Database management device, database management system and database management method

Abstract

To promote identification of information having smaller semantic content and exclusion of such information from a knowledge base, along with visualization of knowledge stored in the knowledge base.SOLUTION: A database management device comprises: a tag management unit that, for each text chunk of a set of first text chunks in a first document, allocates a descriptive tag that describes structure characteristics of the text chunk in the first document; an expression generation unit that generates a first graph expression which indicates a set of n-dimensional data points corresponding to the set of the first text chunks and having a spatial relation based on semantic similarity between text chunks; a classification unit that generates a second graph expression to which a semantic label or a structural label has been allocated for each data point by comparing the descriptive tag with a classification rule; and an exclusion unit that excludes a first text chunk corresponding to the first data point of the set of data points to which the structural label has been allocated from the first document.SELECTED DRAWING: Figure 3

Description

[Technical Field]

[0001] The present disclosure relates to a database management device, a database management system, and a database management method. [Background technology]

[0002] In recent years, with the rapid development of the digital age, there has been an increasing demand to store vast amounts of information and maintain it in a usable form. One of the tools that facilitates the organization, retrieval, and retrieval of information is the so-called knowledge base.

[0003] US Pat. No. 10,956,790 (Patent Document 1) discloses a method for visualizing data contained in a knowledge base and presenting it to a user. Patent Document 1 describes a technology related to a graphical user interface including: a data interface for receiving a series of text documents; a calculation unit for converting each received text document into a vector of an n-dimensional manifold (n is greater than 2), converting each n-dimensional manifold vector into two-dimensional data, generating a first mapping indicating which two-dimensional data corresponds to which n-dimensional manifold vector, and generating a second mapping indicating which two-dimensional data corresponds to which received text document; a first user display configured to display each two-dimensional data on a 2D plane; a first user interface configured to allow a user to select a subset of the two-dimensional data to be displayed and, when the subset of two-dimensional data is selected, automatically select an n-dimensional manifold vector corresponding to the subset of two-dimensional data; and a second user display configured to display a list of text documents corresponding to the subset of two-dimensional data. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] U.S. Patent No. 10,956,790 Summary of the Invention [Problem to be solved by the invention]

[0005] By the way, to perform automatic reasoning on the information contained in the knowledge base, the knowledge base is sometimes used in conjunction with RAG (Retrieval Augmented Generation) means. RAG is a means of searching the information stored in the knowledge base and having a large-scale language model (LLM) generate an answer. By supplementing the knowledge held by the LLM with the knowledge base, it is possible to obtain more reliable and accurate answers.

[0006] To improve the accuracy of RAG, it is necessary to add the latest knowledge to the knowledge base and to maintain the knowledge contained in the knowledge base. More specifically, the more information a knowledge base contains, the less accurate the search for identifying relevant knowledge becomes. Therefore, it is desirable to remove information from the knowledge base that is not necessary for generating answers using LLM.

[0007] However, conventional methods such as those described in Patent Document 1 are not designed to identify and remove unnecessary information from the knowledge base. This can lead to a huge amount of information in the knowledge base, making it difficult to search for meaningful knowledge and limiting the accuracy of RAG.

[0008] Therefore, an object of the present disclosure is to provide a database management means that can visualize the knowledge stored in a knowledge base and facilitate the identification and exclusion of information with less semantic content from the knowledge base. [Means for solving the problem]

[0009] In order to solve the above problem, a representative database management device of the present invention includes a processor, a memory, and a storage unit for storing a knowledge database including one or more documents, wherein the memory processes a first set of text chunks in a first document included in the documents using a predetermined classification model, thereby assigning to each text chunk included in the first set of text chunks a description tag that describes a structural feature of the text chunk in the first document; and a tag management unit processes the first set of text chunks using a predetermined natural language processing means to generate n-dimensional data points corresponding to the first set of text chunks and having a spatial relationship based on semantic similarity between the text chunks. a classification unit that generates and outputs a second graph representation in which a semantic label or a structural label is assigned to each data point included in the set of data points in the first graph representation by comparing the descriptive tags assigned to the first set of text chunks with classification rules that specify whether a text chunk in a document corresponds to a semantic element or a structural element of the document; and processing instructions that cause the processor to operate as an exclusion unit that excludes from the first document, first text chunks of the set of data points that correspond to first data points to which the structural label is assigned, based on a user instruction input to the second graph representation. [Effects of the Invention]

[0010] According to the present disclosure, it is possible to provide a database management means that can visualize the knowledge stored in a knowledge base and facilitate the identification and removal from the knowledge base of information with less semantic content. Problems, configurations, and effects other than those described above will become apparent from the following description of the preferred embodiment of the invention. [Brief explanation of the drawings]

[0011] [Figure 1]FIG. 1 is a diagram illustrating a computer system for implementing an embodiment of the present disclosure. [Figure 2] FIG. 2 is a diagram illustrating an example of a configuration of a database management system according to an embodiment of the present disclosure. [Figure 3] FIG. 3 is a diagram illustrating an example of a configuration of a database management device according to an embodiment of the present disclosure. [Figure 4] FIG. 4 is a diagram showing an example of the overall flow of the database management means according to an embodiment of the present disclosure. [Figure 5] FIG. 5 is a diagram illustrating an example of the configuration of a knowledge base according to an embodiment of the present disclosure. [Figure 6] FIG. 6 is a diagram illustrating an example of a configuration of a classification rule according to an embodiment of the present disclosure. [Figure 7] FIG. 7 is a diagram illustrating an example of a prompt input to a classification model according to an embodiment of the present disclosure. [Figure 8] FIG. 8 is a diagram illustrating an example of the flow of a graph representation generation process according to an embodiment of the present disclosure. [Figure 9] FIG. 9 is a diagram illustrating an example of the flow of a text chunk classification process according to an embodiment of the present disclosure. [Figure 10] FIG. 10 is a diagram illustrating an example of the flow of the exclusion process according to an embodiment of the present disclosure. [Figure 11] FIG. 11 is a diagram illustrating an example of a graph representation display screen according to an embodiment of the present disclosure. [Figure 12] FIG. 12 is a diagram illustrating an example of a details confirmation screen according to an embodiment of the present disclosure. DETAILED DESCRIPTION OF THE INVENTION

[0012] Hereinafter, an embodiment of the present invention will be described with reference to the drawings. Note that the present invention is not limited to this embodiment. In addition, in the description of the drawings, the same parts are designated by the same reference numerals. Furthermore, although terms such as "first," "second," and "third" may be used to describe various elements or components in this disclosure, it will be understood that these elements or components should not be limited by these terms. These terms are used only to distinguish one element or component from another. Thus, a first element or component discussed below could also be referred to as a second element or component without departing from the teachings of the inventive concept.

[0013] Next, referring to FIG. 1, a computer system 100 for implementing embodiments of the present disclosure will be described. The mechanisms and devices of various embodiments disclosed herein may be applied to any suitable computing system. The main components of the computer system 100 include one or more processors 102, memory 104, a terminal interface 112, a storage interface 113, an I / O (input / output) device interface 114, and a network interface 115. These components may be interconnected via a memory bus 106, an I / O bus 108, a bus interface unit 109, and an I / O bus interface unit 110.

