Knowledge construction assistance device and method

Abstract

In an embodiment of the present invention, a component element set is generated by, first, dividing a resource document in a field related to an application included in at least the content of a user request into a plurality of component elements for each level in a document structure, vectorizing each of the plurality of divided component elements, and then associating the component elements with the highest level. Next, if a question sentence representing the application and a requirement included in the content of the user request is input, the question sentence is vectorized, and then a candidate for a component element having relevance to the question sentence is retrieved from the component element set on the basis of the question sentence and attribute information representing the location of the candidate for the component element in the component element set is acquired. A language generation model created in advance is used to generate an answer sentence based on the question sentence and the candidate for the component element, and assistance information including the answer sentence and the attribute information is presented to the user.

Description

Knowledge building support device and method

[0001] One aspect of the present invention relates to a knowledge construction support device and method used to quantify a user's request (Intent) in a system that provides a service that autonomously controls a network or computing resources in response to the user's Intent.

[0002] In recent years, Intent-based Management, which autonomously controls network and computing resources according to user requests (intents), has been studied. To realize this system, it is necessary to identify the specific requirements that users actually place on networks and services according to their vague intents, and to quantify the intents according to the identified requirements. Many requirements are defined, such as quality requirements, availability requirements, and priority requirements, depending on the application, and ChatBots are used as a method for identifying these requirements.

[0003] By the way, to identify requirements and quantify intent, it is necessary to query the knowledge graph and build appropriate knowledge. Knowledge generally exists in technical documents, operation systems, etc., and the following methods are known to build knowledge from these document files, etc.

[0004] One of these is a method of building knowledge based on grammar rules and statistics. While this method has the advantage of being easy to execute and fast, it suffers from the drawback of low accuracy. On the other hand, there are also methods of building knowledge using machine learning models such as neural networks and attention (see, for example, Non-Patent Document 1). This method allows for more accurate extraction than methods that use grammar rules and statistics, but it requires the collection of a huge amount of training data to build the model.

[0005] Another method is to use a language generation model (see, for example, Non-Patent Document 2). This method has the advantage that it does not require a dataset and can omit fine tuning. However, it generally has the drawback of low accuracy and requiring a large amount of computational resources.

[0006] Lingfeng Zhong, Jia Wu, Qian Li, Hao Peng, Xindong Wu: “A Comprehensive Survey on Automatic Knowledge Graph Construction”, Internet <URL: https: / / arxiv.org / abs / 2302.05019>, 10 Feb 2023Yuqi Zhu, Xiaohan Wang, Jing Chen, Shuofei Qiao, Yixin Ou, Yunzhi Yao, Shumin Deng, Huajun Chen, Ningyu Zhang: “LLMs for Knowledge Graph Construction and Reasoning: Recent Capabilities and Future Opportunities”, Internet <URL: https: / / arxiv.org / abs / 2305.13168>, 22 May 2023

[0007] As mentioned above, each of the knowledge construction methods proposed so far has problems that need to be solved, and improvements are eagerly awaited.

[0008] The present invention has been made in light of the above circumstances, and aims to provide a technology that reduces the cost of creating training data and enables the construction of highly versatile and highly accurate knowledge.

[0009] To solve the above-mentioned problems, one aspect of a knowledge construction support device or method according to the present invention supports the construction of knowledge used to identify and quantify requirements from a user's request by first dividing document documents in a field related to at least the application included in the user's request into multiple components based on the document structure hierarchy, vectorizing each of the multiple divided components, and linking them to the highest hierarchy to generate a component set. Next, when a question representing the application and requirements included in the user's request is input, the question is vectorized, and then, based on the question, component candidates related to the question are searched from the component set, and attribute information representing the position of the component candidates in the component set is obtained. Then, using a pre-created language generation model, an answer based on the question and the component candidates is generated, and support information including the answer and the attribute information is presented to the user.

