reasoning based on knowledge of local similarities or fixed relationships. This makes it impossible to capture cross-document knowledge and thus unable to adapt to the complex interactions between various functions within the vehicle. Summary of the Invention

[0004] To address the aforementioned issues, this application provides a method, apparatus, and product for establishing a knowledge graph, thereby improving the comprehensiveness and accuracy of the knowledge graph to meet in-vehicle usage requirements.

[0005] This application discloses a method for knowledge graph construction, the method comprising:

[0006] The knowledge in the original document is broken down into multiple knowledge points using a large language model; each knowledge point corresponds to a vehicle function.

[0007] The knowledge points are inserted into the initial knowledge graph, and when new knowledge points appear in the initial knowledge graph, the large language model is used to identify the association between the new knowledge points and the existing knowledge points in the initial knowledge graph; the association has different types and strengths.

[0008] The relationships are constructed as edges between the newly added knowledge points and the existing knowledge points to establish a knowledge graph.

[0009] Optionally, the step of using the large language model to identify the association between the newly added knowledge points and existing knowledge points in the initial knowledge graph includes:

[0010] Based on the application scenario of the knowledge graph, determine the preset association relationships;

[0011] Prompt words are written based on a preset method for determining association relationships;

[0012] The prompt words are input into the large language model, and the large language model is invoked to identify the type and strength of the association; the type and strength are generated based on the application scenario of the vehicle function, and the strength is generated based on the semantic recognition result of the knowledge point.

[0013] Optionally, identifying the type of the association includes:

[0014] If there is an execution order between the first vehicle function corresponding to the newly added knowledge point and the second vehicle function corresponding to the existing knowledge point, the type is determined as a dependency.

[0015] If the execution effects of the first vehicle function and the second vehicle function are similar, the type is determined to be similar;

[0016] If the first vehicle function and the second vehicle function cannot be executed simultaneously, the type is determined to be mutually exclusive.

[0017] Optionally, the strength of the association is identified, including:

[0018] Perform semantic recognition on the newly added knowledge points, the existing knowledge points, and the relationships, and obtain the semantic recognition results;

[0019] If the semantic recognition result is that the association is recorded in the newly added knowledge point and / or the existing knowledge point, the intensity is determined to be the first-level intensity.

[0020] If the semantic recognition result is that the relationship is not recorded in the newly added knowledge point and / or the existing knowledge point but is represented in the form of intent, the intensity is determined to be a secondary intensity; the secondary intensity is lower than the primary intensity.

[0021] Optionally, the step of inputting the prompt word into the large language model and calling the large language model to identify the type and strength of the association includes:

[0022] The prompt words corresponding to the capabilities of the large language model are input into the large language model, and the large language model is invoked to identify the type and strength of the association. The prompt words corresponding to the capabilities of the large language model are used to instruct the large language model to identify multiple types of the association at once, or to instruct the large language model to identify each type of the association separately.

[0023] Optionally, the step of using a large language model to break down the knowledge in the original document into multiple knowledge points includes:

[0024] If the original document contains multiple levels of directories, the original document is split into multiple sub-documents, with the lowest level chapter in the multiple levels of directories as the node;

[0025] The large language model is used to identify whether the sub-document contains knowledge of multiple different vehicle functions;

[0026] If so, then using one of the vehicle functions as a node, the sub-document is split to obtain multiple knowledge points;

[0027] If not, then a sub-document will be considered as one knowledge point.

[0028] Optionally, after establishing the knowledge graph, the method further includes:

[0029] Input the query request corresponding to the first knowledge point into the knowledge graph;

[0030] Obtain the first knowledge point and the knowledge points that have a relationship with the first knowledge point corresponding to the query requirement.

[0031] Based on the above-mentioned method for establishing a knowledge graph, this application also discloses an apparatus for establishing a knowledge graph, including: a splitting unit, an insertion unit, and an establishment unit;

[0032] The splitting unit is used to split the knowledge in the original document into multiple knowledge points using a large language model; each knowledge point corresponds to a vehicle function.