[0014] Computer system 100 may include one or more general-purpose programmable central processing units (CPUs) 102A and 102B, collectively referred to as processor 102. In some embodiments, computer system 100 may include multiple processors, while in other embodiments, computer system 100 may be a single CPU system. Each processor 102 executes instructions stored in memory 104 and may include an on-board cache. In some embodiments, computer system 100 may include a graphics processing unit (GPU) in addition to processor 102. In some embodiments, computer system 100 may include a tensor processing unit (TPU) or specialized hardware specialized for AI in addition to processor 102. In this manner, using a GPU, TPU, or specialized hardware specialized for AI can speed up the processing of machine learning models, etc., used in database management application 150, described below.

[0015] In one embodiment, memory 104 may include random-access semiconductor memory, storage devices, or storage media (either volatile or non-volatile) for storing data and programs. Memory 104 may store all or part of the programs, modules, and data structures that implement the functions described herein. For example, memory 104 may store a database management application. In one embodiment, the database management application may include instructions or descriptions that execute the functions described below on processor 102.

[0016] In some embodiments, the database management application may be implemented in hardware via semiconductor devices, chips, logic gates, circuits, circuit cards, and / or other physical hardware devices instead of or in addition to a processor-based system. In some embodiments, the database management application may include data other than instructions or descriptions. In some embodiments, cameras, sensors, or other data input devices (not shown) may be provided to communicate directly with bus interface unit 109, processor 102, or other hardware in computer system 100.

[0017] Computer system 100 may include a bus interface unit 109 that facilitates communication between processor 102, memory 104, display system 124, and I / O bus interface unit 110. I / O bus interface unit 110 may couple to an I / O bus 108 for transferring data to and from various I / O units. I / O bus interface unit 110 may communicate via I / O bus 108 with multiple I / O interface units 112, 113, 114, and 115, also known as I / O processors (IOPs) or I / O adapters (IOAs).

[0018] Display system 124 may include a display controller, a display memory, or both. The display controller may provide video, audio, or both data to display device 126. Computer system 100 may also include one or more sensors or other devices configured to collect data and provide the data to processor 102.

[0019] For example, computer system 100 may include biometric sensors that collect heart rate data, stress level data, etc., environmental sensors that collect humidity data, temperature data, pressure data, etc., and motion sensors that collect acceleration data, movement data, etc. Other types of sensors may also be used. Display system 124 may be connected to a display device 126, such as a standalone display screen, a television, a tablet, or a handheld device.

[0020] The I / O interface unit provides functionality for communicating with various storage or I / O devices. For example, the terminal interface unit 112 may be attached to user I / O devices 116, such as user output devices such as a video display, a television with speakers, and user input devices such as a keyboard, a mouse, a keypad, a touchpad, a trackball, buttons, a light pen, or other pointing device. A user may use a user interface to enter input data or instructions into the user I / O devices 116 and the computer system 100, and receive output data from the computer system 100, by operating the user input devices. The user interface may be displayed on a display, played through speakers, or printed via a printer via the user I / O devices 116, for example.

[0021] Storage interface 113 allows attachment of one or more disk drives or direct access storage device 117 (typically a magnetic disk drive storage device, but may also be an array of disk drives or other storage devices configured to appear as a single disk drive). In some embodiments, storage device 117 may be implemented as any secondary storage device. The contents of memory 104 may be stored in storage device 117 and retrieved as needed from storage device 117. I / O device interface 114 may provide an interface to other I / O devices, such as printers, fax machines, etc. Network interface 115 may provide a communications path that allows computer system 100 and other devices to communicate with each other. This communications path may be, for example, network 130.

[0022] In some embodiments, computer system 100 may be a device that receives requests from other computer systems (clients) without a direct user interface, such as a multi-user mainframe computer system, a single-user system, or a server computer. In other embodiments, computer system 100 may be a desktop computer, a portable computer, a laptop, a tablet computer, a pocket computer, a telephone, a smartphone, or any other suitable electronic device.

[0023] Next, a database management system according to an embodiment of the present disclosure will be described with reference to FIG.

[0024] Fig. 2 is a diagram illustrating an example of a configuration of a database management system 200 according to an embodiment of the present disclosure. The database management system 200 according to an embodiment of the present disclosure is a system for visualizing knowledge stored in a knowledge base and promoting the identification and exclusion of information with less semantic content from the knowledge base, and as shown in Fig. 2, mainly includes a database management device 210, a classification model 230, a control terminal 240, and a user terminal 250. The database management device 210, the classification model 230, the control terminal 240, and the user terminal 250 may be connected to each other via a communication network 260.

[0025] The database management device 210 is a device that visualizes knowledge stored in a knowledge base and facilitates the identification and exclusion of information with less semantic content from the knowledge base, and as shown in Figure 2, mainly includes a storage unit 211, a calculation unit 212, a network interface 213, and a memory 214. The details of the configuration of the database management device 210 will be described later with reference to FIG. 3, and therefore will not be described here.

[0026] The classification model 230 is a machine learning model for processing text chunks included in documents stored in the storage unit 211 or the like of the database management device 210, and assigning description tags that describe the structural features of the documents to the text chunks. For example, the classification model 230 may be a large-scale language model (LLM), a neural network trained to assign description tags to text chunks, or the like, and is not particularly limited here. As an example, when the classification model 230 receives a query from the database management device 210 requesting the assignment of description tags to a specific text chunk, the classification model 230 may provide the database management device 210 with a response in which description tags are assigned to the text chunk. The classification model 230 shown in FIG. 2 may be implemented on a cloud server accessible from the database management device 210, or may be implemented within the database management device 210, and is not particularly limited here.

[0027] The control terminal 240 is a terminal used to manage the knowledge database stored in the storage unit 211 of the database management device 210. The control terminal 240 may be used to send commands to the database management device 210 to store new documents in the storage unit 211 of the database management device 210 or to make changes to existing documents stored in the storage unit 211 of the database management device 210. In one embodiment, the control terminal 240 may also be used to check an n-dimensional graphical representation (where "n" may be 2 or 3) that visually shows the contents of documents stored in the storage unit 211 of the database management device 210. As shown in FIG. 2, the control terminal 240 may include a network interface 241 for communicating with external devices via a communication network 260, an arithmetic unit 244 such as a CPU for executing processing, an input unit 242 for accepting information from a user, and a display unit 243 for outputting various information to the user.

[0028] Note that Figure 2 shows an example in which the database management device 210 and the control terminal 240 are separate, independent devices, but the present disclosure is not limited to this; it is also possible for the database management device 210 and the control terminal 240 to be implemented within the same device, and this disclosure is not particularly limited thereto.

[0029] User terminal 250 is a terminal used by a user of database management according to an embodiment of the present disclosure. In an embodiment, a user can input a query to a Retrieval Augmented Generation (RAG) service provided by database management device 210 via user terminal 250 and receive a response to the query. As an example, user terminal 250 may include, but is not limited to, a smartphone, smartwatch, tablet, or personal computer of a user who subscribes to the RAG service provided by database management device 210. For ease of explanation, FIG. 2 illustrates an example of a configuration including one user terminal 250, but the number of user terminals 250 is not limited, and a configuration including multiple user terminals 250 is also possible.

[0030] The communications network 260 may include, for example, a local area network (LAN), a wide area network (WAN), a satellite network, a cable network, a WiFi network, or any combination thereof.