[0010] According to one aspect of the present invention, answers corresponding to questions are generated using an existing language generation model, eliminating the need to prepare a large amount of training data compared to creating a new machine learning model for knowledge construction, thereby reducing development costs. Furthermore, by using a language generation model, it is possible to provide a system that is more versatile and accurate than when knowledge is extracted based on, for example, grammar rules or statistics.

[0011] That is, according to one aspect of the present invention, it is possible to provide a technology that reduces the cost of creating training data and enables the construction of highly versatile and highly accurate knowledge.

[0012] FIG. 1 is a diagram illustrating an example of a system including a knowledge building support device according to an embodiment of the present invention. FIG. 2 is a block diagram illustrating an example of a hardware configuration of the knowledge building support device of the system illustrated in FIG. 1. FIG. 3 is a block diagram illustrating an example of a software configuration of the knowledge building support device of the system illustrated in FIG. 1. FIG. 4 is a flowchart illustrating an example of a processing procedure and processing content of a process executed by a control unit of the knowledge building support device illustrated in FIG. 3 in a pre-setting mode. FIG. 5 is a flowchart illustrating an example of a processing procedure and processing content of a knowledge building support process executed by a control unit of the knowledge building support device illustrated in FIG. 3 in an operation mode. FIG. 6 is a diagram illustrating an example of a document decomposition process included in the pre-setting process illustrated in FIG. 4. FIG. 7 is a diagram illustrating an example of an answer sentence candidate generated by the knowledge building support process illustrated in FIG. 5. FIG. 8 is a diagram illustrating an example of a case where an answer sentence candidate is not generated in the knowledge building support process illustrated in FIG. 5.

[0013] Hereinafter, an embodiment of the present invention will be described with reference to the drawings.

[0014] [One Embodiment] (Configuration Example) (1) System FIG. 1 is a diagram showing an example of a system including a knowledge building support device according to one embodiment of the present invention.

[0015] The system of one embodiment enables the transmission of information data between a knowledge construction support device SV and multiple material databases DB1 to DBn, and between the knowledge construction support device SV and multiple user terminals UT1 to UTk, via a network NW.

[0016] The document databases DB1 to DBn are composed of database servers or terminal devices, and each stores data on document materials such as technical documents.

[0017] The user terminals UT1 to UTk consist of personal computers, tablet terminals or smartphones used by the users, and have the function of allowing the user to input question data and send it to the knowledge construction support device SV, and the function of receiving construction support information sent from the knowledge construction support device SV.

[0018] The network NW includes, for example, a wide area network centered on the Internet and an access network for accessing this wide area network. Examples of the access network include, but are not limited to, a public communication network using wired or wireless connections, a local area network (LAN) using wired or wireless connections, and a cable television (CATV) network. Note that, if the service provision area is limited to a company's business establishment or office, the network NW may be configured solely with a LAN or wireless LAN.

[0019] (2) Knowledge Building Support Device SV The knowledge building support device SV is configured by, for example, a server computer installed on the cloud or the Web. Note that the knowledge building support device SV may also be configured by, for example, a personal computer used by a system administrator.

[0020] 2 and 3 are block diagrams showing an example of the hardware and software configurations of the knowledge building support device SV, respectively.

[0021] The knowledge construction support device SV has a control unit 1 that uses a hardware processor such as a central processing unit (CPU), and this control unit 1 is connected via a bus 5 to a storage unit having a program storage unit 2 and a data storage unit 3, and a communication interface (hereinafter, interface will be referred to as I / F) unit 4.

[0022] The communication I / F section 4 uses a communication protocol defined in the network NW to transmit and receive information data to and from the material databases DB1 to DBn and the user terminals UT1 to UTk.

[0023] The program storage unit 2 is a combination of a nonvolatile memory that can be written to and read from at any time, such as an SSD (Solid State Drive), and a nonvolatile memory such as a ROM (Read Only Memory), and stores middleware such as an OS (Operating System) as well as application programs required to execute various controls according to an embodiment. Hereinafter, the OS and each application program will be collectively referred to as the program.