[0033] The insertion unit is used to insert the knowledge point into the initial knowledge graph, and when a new knowledge point appears in the initial knowledge graph, the large language model is used to identify the association between the new knowledge point and the existing knowledge points in the initial knowledge graph; the association has different types and strengths.

[0034] The establishment unit is used to construct the association relationship as edges between the newly added knowledge points and the existing knowledge points, thereby establishing a knowledge graph.

[0035] Optionally, the insertion unit includes:

[0036] The sub-unit is determined to identify preset relationships based on the application scenario of the knowledge graph.

[0037] Write sub-units to generate prompt words based on preset association judgment methods;

[0038] The identification subunit is used to input the prompt word into the large language model and call the large language model to identify the type and strength of the association relationship; the type and the strength are generated based on the application scenario of the vehicle function, and the strength is generated based on the semantic recognition result of the knowledge point.

[0039] Optionally, the identification subunit includes:

[0040] The dependency determination subunit is used to determine the type as a dependency when there is an execution order between the first vehicle function corresponding to the newly added knowledge point and the second vehicle function corresponding to the existing knowledge point.

[0041] A similarity determination subunit is used to determine the type as similar when the execution effects of the first vehicle function and the second vehicle function are similar;

[0042] The mutual exclusion determination subunit is used to determine the type as mutually exclusive when the first vehicle function and the second vehicle function cannot be executed simultaneously.

[0043] Optionally, the identification subunit includes:

[0044] A semantic recognition subunit is used to perform semantic recognition on the newly added knowledge points, the existing knowledge points, and the relationships, and to obtain the semantic recognition results.

[0045] A first-level determining subunit is used to determine the strength as a first-level strength when the semantic recognition result is that the association is recorded in the newly added knowledge point and / or the existing knowledge point;

[0046] The secondary determination subunit is used to determine the intensity as secondary intensity when the semantic recognition result is that the relationship is not recorded in the newly added knowledge point and / or the existing knowledge point but is represented in the form of intent; the secondary intensity is lower than the primary intensity.

[0047] Optionally, the identification subunit includes:

[0048] The classification and recognition subunit is used to input the prompt words corresponding to the capabilities of the large language model into the large language model, and call the large language model to identify the type and strength of the association relationship; the prompt words corresponding to the capabilities of the large language model are used to instruct the large language model to identify multiple types of the association relationship at once, or to instruct the large language model to identify each type of the association relationship individually.

[0049] Optionally, the splitting unit includes:

[0050] The chapter splitting subunit is used to split the original document into multiple sub-documents by taking the chapter at the lowest level in the multi-level directory as the node when there are multiple levels of directories in the original document.

[0051] The function identification subunit is used to identify whether the sub-document contains knowledge of multiple different vehicle functions using the large language model;

[0052] The functional decomposition subunit is used to split the sub-document with one of the vehicle functions as a node to obtain multiple knowledge points;

[0053] The sub-units are defined to treat a sub-document as a knowledge point.

[0054] Optionally, the device further includes:

[0055] A query unit is used to input the query request corresponding to the first knowledge point into the knowledge graph;

[0056] The relationship query unit is used to obtain the first knowledge point and the knowledge points that have a relationship with the first knowledge point corresponding to the query requirement.

[0057] Based on the above-described method for establishing a knowledge graph, this application also discloses a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the above-described method.

[0058] Based on the above-described method for establishing a knowledge graph, this application also discloses a computer program product, characterized in that the computer program product includes a computer program, which, when executed by a processor, is used to implement the above-described method.

[0059] This application discloses a method, apparatus, and product for building a knowledge graph. It utilizes a large language model to break down the knowledge in the original document into multiple knowledge points, assigning each knowledge point to a vehicle function. This ensures that each node contains a single, complete function, improving information integrity and avoiding the fragmentation problem caused by paragraph-based splitting in traditional methods. These knowledge points are then inserted into an initial knowledge graph. When a new knowledge point appears in the initial knowledge graph, the large language model identifies the relationship between the new knowledge point and existing knowledge points in the initial knowledge graph, overcoming the limitation of traditional knowledge graphs that only support triples and breaking down system and functional barriers. This relationship has different types and strengths, flexibly adapting to the complex scenarios of various in-vehicle functional applications. The relationships are constructed as edges between new and existing knowledge points to build the knowledge graph. This knowledge graph contains a comprehensive and accurate knowledge network, simultaneously considering the relationships between multiple vehicle functions, minimizing the risk of information loss, and avoiding the misleading effects of knowledge networks generated by fixed relationship patterns. Attached Figure Description

[0060] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0061] Figure 1 This is a flowchart illustrating a method for establishing a knowledge graph as disclosed in an embodiment of this application.