[0031] 2, the database management system 200 can visualize the knowledge stored in the knowledge base and facilitate the identification and elimination of information with less semantic content from the knowledge base. Furthermore, by excluding information with less semantic content from documents stored in the knowledge base, the accuracy of retrieval of appropriate knowledge to answer a query can be improved, thereby improving the accuracy of the RAG means.

[0032] Next, the configuration of the database management device according to the embodiment of the present disclosure will be described with reference to FIG.

[0033] Fig. 3 is a diagram illustrating an example of the configuration of a database management device 210 according to an embodiment of the present disclosure. As illustrated in Fig. 3, the database management device 210 mainly includes a storage unit 211, a calculation unit 212, a network interface 213, and a memory 214. The storage unit 211, the calculation unit 212, the network interface 213, and the memory 214 may be connected via a bus 305. In one embodiment, database management device 210 may be implemented by computer system 100 shown in FIG.

[0034] The memory unit 211 is a memory area that contains a database (hereinafter referred to as "DB") 310 for storing various information related to an embodiment of the present disclosure, and may be, for example, an HDD (Hard Disk Drive), an SSD (Solid State Drive), cloud storage, etc., and is not particularly limited here.

[0035] As shown in FIG. 2, database 310 may include knowledge base 311 , classification rules 312 , and prompts 313 . The knowledge base 311 is a database for managing the organization, retrieval, and organization of information, and may store, for example, knowledge used to generate answers to queries input to the RAG service. The knowledge base 311 may store one or more documents composed of text in a natural language. The text contained in each document may be stored in the knowledge base 311 in a form divided into text chunks of a predetermined size (a predetermined number of characters, words, lines, etc.). In one embodiment, the knowledge base 311 may store the text chunks composed of natural language, embedding vectors obtained by converting the text chunks into vector format, and metadata related to the text chunks. As described in this disclosure, the more information contained in the knowledge base 311, the lower the accuracy of searches to identify appropriate knowledge. Therefore, it is desirable to remove from the knowledge base information that is not necessary for generating answers using the LLM or RAG. The configuration of the knowledge base 311 will be explained with reference to FIG. 5, and therefore will not be explained here.

[0036] The classification rules 312 are rules that specify, for each text chunk of a document stored in the knowledge base 311, whether the text chunk corresponds to a semantic element or a structural element of the document. In one embodiment, the structural characteristics of the text chunk may be described, and the descriptive tag assigned to the text chunk may be compared with the classification rules 312 to determine whether the text chunk corresponds to a semantic element or a structural element of the document. The classification rules 312 will be described in detail with reference to FIG. 6, and therefore will not be described here.

[0037] Prompt 313 is a prompt for requesting a predetermined classification model (eg, classification model 230 shown in FIG. 2) to assign descriptive tags to each text chunk of a document stored in knowledge base 311 . The details of the prompt 313 will be explained with reference to FIG. 7, and therefore will not be explained here.

[0038] The calculation unit 212 is a processing unit for executing processing instructions that perform the functions of various functional units that make up the database management application 150 and RAG application 226 described below, and may substantially correspond to, for example, the processor 102 of the computer system 100 shown in Figure 1. The network interface 213 is a network interface for communicating between the database management device 210 and an external device (such as the user terminal 250 or control terminal 240 shown in FIG. 2) via a predetermined communication network such as the Internet, and may substantially correspond to the network interface 115 of the computer system 100 shown in FIG. 1, for example.

[0039] The memory 214 may be a memory for storing a database management application 150 and a RAG application 226 for implementing the functions of the database management means according to an embodiment of the present disclosure. The database management application 150 may include an embedding manager 221, a tag manager 222, a representation generator 223, a classifier 224, and an excluder 225, as shown in FIG.

[0040] The embedding management unit 221 is a functional unit that converts text chunks contained in documents written in natural language into embedding vectors in vector format and stores them in the knowledge base 311 . The embedding management unit 221 will be described in detail later, and therefore will not be described here.

[0041] The tag management unit 222 is a functional unit that processes a set of text chunks contained in documents stored in the knowledge base 311 using a predetermined classification model, and assigns to each text chunk a descriptive tag that describes the structural characteristics of the text chunk in the document. The tag management unit 222 will be described in detail later, and therefore will not be described here.

[0042] The representation generator 223 is a functional unit that processes text chunks of a document using a predetermined natural language processing means to generate a first graphical representation showing a set of n-dimensional data points corresponding to the text chunks and having spatial relationships based on semantic similarities between the text chunks. Here, the first graphical representation may be two-dimensional or three-dimensional. The expression generating unit 223 will be described in detail later, and therefore will not be described here.

[0043] The classification unit 224 is a functional unit that determines whether a text chunk in a document corresponds to a semantic element or a structural element of the document by using the descriptive tags assigned to the text chunks by the tag management unit 222 and the classification rules 312 stored in the storage unit 211. The classification unit 224 may assign semantic labels to data points corresponding to the semantic elements of the document and may assign structural labels to data points corresponding to the structural elements of the document. The classification unit 224 may also generate a second graph representation in which the semantic labels and structural labels are assigned to the data points and output the second graph representation via a user interface. The details of the classification unit 224 will be described later, and therefore will not be described here.

[0044] The exclusion unit 225 is a functional unit that excludes text chunks corresponding to data points to which structural labels have been assigned from documents stored in the knowledge base 311, based on a user instruction input for the second graph representation output from the classification unit 224. In this disclosure, the expression "exclude" means deleting, removing, or excluding a text chunk from the knowledge base. The details of the exclusion unit 225 will be described later, and therefore will not be described here.

[0045] As shown in FIG. 3, the RAG application 226 may include a query receiving unit 227, an information searching unit 228, and a response output unit 229. The query receiving unit 227 is a functional unit that receives a query input via, for example, the user terminal 250 shown in Fig. 2. The query here may be, for example, a query for the RAG service that uses a large-scale language model and the knowledge base 311 according to an embodiment of the present disclosure. The content of the query is not particularly limited, and may be, for example, "Tell me about famous physicists of the 20th century" or "Tell me the cause of the collapse of the Roman Empire."

[0046] The information search unit 228 is a functional unit that searches the knowledge base 311 for information to answer a query based on the query received by the query receiving unit 227. The search here is not particularly limited, and any existing search means may be used.

[0047] Answer output unit 229 is a functional unit that generates an answer to the query received by query receiving unit 227, based on the information searched by information search unit 228. The means for generating an answer based on the information searched from knowledge base 311 is not particularly limited, and a means used in an existing RAG means may be used.

[0048] 3 described above, it is possible to visualize the knowledge stored in the knowledge base and facilitate the identification and removal of information with less semantic content from the knowledge base. Furthermore, by removing information with less semantic content from the documents stored in the knowledge base, it is possible to improve the accuracy of searching for appropriate knowledge to answer a query and improve the accuracy of the RAG means.

[0049] Next, with reference to FIG. 4, the overall flow of the database management means according to the embodiment of the present disclosure will be described.

[0050] 4 is a diagram showing an example of an overall flow 400 of a database management means according to an embodiment of the present disclosure. As described above, the database management means according to an embodiment of the present disclosure is a means for visualizing knowledge stored in a knowledge base and promoting the identification and elimination of information with less semantic content from the knowledge base, and may be implemented by the database management device 210, classification model 230, control terminal 240, user terminal 250, etc., described with reference to FIGS.