[0024] The data storage unit 3 is a combination of a non-volatile memory such as an SSD that can be written to and read from at any time, and a volatile memory such as a RAM (Random Access Memory), and its storage area includes a material data storage unit 31, an element set storage unit 32, and a language generation model storage unit 33.

[0025] The document data storage unit 31 is used to store data of document documents collected from the document databases DB1 to DBn. ​​The element set storage unit 32 is used to store sets of component elements obtained by dividing the document documents into a plurality of component elements by hierarchy and then vectorizing them. The language generation model storage unit 33 stores language generation models.

[0026] The control unit 1 has the following processing functions necessary to implement one embodiment of the present invention: a material data acquisition processing unit 11, a document segmentation processing unit 12, a vectorization processing unit for material documents 13, a question sentence data acquisition processing unit 14, a vectorization processing unit for question sentences 15, a candidate sentence search processing unit 16, an answer sentence generation processing unit 17, and a construction support information output processing unit 18.

[0027] The processing units 11 to 18 are all realized by causing a hardware processor in the control unit 1 to execute an application program stored in the program storage unit 2. Note that some or all of the processing units 11 to 18 may be realized using hardware such as an LSI (Large Scale Integration) or an ASIC (Application Specific Integrated Circuit).

[0028] The document data acquisition processing unit 11 accesses the document databases DB1 to DBn via the network NW in a state where the preset mode is specified, thereby collecting document data of document documents related to a predetermined field, and then stores the collected document data in the document data storage unit 31.

[0029] The document segmentation processing unit 12 segments the document stored in the document data storage unit 31 into a plurality of components according to the hierarchical level of the document structure. The smallest unit of segmentation is, for example, a "sentence." The document segmentation processing unit 12 then links the segmented components to the highest hierarchical level, thereby generating a set of components that maintains the document structure. An example of the segmentation process for a document will be described in the operation example.

[0030] The vectorization processing unit 13 of the document vectorizes each of the elements included in the generated set of elements, and stores the set of vectorized elements in the element set storage unit 32 .

[0031] With the operation mode specified, the question data acquisition processing unit 14 receives, via the communication I / F unit 4, question data created by the user, which is transmitted from the user terminals UT1 to UTk via the network NW.

[0032] The question vectorization processing unit 15 vectorizes the acquired question sentence and provides it to the candidate sentence search processing unit 16 .

[0033] Based on the vectorized question sentence, the candidate sentence search processor 16 searches the element set storage unit 32 for components highly related to the question sentence, such as candidate sentences. The candidate sentence search processor 16 also searches for attribute information on the candidate components. For example, if the candidate component is a sentence, the attribute information includes identification information for the document containing the component and information indicating the location of the sentence.

[0034] The answer sentence generation processing unit 17 inputs the vectorized question sentence and the vectorized candidate components into a language generation model stored in the language generation model storage unit 33, and generates an answer sentence corresponding to the question sentence from the language generation model.

[0035] The construction support information output processing unit 18 generates construction support information including the generated answer statement and attribute information of the component that was the basis for generating the answer statement, in other words, identification information of the document in which the answer statement is written and information indicating the position where the answer statement is written, and transmits the generated construction support information from the communication I / F unit 4 to the requesting user terminals UT1 to UTk via the network NW.

[0036] (Example of Operation) Next, an example of operation of the knowledge building support device SV configured as above will be described.

[0037] (1) Presetting Mode FIG. 4 is a flowchart showing an example of the processing procedure and processing contents of the material data presetting processing executed by the control unit 1 of the knowledge building support device SV.

[0038] (1-1) Acquisition of Material Documents When the control unit 1 of the knowledge building support device SV receives a request to specify a preset mode in step S10, under the control of the material data acquisition processing unit 11, in step S11, the control unit 1 selectively accesses the material databases DB1 to DBn and receives the material document data transmitted from the material databases DB1 to DBn via the network NW at the communication I / F unit 4. The received material document data is then stored in the material data storage unit 31. The material data acquisition processing unit 11 repeatedly performs the material document acquisition process until a predetermined number of material documents are collected.