[0062] Figure 2 A flowchart illustrating another method for establishing a knowledge graph disclosed in an embodiment of this application;

[0063] Figure 3 This is a schematic diagram of the structure of a knowledge graph creation device disclosed in an embodiment of this application. Detailed Implementation

[0064] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0065] Example 1: This application discloses a method for establishing a knowledge graph.

[0066] For details, please refer to Figure 1 The knowledge graph construction method disclosed in this embodiment includes the following steps:

[0067] Step 101: Use a large language model to break down the knowledge in the original document into multiple knowledge points.

[0068] Current methods typically split the original document by paragraphs or word count, which has a significant drawback: it's difficult to guarantee the completeness of knowledge point information. Therefore, the method described in this embodiment employs a more scientific and reasonable approach to split knowledge points, aiming for greater accuracy and practicality. Each resulting knowledge point corresponds to only one vehicle function, facilitating subsequent knowledge graph construction.

[0069] In the method described in this embodiment, the original document (such as a technical manual, product manual, etc.) may contain multiple chapters organized with multi-level directories. To improve the efficiency and accuracy of knowledge decomposition, it can be decomposed first based on chapters. As an feasible solution, the original document is decomposed into multiple sub-documents using the lowest-level chapter in the multi-level directory as the node. This decomposition method helps to initially simplify complex original documents, laying the foundation for subsequent more detailed knowledge point decomposition.

[0070] In the method described in this embodiment, when the original document has complete chapter numbers and clear chapter divisions, each chapter focuses on a specific function. For example, in a functional design document describing the in-car Bluetooth phone function, different chapters will introduce multiple functions such as logging into the address book, connecting to Bluetooth devices, and making calls. In practice, if a chapter contains many sub-chapter contents, and the parent chapter does not contain specific knowledge information, then only the sub-chapter content can be retained. The specific method for splitting chapters can be tailored to requirements, ensuring that each sub-document contains the complete knowledge of one and only one minimal chapter.

[0071] In the method described in this embodiment, after completing the chapter splitting, the knowledge points of each sub-document can be further analyzed. If a sub-document is identified as containing multiple independent knowledge points, then the sub-document is split again using one knowledge point as a node, thereby obtaining multiple knowledge points. For example, if a sub-document introduces both the vehicle's automatic parking function and adaptive cruise control function, and these two functions are independent and have complete knowledge information, then this sub-document can be split into two knowledge points: "automatic parking function" and "adaptive cruise control function".

[0072] In the method described in this embodiment, if a sub-document is identified as containing knowledge about only one vehicle function, then there is no need for further splitting; the sub-document can be directly treated as a knowledge point. For example, if a sub-document specifically details the vehicle's tire pressure monitoring function, and all content revolves around this function, then this sub-document can be directly treated as a knowledge point about the "tire pressure monitoring function."

[0073] The identification of knowledge points can be achieved using the semantic understanding and analysis capabilities of a large language model. For example, in the CarPlay function introduction section of an in-vehicle IVI system, multiple applications such as calls, navigation, and music playback are included. These applications are relatively independent and each has unique knowledge information. The large language model can perform semantic analysis on the sub-documents to identify similar but different knowledge points. Then, based on each independent knowledge point, the sub-documents describing CarPlay functions are further split into multiple knowledge points such as CarPlay-call, CarPlay-navigation, and CarPlay-music playback. This splitting avoids situations where a sub-document contains multiple unrelated knowledge points or a knowledge point is irrationally split, making the content of the knowledge points clearer and facilitating subsequent exploration of the relationships between knowledge points.