[0051] First, in step S410, the embedding manager 221 (not shown in FIG. 4) according to an embodiment of the present disclosure performs a predetermined vector conversion means, such as an embedding model like a Sentence Transformer, on a document written in a natural language to convert a set of text chunks included in the document into a set of first embedding vectors, and stores the first embedding vectors in the knowledge base 311. As described above, the document here may be a document selected by the user of the control terminal 240 to be added to the knowledge base 311. Also, the text chunk here refers to a sentence of a predetermined size (a predetermined number of characters, words, lines, etc.) in the document.

[0052] Next, in step S420, the embedding manager 221 generates a set of dimension-reduced first embedding vectors by applying a predetermined dimension reduction method to the set of first embedding vectors generated in step S410. This set of dimension-reduced first embedding vectors includes a set of data points that represent the features (semantic features or structural features) of each text chunk in the document in the form of an n-dimensional vector.

[0053] Next, in step S430, the tag management unit 222 (not shown in FIG. 4) assigns, to each text chunk corresponding to each data point in the first set of embedding vectors, a description tag that describes the structural characteristics of the text chunk in the document. The structural characteristics here are information that indicates the role, function, or purpose of the text chunk in the document, and may be, for example, "body," "heading," "header," "footer," etc.

[0054] In step S440, the representation generator 223 (not shown in FIG. 4) processes the first set of embedding vectors using a predetermined natural language processing means to generate a first graphical representation showing a set of n-dimensional data points having spatial relationships based on semantic similarities between the text chunks, where the first graphical representation may be two-dimensional or three-dimensional.

[0055] Next, in step S450, the classification unit 224 (not shown in FIG. 4) generates a second graph representation in which a semantic label or a structural label is assigned to each data point included in the set of data points in the first graph representation by comparing the descriptive tags assigned to the text chunks by the tag management unit 222 described above with classification rules that define whether a text chunk in the document corresponds to a semantic element or a structural element of the document, and outputs the second graph representation to the user terminal 250 (not shown in FIG. 4). Here, the classification unit 224 may display a single-document representation showing a graph representation of a single document, or a multiple-document representation in which graph representations of multiple documents are overlaid.

[0056] A semantic label here indicates that a text chunk corresponding to a specific data point is a sentence related to the essential semantic content of a document. A structural label here indicates that a text chunk corresponding to a specific data point is a sentence related to the format or structural content of a document. By assigning a "semantic label" or "structural label" to a text chunk corresponding to a specific data point, it becomes possible to distinguish between a text chunk containing meaningful semantic content that can be used by a large-scale language model to generate a user query and a text chunk containing structural content that is unnecessary for the large-scale language model to generate an answer.

[0057] In some embodiments, the classifier 224 may also use a distance algorithm to group each data point in the second graph representation into a set of clusters and assign a semantic or structural label to each cluster based on the number of semantic and structural labels in each cluster, thereby enabling the determination of text chunks that are desirable to exclude from the document for each cluster that contains a large number of data points.

[0058] Next, in step S460, the exclusion unit 225 (not shown in FIG. 4) excludes from the document text chunks corresponding to clusters (or data points) to which structural labels have been assigned, based on user instructions input with respect to the second graph representation generated in step S450 and output to the user terminal 250. More specifically, for each cluster (or data point) displayed in the second graph representation, the user of the user terminal 250 determines whether the cluster is a text chunk containing meaningful semantic content for use in the large-scale language model. If the user inputs a user instruction indicating that the text chunk corresponding to a particular cluster or data point is determined to be a text chunk containing structural content unnecessary for answer generation by the large-scale language model rather than a text chunk containing meaningful semantic content, the exclusion unit 225 excludes the text chunk from the documents stored in the knowledge base 311.

[0059] 4, the database management means can distinguish between text chunks containing meaningful semantic content that can be used in a large-scale language model to generate a user query and text chunks containing structural content that is unnecessary for generating an answer using the large-scale language model, and can exclude structural information with less semantic content, such as document headers and footers, from the documents stored in the knowledge base. This can improve the accuracy of searching for appropriate knowledge to answer a query and the accuracy of the RAG means.

[0060] Next, with reference to FIG. 5, the configuration of the knowledge base according to the embodiment of the present disclosure will be described.

[0061] 5 is a diagram illustrating an example of a configuration of a knowledge base 311 according to an embodiment of the present disclosure. As illustrated in FIG. 5, the knowledge base 311 may store information on text chunks 510, embedding vectors 520, and metadata 530.

[0062] A text chunk 510 is information indicating a sentence of a predetermined size (number of characters, number of words, etc.) composed of natural language contained in a particular document. For example, if a particular document is divided into text chunks of 50 words, each of the 50-word text chunks may be stored in the knowledge base 311. As described above, these text chunks 510 include text chunks related to the essential semantic content of the document (sometimes referred to as "semantic text chunks" in this disclosure) and text chunks for defining the structure and format of the text in the document or for providing auxiliary information (sometimes referred to as "structural text chunks" in this disclosure). While the semantic text chunks contain meaningful semantic content that is used by the large-scale language model to generate a user query, the structural text chunks may contain information that is unnecessary for the large-scale language model to generate an answer. Therefore, one aspect of the present disclosure relates to identifying these semantic text chunks and structural text chunks and excluding the structural text chunks from the document. This improves the accuracy of searching for appropriate knowledge to answer a query and improves the accuracy of the RAG means.

[0063] The embedding vector 520 is information obtained by converting the text chunk 510 into a vector format. Converting the text chunk 510 into the embedding vector 520 and performing a predetermined dimensionality reduction method enables statistical analysis using a clustering method or visualization using an n-dimensional (e.g., 2D or 3D) graph representation. In particular, processing the high-dimensional embedding vector 520 using a dimensionality reduction method such as UMAP (Uniform Manifold Approximation and Projection) can compress the high-dimensional data into a low-dimensional space while maintaining important features of the data chunk corresponding to the embedding vector 520, thereby reducing the processing load.

[0064] Metadata 530 is information that indicates attributes or characteristics of text chunk 510. For example, in one embodiment, metadata 530 may be information that indicates the document from which text chunk 510 originated, the location of text chunk 510 in that document, etc.

[0065] By using the knowledge base 311 shown in Figure 5 in conjunction with any RAG tool, the knowledge contained in the large-scale language model can be supplemented with text chunks stored in the knowledge base 311, facilitating more reliable and accurate answers.

[0066] Next, classification rules according to an embodiment of the present disclosure will be described with reference to FIG.

[0067] 6 is a diagram illustrating an example of the configuration of the classification rules 312 according to an embodiment of the present disclosure. As described above, the classification rules 312 according to an embodiment of the present disclosure are rules that specify, for each text chunk of a document stored in the knowledge base 311 described with reference to FIG. 5, whether the text chunk corresponds to a semantic element or a structural element of the document. In an embodiment, by describing the structural characteristics of the text chunk and comparing the descriptive tag assigned to the text chunk with the classification rules 312, it is possible to determine whether the text chunk corresponds to a semantic element or a structural element of the document.