[0039] The source documents to be acquired are selected from, for example, source documents in a field related to the expected user question, but source documents in an unspecified field may also be selected. The number of source documents to be acquired can also be set arbitrarily.

[0040] (1-2) Dividing the source document When the source document is acquired, the control unit 1 of the knowledge construction support device SV, under the control of the document division processing unit 12, performs a process of dividing the source document into multiple components according to the hierarchy of the document structure in step S12.

[0041] For example, as shown in Figure 6, the document segmentation processing unit 12 first divides the document structure into "chapters," "paragraphs," and "sentences" at each level. Then, for each level, the document is divided into a plurality of chapters based on, for example, headings, and each chapter is divided into a plurality of paragraphs, and each paragraph is further divided into a plurality of sentences. Then, for each chapter, the document segmentation processing unit 12 associates the paragraphs and sentences contained in the chapter with, for example, the chapter's identification number or title, thereby generating a set of sentences for each chapter that maintains the document structure, and temporarily stores the generated set of sentences in a storage area within the data storage unit 3.

[0042] In the above example, the minimum unit of division is a "sentence," but it may be a "phrase" or a "clause."

[0043] (1-3) Vectorization of Documents Next, in step S13, the control unit 1 of the knowledge construction support device SV, under the control of the vectorization processing unit 13, reads the sentence set from the data storage unit 3 one by one and vectorizes each of the multiple sentences constituting the sentence set while maintaining their semantic information. The vectorization processing unit 13 then stores the vectorized sentence set in the element set storage unit 32 of the data storage unit 3. Note that various well-known methods can be used for vectorization, such as those using morphological analysis or a method of converting sentences into vectors using distributed representations using a machine learning model.

[0044] At this time, the vectorization processing unit 13 also associates the identification information of the source document in which the sentence is written and the chapter identification number or title with the vectorized sentence and stores them as attribute information of the sentence. The attribute information may include paragraph numbers and line numbers in addition to the chapter identification number or title.

[0045] (1-4) Repeating the Pre-setting Process The control unit 1 of the knowledge building support device SV monitors the input of a setting end instruction in step S14. Then, until the setting end instruction is input, the control unit 1 returns to step S11 and repeatedly executes the series of document setting processes described above in steps S11 to S13. Then, when the input of the setting end instruction is detected, the control unit 1 returns to the standby state.

[0046] (2) Operation Mode FIG. 5 is a flowchart showing an example of the procedure and content of the knowledge building support process executed by the control unit 1 of the knowledge building support device SV.

[0047] When an operation mode is designated and this designation of the operation mode is recognized in step S20, the control unit 1 of the knowledge construction support device SV executes the knowledge construction support process as follows.

[0048] (2-1) Acquiring a question and vectorizing it When a user inputs a question expressing his / her desire (intent) into a user terminal UT1 to UTk, the data of the input question is transmitted from the user terminal UT1 to UTk along with a query to the knowledge construction support device SV.

[0049] In response to this, when the control unit 1 of the knowledge construction support device SV receives the query in step S21, under the control of the question data acquisition processing unit 14, in step S22 it receives the question data sent from the user terminals UT1 to UTk via the communication I / F unit 4.

[0050] Upon receiving the question data, the control unit 1 of the knowledge building support device SV vectorizes the question in step S23 under the control of the vectorization processing unit 15. For the vectorization process, a well-known vectorization method can be used, as in the vectorization process of the document text described above.

[0051] (2-2) Searching for candidate source sentences highly relevant to the question sentence Once the vectorized question sentence is obtained, the control unit 1 of the knowledge building support device SV, under the control of the candidate sentence search processing unit 16, searches for candidate vectorized source sentences highly relevant to the vectorized question sentence from each element set stored in the element set storage unit 32 based on the vectorized question sentence in step S24. More specifically, the control unit 1 calculates, for example, the similarity between the vectorized question sentence and each of the vectorized source sentences, and selects vectorized source sentences whose similarity is equal to or greater than a preset threshold as candidate source sentences highly relevant to the question sentence.