[0074] Step 102: Insert the knowledge point into the initial knowledge graph, and when a new knowledge point appears in the initial knowledge graph, use the large language model to identify the relationship between the new knowledge point and the existing knowledge point in the initial knowledge graph.

[0075] In this embodiment, knowledge points are inserted into an initial knowledge graph (or graph database), and the knowledge points are gradually connected according to their relationships to enrich the initial knowledge graph. Specifically, the first knowledge point out of n knowledge points is inserted into the initial knowledge graph, then the second knowledge point out of n knowledge points is inserted into the initial knowledge graph, while identifying the relationship between the second knowledge point and the first knowledge point. This process continues until the nth knowledge point is inserted into the initial knowledge graph, while identifying the relationship between the nth knowledge point and the (n-1), (n-2), ..., the first knowledge point. It should be noted that knowledge points can be acquired all at once and inserted into the initial knowledge graph one by one. Alternatively, each acquired knowledge point can be inserted into the initial knowledge graph, facilitating subsequent updates and iterations of knowledge points.

[0076] In the method of this embodiment, the relationships between knowledge points can be identified using a large language model. As one feasible approach, pre-defined relationships can be determined based on the application scenario of the knowledge graph. Then, prompt words are written based on the pre-defined relationship judgment method and input into the large language model. These prompt words need to be written in an expression that the large language model can understand, avoiding self-contradictory content and vague expressions as much as possible. Alternatively, the pre-defined relationship judgment method can be embedded into the large language model through model pre-training or post-training fine-tuning.

[0077] The types and strengths of the relationships are primarily generated based on the application scenarios of vehicle functions. Due to the different application scenarios of vehicle functions, the types and strengths of the relationships between them are complex and diverse. For example, in the "play music" application scenario, the relationship between local music and cloud music is similarity; in the "find music" application scenario, the relationship between local music and cloud music is mutual exclusion. Another example is that in the "remote control" application scenario, Bluetooth connection must be executed before the trunk opening function can be executed. Therefore, the relationship between these two vehicle functions is dependency, with a strength of level one. In the "in-car control" application scenario, Bluetooth connection can be executed or not, and the trunk opening function can be executed in either case. Therefore, the relationship between these two vehicle functions is dependency, with a weaker strength of level two. Therefore, for the application scenarios of this knowledge graph, various relationship judgment methods can be predefined based on actual needs.

[0078] As a feasible solution, when the execution order of the vehicle functions corresponding to two knowledge points exists, the relationship between the two knowledge points is defined as dependency. When the execution effects of the vehicle functions corresponding to two knowledge points are similar, the relationship between the two knowledge points is defined as similarity. Specifically, similarity can be determined by the similarity of the descriptions of the two knowledge points, their descriptions of the same vehicle function, and their recorded operation procedures. When the vehicle functions corresponding to two knowledge points cannot be executed simultaneously, the relationship between the two knowledge points is defined as mutual exclusion. Furthermore, when the content recorded in the first knowledge point completely includes the content of the second knowledge point, the two knowledge points can be judged as having an inclusion / inclusion relationship. When the content recorded in the first knowledge point is the cause of the content recorded in the second knowledge point, the two knowledge points can be judged as having a causal relationship.

[0079] For example, knowledge points related to the in-car Bluetooth phone function, such as logging into the address book, connecting Bluetooth devices, and making calls, are inserted into the initial knowledge graph. Similarly, multiple knowledge points related to the CarPlay function, such as CarPlay-phone, CarPlay-navigation, and CarPlay-music playback, are also imported into the same initial knowledge graph. Where there is a dependency between the knowledge points of connecting Bluetooth devices and making calls, this relationship is identified as dependency in the initial knowledge graph. Likewise, if the functions described by the call knowledge point and the CarPlay-phone knowledge point are similar, this relationship is identified as similarity in the knowledge graph.

[0080] In the method of this embodiment, depending on the capabilities of the large language model, it can either call the large language model at once to determine multiple types of association relationships, or it can determine each type of association relationship independently. Specifically, it can involve writing prompt words corresponding to the capabilities of the large language model, which instruct the large language model to identify multiple types of association relationships at once, or instruct the large language model to identify each type of association relationship separately.