[0068] As shown in FIG. 6, in one embodiment, classification rules 312 may include descriptive tag 610 and category 620 information. The description tag 610 is information describing the structural characteristics of the text chunk in the document. The structural characteristics here are information indicating the role, function, or purpose of the text chunk in the document, and may be, for example, "body," "heading," "header," "footer," etc. In one embodiment, the description tag 610 may be an HTML tag assigned to the text chunk by a predetermined classification model such as a large-scale language model. The HTML tag here may be, for example, as shown in FIG. 6: <article> 、 、 <h1>~< / h1> <h6>、 <form> 、 <figcaption> 、 <header> 、 <footer> 、 <pre>、 <address>、 and

[0069] <article>is a tag that indicates the body of text in a document, identifying a self-contained section, such as an article, blog, or comment. is a tag that indicates a paragraph in a document. <h1>~< / h1> <h6> is a tag that indicates a heading in a document, and is expressed in six levels of importance, from low to high.< / h6> <h1> indicates the most important heading,< / h1> <h6>indicates the heading with the lowest importance. is a tag that indicates a form intended for data submission from a user. The form here may include form controls such as input fields, buttons, radio buttons, etc. <figcaption>is a tag indicating a caption for a figure, illustration, etc. in a document, and may include information describing, for example, the caption of the image or figure. <header>is a tag that indicates a header within a document, and may include, for example, a heading at the top of a page, navigation links, a search form, etc. <footer>is a tag that indicates a footer within a document, which may contain, for example, copyright information, contact information, a site map, etc., at the bottom of the page. <pre>is a tag that indicates preformatted text. <address>is a tag that indicates contact information such as the creator or owner of the document. is a tag that indicates an unordered list in a document and may contain, for example, the items of the list. < / address> < / pre> < / footer> < / header> < / figcaption> < / h6> < / article> may include, and is not particularly limited here.

[0070] <article>and The HTML tags are mapped to "semantic elements" and <h1>~< / h1> <h6>、、 <figcaption> 、 <header> 、 <footer> 、 <pre>、 <address>、 and < / address> < / pre> < / footer> < / header> < / figcaption> < / h6> < / article> is a tag indicating a table in a document. Category 620 is information indicating whether the above-mentioned description tag 610 corresponds to a "semantic element" or a "structural element." As described above, the text chunks contained in a document include semantic text chunks related to the essential semantic content of the document and structural text chunks for defining the structure and format of the text in the document and for recording auxiliary information. By excluding structural text chunks from documents in a knowledge base, the accuracy of the RAG means using the knowledge base can be improved. As shown in Figure 6, in one embodiment,

[0071]

[0072]

[0073]

[0074] <article> 、 、、 <h1>~< / h1> <h6>、< / h6> < / article> may be associated with "structural elements." By comparing descriptive tags, such as HTML tags, assigned to text chunks in a document with the classification rules 312 shown in FIG. 6 , it is possible to determine whether each text chunk corresponds to a semantic element or a structural element of the document, and identify text chunks that are desirable to exclude from the document. Next, with reference to FIG. 7 , a prompt input to a classification model according to an embodiment of the present disclosure will be described. FIG. 7 is a diagram illustrating an example of a prompt 313 input to a classification model according to an embodiment of the present disclosure. The prompt 313 according to an embodiment of the present disclosure is information for requesting the classification model to assign descriptive tags to text chunks included in a document. As described above, the classification model 230 here may be a large-scale language model such as GPT-4, or a neural network trained to assign descriptive tags to text chunks. As shown in FIG. 7 , the prompt 313 may include information indicating the text chunk to which a descriptive tag is desired to be assigned, candidate descriptive tags to be assigned, and an instruction to request the assignment of the descriptive tag to the text chunk. As an example, when HTML tags are used as descriptive tags, the prompt 313 may be: <header> 、 <footer> 、 <figcaption> 、 <pre>、 , and <address>However, the present disclosure is not limited to this, and other HTML tags and descriptive tags other than HTML tags may also be freely defined. Note that while prompt 313 shown in FIG. 7 illustrates an example of requesting the assignment of a descriptive tag to one text chunk, the present disclosure is not limited thereto, and prompt 313 may request the assignment of descriptive tags to multiple text chunks.

[0075] Upon receiving the prompt 313, the classification model executes natural language processing on the text chunk, analyzes it, and selects and assigns a description tag that appropriately describes the constituent features of the text chunk from the candidate description tags defined in the prompt 313. By performing this process for each text chunk included in the document, a description tag can be assigned to each text chunk in the document. As described above, by comparing the description tag with the classification rules according to embodiments of the present disclosure, it is possible to determine whether each text chunk corresponds to a semantic element or a structural element.

[0076] Next, with reference to FIG. 8, the flow of a graph representation generation process according to an embodiment of the present disclosure will be described.

[0077] Fig. 8 is a diagram showing an example of the flow of a graph representation generation process 800 according to an embodiment of the present disclosure. The graph representation generation process 800 shown in Fig. 8 is a process for visualizing text chunks included in a document and generating a graph representation shown as n-dimensional data points (where "n" may be 2 or 3), and is performed by the embedding management unit 221, tag management unit 222, and representation generation unit 223 of the database management device 210 shown in Figs. 2 and 3.

[0078] First, in step S810, the embedding management unit 221 according to an embodiment of the present disclosure performs a predetermined vector conversion means, such as an embedding model like a Sentence Transformer, on a document composed of natural language, thereby converting a set of text chunks contained in the document into a set of first embedding vectors and storing them in the knowledge base 311.

[0079] Next, in step S820, the embedding manager 221 applies a predetermined dimension reduction method to the set of first embedding vectors generated in step S810 to generate a set of dimension-reduced first embedding vectors. The dimension reduction method may be, for example, UMAP (Uniform Manifold Approximation and Projection). By applying the predetermined dimension reduction method to the set of first embedding vectors to generate a set of dimension-reduced first embedding vectors, it is possible to compress high-dimensional data into a low-dimensional space while maintaining important features of the data chunks, thereby reducing the processing load. This first set of dimension-reduced embedding vectors includes a set of data points that represent the features (semantic features or structural features) of each text chunk in the document in the form of a two-dimensional vector.

[0080] Next, in step S830, the tag manager 222 assigns, to the text chunks corresponding to each data point in the first set of embedding vectors, a description tag that describes the structural characteristics of the text chunk in the document. In one embodiment, the tag manager 222 inputs the prompt 313 described with reference to FIG. 6 into a predetermined classification model, such as a large-scale language model, and assigns the following HTML tag for the text chunk: <article> 、 、 <h1>~< / h1> <h6>、、 <figcaption> 、 <header> 、 <footer> 、 <pre>、 <address>、 and < / address> < / pre> < / footer> < / header> < / figcaption> < / h6> < / article> < / address> < / pre> < / figcaption> < / footer> < / header> 、

[0081]

[0082]

[0083]

[0084]

[0085]

[0086]