[0052] In addition, the candidate sentence search processing unit 16 searches for attribute information linked to the searched candidate source sentence, i.e., the identification information of the source document in which the candidate source sentence is written, together with the chapter identification number or title.

[0053] (2-3) Generation of Answer Sentence Subsequently, in step S25, under the control of the answer sentence generation processing unit 17, the control unit 1 of the knowledge construction support device SV inputs the vectorized question sentence and the searched vectorized material sentence candidates into the language generation model stored in the language generation model storage unit 33. Then, the language generation model generates an answer sentence corresponding to the vectorized question sentence and the vectorized material sentence candidates, and the generated answer sentence is acquired by the answer sentence generation processing unit 17.

[0054] As a language generation model, an existing general-purpose large-scale language model can be used, which is trained in advance on a large amount of text data using a machine learning model and can output various language processing tasks by simply providing a few example tasks. One example is GPT-4 (Generative Pre-Training-4).

[0055] (2-4) Generation and Output of Construction Support Information Finally, in step S26, under the control of the construction support information output processing unit 18, the control unit 1 of the knowledge construction support device SV generates construction support information including the generated answer sentence and attribute information linked to the candidate source sentence from which the answer sentence was generated, i.e., identification information of the source document in which the candidate source sentence is written, and the chapter identification number or title. The construction support information output processing unit 18 then transmits the generated construction support information from the communication I / F unit 4 to the requesting user terminals UT1 to UTk.

[0056] The control unit 1 of the knowledge building support device SV executes the series of support processes in steps S21 to S26 described above each time a query and question sentence data is received. Then, when a support end instruction is input in step S27, the process ends.

[0057] 7 shows an example of a response to a question. In this example, when a user inputs a question such as "What is the priority of railway remote control?", the knowledge construction support device SV generates an example of a response, and portions related to the question are shaded in grayscale, for example.

[0058] Therefore, in this case, the user determines whether the answer sentence is appropriate based on the answer sentence presented on the user terminals UT1 to UTk and the chapter number (Section No.) indicating its location, and if appropriate, adds the chapter of this document to the knowledge.

[0059] 8 shows another example of a response to a question. In this example, no sentence related to the question was found in the chapter of the source document, and the response in this case is "NO_OUTPUT."

[0060] Therefore, in this case, the user recognizes that the chapter of the source document in question does not contain any sentences related to the question, and does not add this chapter of the source document to the knowledge.

[0061] (Effects) As described above, in one embodiment, first, source document data in a field related to the anticipated question is selectively acquired from the source databases DB1 to DBn, the acquired source document is divided into chapters, paragraphs, and sentences, and vectorized. The divided paragraphs and sentences are then linked to chapters to generate and store a set of sentences that maintain the document structure. Then, when a user inputs a source sentence as an Intent in this state, the source sentence is vectorized, and the set of sentences is accessed based on the vectorized source sentence to search for candidate source sentences that are highly relevant to the source sentence. Then, the question sentence and the retrieved source sentence are input into a language generation model to generate an answer sentence corresponding to the question sentence, and the generated answer sentence is presented to the requesting user along with identification information for the source document and chapter containing the answer sentence.

[0062] Therefore, answers corresponding to questions are generated using an existing language generation model, which eliminates the need for a large amount of training data compared to creating a new machine learning model for knowledge extraction, thereby reducing development costs. Furthermore, by using a language generation model, it is possible to provide a system that is more versatile and accurate than when knowledge is constructed based on, for example, grammar rules or statistics.

[0063] [Other Embodiments] In one embodiment, a user inquires about "railway remote control priority" in a use case in the communications field. However, the present invention can be applied to any requirements defined for each field. For example, with respect to the above-mentioned railway remote control, in addition to priority, other requirements may be assumed, such as time synchronization, number of terminals, availability, data volume, area, delay, and communication frequency, and answer sentences may be generated corresponding to each of these requirements. Furthermore, the use case is not limited to the communications field, and may be in other fields such as information processing, medicine, materials, and construction and civil engineering.