[0081] In the method of this embodiment, a large language model can also be invoked to identify the strength of associations, facilitating the filtering of associations of specific strengths during searches. This strength can be generated based on the application scenario of vehicle functions and the semantic recognition results of knowledge points. Specifically, semantic recognition can be performed on new knowledge points, existing knowledge points, and associations to obtain semantic recognition results. If the semantic recognition result indicates that the association is recorded in new knowledge points and / or existing knowledge points, the strength is determined to be a higher level (Level 1). Correspondingly, if the semantic recognition result indicates that the association is not recorded in new knowledge points and / or existing knowledge points but is represented by intent, the strength is determined to be a lower level (Level 2).

[0082] For example, taking the dependency relationship between connecting a Bluetooth device and making a call in Bluetooth phone functionality as an example, Level 1 strength indicates that there may be a dependency between the two knowledge points, but this dependency is not explicitly recorded in the knowledge content. Level 2 strength, on the other hand, indicates that the knowledge content explicitly records that a Bluetooth device needs to be connected before a call can be made, so the dependency between the two is obvious.

[0083] In the method of this embodiment, the type and strength of the association in the above steps are only examples, and the specific content or value of the type and strength are not limited here, as long as the method of this embodiment can be implemented.

[0084] Step 103: Construct the relationships as edges between the newly added knowledge points and the existing knowledge points to establish a knowledge graph.

[0085] In the method of this embodiment, after the knowledge graph is constructed, the user can input a query request corresponding to the first knowledge point into the knowledge graph. The knowledge graph will query and return the first knowledge point, as well as all knowledge points that are related to the first knowledge point, to the user. Of course, the user can also input requests such as "query knowledge points that are dependent on the first knowledge point" or "query knowledge points that have a first-level strong relationship with the first knowledge point," and the knowledge graph can flexibly query and return data according to the user's needs.

[0086] The method described in this embodiment supports the splitting of knowledge points by chapters and independent vehicle functions, ensuring the integrity of knowledge at individual nodes and reducing knowledge fragmentation. Simultaneously, the split knowledge points are uniformly stored in the same knowledge graph, eliminating information silos and supporting cross-functional knowledge retrieval. It also supports large language models to recognize various types of relationships and is more suitable for concrete real-world business application scenarios, addressing the diverse needs of practical applications. Furthermore, it employs a relationship strength grading mechanism, supporting on-demand, customized, and precise queries. This allows for the rapid and efficient retrieval of other knowledge points with different types and strengths of relationships with specified knowledge nodes, avoiding multiple rounds of queries, improving business response efficiency, and eliminating the uncertainty of relying on large model inference. In practical business applications, it can efficiently retrieve the required knowledge content. Compared to existing RAG methods, the method in this embodiment improves the recall rate for various types of knowledge and supports queries of combined knowledge, enabling cross-knowledge point collaborative reasoning with a recall rate of 80%.

[0087] Example 2: This application discloses another method for knowledge graph construction. Please refer to [link / reference]. Figure 2 This embodiment describes the process of knowledge graph construction.

[0088] Step 201: Split the knowledge in the original document into multiple sub-documents by chapter.

[0089] Step 202: For each sub-document, use a large language model to identify the knowledge within it, and take the knowledge corresponding to a vehicle function as a knowledge point.

[0090] Step 203: Insert a knowledge point into the initial knowledge graph.

[0091] Step 204: Write prompt words according to the preset association type and strength.

[0092] In the method of this embodiment, the writing of prompt words can also be performed simultaneously with steps 201 to 203, and the execution time of step 204 is not specifically limited here.

[0093] Step 205: Input the prompt word into the large language model to identify the relationship between the knowledge point and the existing knowledge points in the initial knowledge graph.

[0094] In the method of this embodiment, step 205 can be skipped when there are no existing knowledge points in the initial knowledge graph.

[0095] Step 206: Construct the relationship between this knowledge point and existing knowledge points as edges.

[0096] Step 207: Determine if the knowledge points can be obtained from the original document. If yes, return to step 201. If no, proceed to step 208.

[0097] Step 208: Obtain the knowledge graph.