[0087] <article>or assigning semantic labels to data points corresponding to text chunks to which the HTML tags of <h1>~< / h1> <h6>、、 <figcaption> 、 <header> 、 <footer> 、 <pre>、 <address>、 and < / address> < / pre> < / footer> < / header> < / figcaption> < / h6> < / article> The document may request assignment of a description tag, etc., to each text chunk included in the document, describing the structural characteristics of the text chunk. As will be described later, by comparing the description tag with the classification rules 312 described above, it is possible to determine whether each text chunk corresponds to a semantic element or a structural element. In step S840, the representation generation unit 223 processes the first set of embedding vectors using a predetermined natural language processing means to generate a first graph representation representing a set of n-dimensional data points having spatial relationships based on the semantic similarity between the text chunks. In this first graph representation, the distribution of data points corresponding to text chunks in the document is determined by the semantic similarity between the text chunks. That is, in the first graph representation, data points corresponding to text chunks with high semantic similarity are arranged closer to each other, and data points corresponding to text chunks with low semantic similarity are arranged farther from each other. According to the graph representation generation process 800 described above, it is possible to visualize text chunks included in a document and generate a graph representation shown as n-dimensional (e.g., two-dimensional or three-dimensional) data points. As will be described later, in this n-dimensional graph representation, text chunks corresponding to semantic elements in the document and text chunks corresponding to structural elements in the document are distinguished from each other and presented to the user, thereby facilitating the elimination of unnecessary information included in the knowledge base that is not used to generate an answer in the RAG means. Next, with reference to FIG. 9, a flow of a text chunk classification process according to an embodiment of the present disclosure will be described. FIG. 9 is a diagram showing an example of the flow of the text chunk classification process 900 according to an embodiment of the present disclosure. The text chunk classification process 900 according to an embodiment of the present disclosure is a process for classifying data points corresponding to text chunks of a document based on classification rules in the first graph representation generated by the graph representation generation process 800 described with reference to FIG. 8 and assigning semantic labels or structural labels, and is performed by the classifier 224 shown in FIGS. 2 and 3.First, in step S910, the classifier 224 groups each data point in the first graph representation into a set of clusters by using a predetermined distance algorithm. The clustering method used to group the data points into the set of clusters is not particularly limited, and may be, for example, K-means or another existing clustering method. Next, in step S920, the classifier 224 assigns a semantic label or a structural label to each data point included in the set of data points in the first graph representation by comparing the descriptive tags assigned to the text chunks in the graph representation generation process 800 described above with classification rules that define whether the text chunks in the document correspond to semantic elements or structural elements of the document. As an example, if the descriptive tags are HTML tags, the classifier 224 may assign a semantic label or a structural label to each data point included in the set of data points in the first graph representation.

[0088] <article>or The text chunks assigned the HTML tag correspond to the main body of the document, which contains the essential semantic content, whereas <h1>~< / h1> <h6>、、 <figcaption> 、 <header> 、 <footer> 、 <pre>、 <address>、 and < / address> < / pre> < / footer> < / header> < / figcaption> < / h6> < / article> A structural label may be assigned to a data point that corresponds to a chunk of text that has been assigned one of the HTML tags:

[0089]

[0090]

[0091]

[0092]

[0093]

[0094]

[0095]

[0096]

[0097]

[0098]

[0099]

[0100]

[0101]

[0102]

[0103]

[0104]

[0105]

[0106]

[0107]

[0108]

[0109]

[0110]

[0111]

[0112]

[0113]

[0114]

[0115]