[0064] In addition, the type and configuration of the knowledge construction support device, the configuration of each processing function provided by the knowledge construction support device, the processing procedures and processing contents when executing these processing functions, the type and configuration of the language generation model, the content of the question sentence, etc. can be modified and implemented in various ways without departing from the spirit of this invention.

[0065] Although the embodiments of the present invention have been described in detail above, the above description is merely an example of the present invention in every respect. It goes without saying that various improvements and modifications can be made without departing from the scope of the present invention. In other words, when implementing the present invention, specific configurations according to the embodiments may be appropriately adopted.

[0066] In short, this invention is not limited to the above-described embodiments, and in the implementation stage, the components can be modified and embodied without departing from the spirit of the invention. Furthermore, various inventions can be formed by appropriately combining multiple components disclosed in the above-described embodiments. For example, some components may be omitted from all the components shown in the embodiments. Furthermore, components from different embodiments may be appropriately combined.

[0067] SV...knowledge construction support device DB1 to DBn...document database UT1 to UTk...user terminal NW...network 1...control unit 2...program storage unit 3...data storage unit 4...communication I / F unit 5...bus 11...document data acquisition processing unit 12...document segmentation processing unit 13...vectorization processing unit for document sentences 14...question sentence data acquisition processing unit 15...vectorization processing unit for question sentences 16...candidate sentence search processing unit 17...answer sentence generation processing unit 18...construction support information output processing unit 31...document data storage unit 32...element set storage unit 33...language generation model storage unit

Claims

1. A knowledge construction support device for supporting the construction of knowledge used to quantify requirements included in a user's request content, comprising: a storage medium storing at least document materials in a field related to the use included in the user's request content; a first processing unit that divides the document materials into a plurality of components according to the hierarchy of the document structure, vectorizes each of the divided components, and generates a set of components associated with a higher-level hierarchy; a second processing unit that acquires a question sentence representing the use and the requirements included in the user's request content, and vectorizes the acquired question sentence; a third processing unit that, based on the vectorized question sentence, searches for candidate components having relevance to the question sentence from the vectorized set of components, and acquires attribute information representing the position of the candidate components in the set of components; a fourth processing unit that generates an answer sentence based on the question sentence and the candidate components using a pre-created language generation model; and a fifth processing unit that generates support information including the answer sentence and the attribute information, and presents the generated support information to the user. A knowledge construction support device comprising the above components.

2. The knowledge construction support device according to claim 1, wherein the first processing unit divides the document materials into a plurality of chapters, a plurality of paragraphs, and a plurality of sentences according to the hierarchy of the document structure, vectorizes the divided sentences, and associates them with the identification information of the chapters, thereby generating a set of sentences for each chapter.

3. The knowledge construction support device according to claim 2, wherein the third processing unit searches for candidate sentences having relevance to the question sentence from the set of sentences based on the vectorized question sentence, and acquires the attribute information including the identification information of the chapter associated with the candidate sentences.

4. A knowledge construction support method executed by an information processing apparatus for supporting the construction of knowledge used to quantify requirements included in a user's request content, the method comprising: a process of acquiring and storing in a storage medium at least a document of a field related to the use included in the user's request content; a process of dividing the document into a plurality of components according to a document structure hierarchy, vectorizing each of the divided components, and generating a set of components associated with a higher hierarchy; a process of acquiring a question sentence representing the use and the requirements included in the user's request content and vectorizing the acquired question sentence; a process of searching, based on the vectorized question sentence, for candidates for components having relevance to the question sentence from the set of vectorized components and acquiring attribute information representing the positions of the candidates for components in the set of components; a process of generating an answer sentence based on the question sentence and the candidates for components using a pre-created language generation model; and a process of generating support information including the answer sentence and the attribute information and presenting the generated support information to the user.

Images