[0098] Based on the knowledge graph construction method disclosed in the above embodiments, this embodiment correspondingly discloses a knowledge graph construction apparatus. Please refer to... Figure 3 The device for establishing a knowledge graph includes: a splitting unit 301, an insertion unit 302, and an establishment unit 303;

[0099] The splitting unit 301 is used to split the knowledge in the original document into multiple knowledge points using a large language model; one knowledge point corresponds to one vehicle function.

[0100] The insertion unit 302 is used to insert the knowledge point into the initial knowledge graph, and when a new knowledge point appears in the initial knowledge graph, the large language model is used to identify the association between the new knowledge point and the existing knowledge points in the initial knowledge graph; the association has different types and strengths.

[0101] The establishment unit 303 is used to construct the association relationship as an edge between the newly added knowledge point and the existing knowledge point to establish a knowledge graph.

[0102] Optionally, the insertion unit 302 includes:

[0103] The sub-unit is determined to identify preset relationships based on the application scenario of the knowledge graph.

[0104] Write sub-units to generate prompt words based on preset association judgment methods;

[0105] The identification subunit is used to input the prompt word into the large language model and call the large language model to identify the type and strength of the association relationship; the type and the strength are generated based on the application scenario of the vehicle function, and the strength is generated based on the semantic recognition result of the knowledge point.

[0106] Optionally, the identification subunit includes:

[0107] The dependency determination subunit is used to determine the type as a dependency when there is an execution order between the first vehicle function corresponding to the newly added knowledge point and the second vehicle function corresponding to the existing knowledge point.

[0108] A similarity determination subunit is used to determine the type as similar when the execution effects of the first vehicle function and the second vehicle function are similar;

[0109] The mutual exclusion determination subunit is used to determine the type as mutually exclusive when the first vehicle function and the second vehicle function cannot be executed simultaneously.

[0110] Optionally, the identification subunit includes:

[0111] A semantic recognition subunit is used to perform semantic recognition on the newly added knowledge points, the existing knowledge points, and the relationships, and to obtain the semantic recognition results.

[0112] A first-level determining subunit is used to determine the strength as a first-level strength when the semantic recognition result is that the association is recorded in the newly added knowledge point and / or the existing knowledge point;

[0113] The secondary determination subunit is used to determine the intensity as secondary intensity when the semantic recognition result is that the relationship is not recorded in the newly added knowledge point and / or the existing knowledge point but is represented in the form of intent; the secondary intensity is lower than the primary intensity.

[0114] Optionally, the identification subunit includes:

[0115] The classification and recognition subunit is used to input the prompt words corresponding to the capabilities of the large language model into the large language model, and call the large language model to identify the type and strength of the association relationship; the prompt words corresponding to the capabilities of the large language model are used to instruct the large language model to identify multiple types of the association relationship at once, or to instruct the large language model to identify each type of the association relationship individually.

[0116] Optionally, the splitting unit 301 includes:

[0117] The chapter splitting subunit is used to split the original document into multiple sub-documents by taking the chapter at the lowest level in the multi-level directory as the node when there are multiple levels of directories in the original document.

[0118] The function identification subunit is used to identify whether the sub-document contains knowledge of multiple different vehicle functions using the large language model;

[0119] The functional decomposition subunit is used to split the sub-document with one of the vehicle functions as a node to obtain multiple knowledge points;

[0120] The sub-units are defined to treat a sub-document as a knowledge point.

[0121] Optionally, the device further includes:

[0122] A query unit is used to input the query request corresponding to the first knowledge point into the knowledge graph;

[0123] The relationship query unit is used to obtain the first knowledge point and the knowledge points that have a relationship with the first knowledge point corresponding to the query requirement.

[0124] Based on the above-described method for establishing a knowledge graph, this application also discloses a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the above-described method.

[0125] Based on the above-described method for establishing a knowledge graph, this application also discloses a computer program product, characterized in that the computer program product includes a computer program, which, when executed by a processor, is used to implement the above-described method.

[0126] The embodiments in this specification are described in a progressive manner. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant details can be found in the method section.