[0116] <article> 、 、 <h1>~< / h1> <h6>、、 <figcaption> 、 <header> 、 <footer> 、 <pre>、 <address>、 and < / address> < / pre> < / footer> < / header> < / figcaption> < / h6> < / article> This is because text chunks assigned one of the HTML tags correspond to formal elements, such as headers, footers, and tables, used to organize text in a document. By assigning a "semantic label" or a "structural label" to a text chunk corresponding to a specific data point, it is possible to distinguish between text chunks containing meaningful semantic content that can be used in a large-scale language model to generate a user query and text chunks containing structural content that is unnecessary for the large-scale language model to generate an answer. In one embodiment, the classifier 224 may generate a second graph representation in which a "semantic label" or a "structural label" is assigned to each data point, and output the second graph representation via the control terminal 240. Next, in step S930, the classifier 224 counts, for each cluster included in the second graph representation, the number of semantic data points indicating the number of data points assigned with a semantic label and the number of structural data points indicating the number of data points assigned with a structural label. If the number of semantic data points is greater than the number of structural data points, the process proceeds to step S935. If the number of structural data points is greater than the number of semantic data points, the process proceeds to step S940. In step S935, semantic data labels are assigned to all data points included in clusters whose semantic data points are greater than the number of structural data points. On the other hand, in step S940, structural data labels are assigned to all data points included in clusters whose structural data points are greater than the number of semantic data points. In addition, in one embodiment, the classifier 224 may display exclusion recommendation information in the second graph representation, recommending the exclusion of data points to which structural labels have been assigned. This exclusion recommendation information may be a message recommending exclusion, or may be a display format recommending exclusion (e.g., red display, highlighting, etc.). This allows a user viewing the graph representation to make an exclusion decision for an entire cluster containing a large number of data points without having to check each data point one by one, thereby saving the user time and effort.According to the text chunk classification process 900 described with reference to FIG. 9 , semantic text chunks related to the essential semantic content of a document are distinguished from structural text chunks for defining the structure and format of the text in the document or for providing auxiliary information in the above-described graph representation, making it easier to identify text chunks (e.g., structural text chunks) that are desirable to exclude from the knowledge base. Next, with reference to FIG. 10 , a flow of the exclusion process according to an embodiment of the present disclosure will be described. FIG. 10 is a diagram showing an example of the flow of the exclusion process 1000 according to an embodiment of the present disclosure. The exclusion process 1000 according to an embodiment of the present disclosure is a process for excluding specific data chunks from a document in the knowledge base based on a user instruction input to an n-dimensional graph representation (second graph representation) in which semantic labels or structural labels are assigned to data points corresponding to text chunks in the document, and may be performed by the classifier 224 and the exclusion unit 225 shown in FIGS. 2 and 3 . First, in step S1010, the classification unit 224 presents to the user the second graph representation generated by the above-described text chunk classification process 900 (i.e., a graph representation in which semantic labels or structural labels are assigned to data points corresponding to text chunks in a document). The classification unit 224 may present the second graph representation via a GUI (Graphical User Interface) displayed on the display unit 243 of the control terminal 240 shown in FIG. 2, for example. In one embodiment, the classification unit 224 may display data points in the second graph representation to which structural labels are assigned, as candidates corresponding to text chunks that are desirable to exclude from the document, in a different color or symbol from that of data points to which semantic labels are assigned. Next, in step S1020, the exclusion unit 225 accepts a user instruction input to the second graph representation presented in step S1010. This user instruction may be an instruction to select a specific cluster from among the clusters displayed in the second graph representation, and may be input by operating a mouse, a keyboard, or a touch panel.Next, in step S1030, the exclusion unit 225 lists, in table form, the data points included in the cluster selected by the user instruction received in step S1020, and accepts the user's selection of a specific data point. When a specific data point is selected by the user, the exclusion unit 225 may present information about the data chunk corresponding to the selected data point. This allows the user to check the content of the text chunk corresponding to the data point of interest and easily consider whether it is desirable to exclude it from the document. Next, in step S1040, the exclusion unit 225 receives a user instruction to select a data point to be excluded from the second graph representation, and excludes the text chunk corresponding to the data point selected for exclusion from the document in the knowledge base. The exclusion unit 225 also updates the second graph representation presented to the user by deleting the data point selected for exclusion. According to the above-described exclusion process 1000, by presenting a user with an n-dimensional graph representation in which semantic or structural labels are assigned to data points corresponding to text chunks in a document, semantic text chunks related to the essential meaning of the document can be distinguished from structural text chunks that define the structure and format of the text in the document or provide auxiliary information. This makes it easier to determine text chunks that should be excluded from the knowledge base. FIGS. 11 and 12 illustrate a user interface according to an embodiment of the present disclosure. FIG. 11 is a diagram illustrating an example of a graph representation display screen 1100 according to an embodiment of the present disclosure. The graph representation display screen 1100 shown in FIG. 11 is a screen for presenting a user with a second graph representation showing data points corresponding to text chunks included in a document stored in the knowledge base. The graph representation display screen 1100 may be output to the user via, for example, the control terminal 240 shown in FIG. 2. As shown in FIG. 11, the graph representation display screen 1100 mainly includes a management menu 1110 and a display window 1120. The management menu 1110 provides tools for filtering and differentiating the data points displayed in the display window 1120 .For example, as shown in FIG. 11 , the management menu 1110 may include a data point color button for changing the display color of a specific data point, a data point sign button for changing the sign of a specific data point, a data point magnitude button for changing the magnitude of a specific data point, a description tag button for selecting all data points assigned a specific description tag, a label button for selecting all data points assigned a specific label (semantic label or structural label), and a chunk size button for selecting all data points corresponding to text chunks having a specific size. The management menu 1110 may also include a filter tool for filtering the number of data points by description tag, label, or text chunk size. The display window 1120 displays a graphical representation (second graphical representation) showing n-dimensional data points corresponding to text chunks included in the document. As described above, in this graphical representation, the distribution of data points corresponding to text chunks in the document is determined by the semantic similarity between the text chunks. That is, in the graphical representation displayed in the display window 1120, data points corresponding to text chunks with high semantic similarity are arranged closer to each other, and data points corresponding to text chunks with low semantic similarity are arranged farther from each other. Note that, like the first graphical representation described above, the second graphical representation may be two-dimensional or three-dimensional. Also, as shown in FIG. 11 , in an embodiment, structural clusters including data points assigned with structural labels and semantic clusters including data points assigned with semantic labels may be displayed with different symbols. Also, in an embodiment, exclusion recommendation information 1122 recommending the exclusion of data points assigned with structural labels may be displayed in the display window 1120. This exclusion recommendation information 1122 may be, for example, a message recommending exclusion, as shown in FIG. 11 . FIG. 12 is a diagram illustrating an example of a detail confirmation screen 1200 according to an embodiment of the present disclosure.12 is a screen for checking details of data points displayed on the graph representation display screen 1100 shown in FIG. 11 and making an exclusion decision, and may be output to a user via, for example, the control terminal 240 shown in FIG. 2. When a user selects a specific data point or cluster on the graph representation display screen 1100 described with reference to FIG. 11, a details window 1230 related to the selected data point or cluster is displayed. As shown in FIG. 12, the details window 1230 may include a text chunk confirmation window 1232 showing part of the content of a text chunk corresponding to the selected data point, a text chunk expansion button 1234 for expanding a specific text chunk in the text chunk confirmation window 1232 to display the full text, an exclusion target selection button 1236 for selecting a text chunk to be excluded, and an exclusion text chunk button 1238 for excluding a text chunk selected with the exclusion target selection button 1236 from the corresponding document. The user interface according to the embodiment of the present disclosure, described with reference to FIGS. 11 and 12 , allows a user managing a knowledge base to easily identify semantic text chunks related to essential semantic content and structural text chunks for defining the structure and format of the text in the document or for providing auxiliary information, and to identify text chunks that should be excluded from the knowledge base. In one aspect of the database management means according to the embodiment of the present disclosure, descriptive tags describing the structural characteristics of a text chunk (e.g., "body," "header," "table") can be assigned to natural language text chunks included in documents stored in the knowledge base. These descriptive tags may be, for example, HTML tags. By comparing the descriptive tags assigned to the text chunks with predetermined classification rules, it is possible to determine whether each text chunk included in a document stored in the knowledge base is a sentence related to the essential semantic content of the document or a sentence used to organize the format and text in the document.By assigning semantic labels to text chunks related to the essential semantic content of a document and assigning structural labels to text chunks related to the essential semantic content of a document, and then generating a graph representation showing n-dimensional data points corresponding to the text chunks and having a distribution based on the semantic similarity between each text chunk in the document, users managing a knowledge base can easily check the semantic and structural elements of documents contained in the knowledge base in a visualized form. This makes it easy to understand the semantic and structural relationships of documents contained in the knowledge base, reducing the effort required to manage the knowledge base. Furthermore, by excluding text chunks assigned structural labels from documents, information unnecessary for generating answers using a large-scale language model can be preferentially removed from the knowledge base. This solves the problem of the knowledge base containing a huge amount of information making it difficult to search for meaningful knowledge and limiting the accuracy of RAG, thereby improving the accuracy of the RAG means. As described above, the database management means according to an embodiment of the present disclosure includes the following aspects.(Aspect 1) A database management device, the database management device comprising: a processor; a memory; and a storage unit for storing a knowledge database including one or more documents, the memory including a tag management unit for processing a first set of text chunks in a first document included in the documents using a predetermined classification model, thereby assigning to each text chunk included in the set of first text chunks a description tag that describes a constituent feature of the text chunk in the first document; and a first tag management unit for processing the set of first text chunks using a predetermined natural language processing means, thereby assigning a first tag that indicates a set of n-dimensional data points that correspond to the set of first text chunks and have a spatial relationship based on semantic similarity between the text chunks. a classification unit that generates and outputs a second graph representation in which semantic labels or structural labels are assigned to each data point included in the set of data points in the first graph representation by comparing the descriptive tags assigned to the first set of text chunks with classification rules that specify whether a text chunk in a document corresponds to a semantic element or a structural element of the document; and an exclusion unit that excludes from the first document, first text chunks in the set of data points that correspond to first data points to which the structural labels are assigned, based on a user instruction input to the second graph representation. (Aspect 2) The database management device according to aspect 1, further includes an embedding management unit that converts the set of first text chunks into a set of first embedding vectors and applies a predetermined dimension reduction means to the set of first embedding vectors to generate a set of dimension-reduced first embedding vectors, and the representation generation unit generates the first graph representation by processing the set of dimension-reduced first embedding vectors using the predetermined natural language processing means.(Aspect 3) The database management device according to aspect 1 or 2, wherein the tag management unit assigns an HTML tag as the description tag to each text chunk included in the set of first text chunks. (Aspect 4) The HTML tag is:

[0117] <article>or assigning said semantic labels to data points corresponding to text chunks to which HTML tags of For each data point in the set of data points in the first graphical representation: <h1>~< / h1> <h6>、、 <figcaption> 、 <header> 、 <footer> 、 <pre>、 <address>、 and < / address> < / pre> < / footer> < / header> < / figcaption> < / h6> < / article> The database management device according to aspect 3 includes: (Aspect 5) wherein the classification unit classifies each of the data points included in the set of data points in the first graph representation into

[0118]

[0119]

[0120]

[0121] and assigning the structural label to a data point corresponding to a text chunk to which any one of the HTML tags in the following is assigned: (Aspect 6) The database management device according to aspect 5, wherein the classifier uses a predetermined distance algorithm to group the set of data points in the first graph representation into a set of clusters, and for a first cluster in the set of clusters, determines a semantic data point count indicating the number of data points to which the semantic label is assigned and a structural data point count indicating the number of data points to which the structural label is assigned, and if the semantic data point count is greater than the structural data point count, assigns the semantic label to each data point included in the first cluster in the second graph representation, and if the structural data point count is greater than the semantic data point count, assigns the structural label to each data point included in the first cluster in the second graph representation. (Aspect 7) The database management device according to any one of Aspects 1 to 6, wherein the classification unit displays, in the second graph representation, exclusion recommendation information that recommends excluding from the first document the text chunk corresponding to the data point to which the structural label is assigned, and presents this to the user via a user interface. While the embodiments of the present invention have been described above, the present invention is not limited to the above-described embodiments, and various modifications are possible without departing from the spirit of the present invention. [Explanation of Symbols] 150 Database management application 210 Database management device 211 Storage unit 212 Arithmetic unit 213 Network interface 214 Memory 221 Embedding management unit 222 Tag management unit 223 Expression generation unit 224 Classification unit 225 Exclusion unit 226 RAG application 227 Query reception unit 228 Information search unit 229 Answer output unit 305 Bus 310 Database 311 Knowledge base 312 Classification rules 313 Prompt < / address> < / pre> < / footer> < / header> < / figcaption> < / form> < / h6> < / article>

Claims

1. a processor; Memory and a memory unit for storing a knowledge database including one or more documents, The memory includes: a tag management unit that processes a first set of text chunks in a first document included in the document using a predetermined classification model, thereby assigning to each text chunk included in the first set of text chunks a description tag that describes a structural feature of the text chunk in the first document; a representation generator that processes the first set of text chunks using a predetermined natural language processing means to generate a first graphical representation showing a set of n-dimensional data points corresponding to the first set of text chunks and having spatial relationships based on semantic similarities between the text chunks; a classifier that generates and outputs a second graph representation in which a semantic label or a structural label is assigned to each data point included in the set of data points in the first graph representation by comparing the descriptive tags assigned to the first set of text chunks with classification rules that define whether a text chunk in a document corresponds to a semantic element or a structural element of the document; an exclusion unit that excludes from the first document a first text chunk corresponding to a first data point in the set of data points to which the structural label is assigned, based on a user instruction input to the second graphical representation; a processing instruction for causing the processor to execute the processing instruction as a database management device;

2. The database management device converting the first set of text chunks into a first set of embedding vectors; an embedding manager that applies a predetermined dimension reduction method to the first set of embedding vectors to generate a dimension-reduced first set of embedding vectors; The expression generation unit generating the first graph representation by processing the dimension-reduced first embedding vector set by the predetermined natural language processing means; 2. The database management device according to claim 1.

3. The tag management unit assigning an HTML tag as the description tag to each text chunk in the first set of text chunks; 2. The database management device according to claim 1.

4. The HTML tag is: Including <article>, , <h1> to <h6>, <form>, <figcaption>, <header>, <footer>, <pre>, <address>, and , 4. The database management device according to claim 3.

5. The classification unit assigning the semantic label to each data point in the set of data points in the first graphical representation that corresponds to a text chunk that is assigned an <article> or HTML tag; assigning the structural label to a data point, among each data point included in the set of data points in the first graph representation, that corresponds to a text chunk to which any one of HTML tags <h1> to <h6>, <form>, <figcaption>, <header>, <footer>, <pre>, <address>, , and is assigned; 5. The database management device according to claim 4.

6. The classification unit grouping the set of data points in the first graphical representation into a set of clusters using a predetermined distance algorithm; determining, for a first cluster in the set of clusters, a semantic data point count indicating the number of data points to which the semantic label is assigned and a structural data point count indicating the number of data points to which the structural label is assigned; if the number of semantic data points is greater than the number of structural data points, assigning the semantic label to each data point included in the first cluster in the second graph representation; if the number of structural data points is greater than the number of semantic data points, assigning the structural label to each data point included in the first cluster in the second graph representation; 6. The database management device according to claim 5.

7. The classification unit displaying, in the second graphical representation, exclusion recommendation information that recommends excluding from the first document the text chunks that correspond to the data points to which the structural labels are assigned, and presenting the information to a user via a user interface; 2. The database management device according to claim 1.

8. The classification unit excluding the first data point from the second graphical representation after excluding the first text chunk from the first document; 2. The database management device according to claim 1.

9. A database management method implemented in a database management device, comprising: The database management device a processor; Memory and a memory unit for storing a knowledge database including one or more documents; The database management method includes: The processing instructions stored in the memory processing a set of first text chunks in a first document included in the document using a predetermined classification model, thereby assigning to each text chunk included in the set of first text chunks one or more of <article>, , <h1> to <h6>, <form>, <figcaption>, <header>, <footer>, <pre>, <address>, and as HTML tags describing the structural characteristics of the text chunk in the first document; processing the first set of text chunks using a natural language processing means to generate a first graphical representation showing a set of n-dimensional data points corresponding to the first set of text chunks and having spatial relationships based on semantic similarities between the text chunks; generating a second graphical representation in which, by comparing the HTML tags assigned to the first set of text chunks with classification rules that define whether a text chunk in a document corresponds to a semantic element or a structural element of the document, semantic labels are assigned to data points in the set of data points in the first graphical representation that correspond to text chunks that are assigned an HTML tag of <article> or , and structural labels are assigned to data points in the set of data points in the first graphical representation that correspond to text chunks that are assigned an HTML tag of any one of <h1> to <h6>, <form>, <figcaption>, <header>, <footer>, <pre>, <address>, , and ; excluding, based on a user instruction input to the second graphical representation, a first text chunk from the first document that corresponds to a first data point in the set of data points to which the structural label is assigned; excluding the first data point from the second graphical representation after excluding the first text chunk from the first document; A database management method characterized by causing the processor to execute the above steps.

10. a database management device; A database management system connected to a user terminal via a communication network, The database management device a processor; Memory and a memory unit for storing a knowledge database including one or more documents; The memory includes: a tag management unit that processes a first set of text chunks in a first document included in the document using a predetermined classification model, thereby assigning to each text chunk included in the first set of text chunks a description tag that describes a structural feature of the text chunk in the first document; a representation generator that processes the first set of text chunks using a predetermined natural language processing means to generate a first graphical representation showing a set of n-dimensional data points corresponding to the first set of text chunks and having spatial relationships based on semantic similarities between the text chunks; a classification unit that generates a second graph representation in which a semantic label or a structural label is assigned to each data point included in the set of data points in the first graph representation by comparing the descriptive tags assigned to the first set of text chunks with classification rules that define whether a text chunk in a document corresponds to a semantic element or a structural element of the document, and outputs the second graph representation to the user terminal; an exclusion unit that excludes from the first document a first text chunk corresponding to a first data point in the set of data points to which the structural label is assigned, based on a user instruction input to the second graphical representation; and a processing instruction for causing the processor to execute the processing instruction as a database management system.

Images