[0127] It should also be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relational relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0128] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0129] The features described in the embodiments of this specification can be substituted for or combined with each other, so that those skilled in the art can implement or use this application.

[0130] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for constructing a knowledge graph, characterized in that, include: The knowledge in the original document is broken down into multiple knowledge points using a large language model; One of the aforementioned knowledge points corresponds to one vehicle function; The knowledge points are inserted into the initial knowledge graph, and when new knowledge points appear in the initial knowledge graph, the large language model is used to identify the association between the new knowledge points and the existing knowledge points in the initial knowledge graph. The associations have different types and strengths; The relationships are constructed as edges between the newly added knowledge points and the existing knowledge points to establish a knowledge graph.

2. The method according to claim 1, characterized in that, The step of using the large language model to identify the association between the newly added knowledge points and the existing knowledge points in the initial knowledge graph includes: Based on the application scenario of the knowledge graph, determine the preset association relationships; Prompt words are written based on a preset method for determining association relationships; The prompt word is input into the large language model, and the large language model is invoked to identify the type and strength of the association; the type and strength are generated based on the application scenario of the vehicle function, and the strength is generated based on the semantic recognition result of the knowledge point.

3. The method according to claim 2, characterized in that, Identifying the type of the association includes: If there is an execution order between the first vehicle function corresponding to the newly added knowledge point and the second vehicle function corresponding to the existing knowledge point, the type is determined as a dependency. If the execution effects of the first vehicle function and the second vehicle function are similar, the type is determined to be similar; If the first vehicle function and the second vehicle function cannot be executed simultaneously, the type is determined to be mutually exclusive.

4. The method according to claim 2, characterized in that, Identifying the strength of the association includes: Perform semantic recognition on the newly added knowledge points, the existing knowledge points, and the relationships, and obtain the semantic recognition results; If the semantic recognition result is that the association is recorded in the newly added knowledge point and / or the existing knowledge point, the intensity is determined to be the first-level intensity. If the semantic recognition result is that the new knowledge point and / or the existing knowledge point does not record the association relationship but represents the association relationship in the form of intent, the intensity is determined to be a secondary intensity; the secondary intensity is lower than the primary intensity.

5. The method according to claim 2, characterized in that, The step of inputting the prompt word into the large language model and calling the large language model to identify the type and strength of the association includes: The prompt words corresponding to the capabilities of the large language model are input into the large language model, and the large language model is invoked to identify the type and strength of the association relationship; the prompt words corresponding to the capabilities of the large language model are used to instruct the large language model to identify multiple types of the association relationship at once, or to instruct the large language model to identify each type of the association relationship individually.

6. The method according to claim 1, characterized in that, The method of using a large language model to break down the knowledge in the original document into multiple knowledge points includes: If the original document contains multiple levels of directories, the original document is split into multiple sub-documents, with the lowest level chapter in the multiple levels of directories as the node; The large language model is used to identify whether the sub-document contains knowledge of multiple different vehicle functions; If so, then using one of the vehicle functions as a node, the sub-document is split to obtain multiple knowledge points; If not, then a sub-document will be considered as one knowledge point.

7. The method according to any one of claims 1-6, characterized in that, After establishing the knowledge graph, the method further includes: Input the query request corresponding to the first knowledge point into the knowledge graph; Obtain the first knowledge point and the knowledge points that have a relationship with the first knowledge point corresponding to the query requirement.

8. An apparatus for establishing a knowledge graph, characterized in that, include: Split cells, insert cells, and create cells; The splitting unit is used to split the knowledge in the original document into multiple knowledge points using a large language model; One of the aforementioned knowledge points corresponds to one vehicle function; The insertion unit is used to insert the knowledge point into the initial knowledge graph, and when a new knowledge point appears in the initial knowledge graph, the large language model is used to identify the association between the new knowledge point and the existing knowledge point in the initial knowledge graph. The associations have different types and strengths; The establishment unit is used to construct the association relationship as edges between the newly added knowledge points and the existing knowledge points, thereby establishing a knowledge graph.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the method as described in any one of claims 1-6.

10. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, is used to implement the method described in any one of claims 1-6.

Images