Information processing systems, information processing methods, programs

Abstract

An information processing device, a control method and a program are provided that enable appropriate use of domain knowledge including tabular information in a system that answers questions by linking a generation AI and a search system. [Solution] In a question answering device system, the question answering device comprises a question acquisition means for acquiring a question, a conversion means for converting table information in a first format contained in search target information into table information in a second format, and a search means for searching for information related to the question acquired by the question acquisition means from the search target information in which the table information in the first format has been converted into table information in the second format by the conversion means, wherein the conversion means further converts the table information in the second format contained in the information searched by the search means into table information in the first format, and comprises an output means for outputting an answer to the question generated based on the information in which the table information in the second format has been converted into table information in the first format by the conversion means.

Description

Technical Field

[0001] The present invention relates to an information processing system, an information processing method, and a program.

Background Art

[0002] In enterprises and public institutions, the introduction of full-text search systems such as enterprise search to find necessary information from a vast amount of accumulated electronic documents is increasing. Also, by introducing a question-and-answer system that performs information search and answer through dialogue, it has become possible to easily reach the necessary information.

[0003] Furthermore, in recent years, with the emergence of generative AI based on large language models such as ChatGPT / GPT, more natural and highly accurate question-and-answer has become possible.

[0004] Also, generative AI has a high response ability for open-domain questions, but it cannot accurately answer domain-specific matters not included in the pre-training data, such as questions regarding specific documents within an enterprise.

[0005] Non-Patent Document 1 reports on a mechanism called RAG (Retrieval-Augmented Generation) that enables answers to unlearned matters by linking a generative AI and a search system. RAG uses a search system registered with domain knowledge to obtain domain knowledge strongly related to the question text, and includes the domain knowledge in the instruction (prompt) for the generative AI to ask questions, enabling the generative AI to answer unlearned matters as well.

[0006] Furthermore, Non-Patent Document 2 reports that by including a structured table in the prompt for a generative AI using a large language model, it is possible to answer questions regarding the table with high accuracy.

Prior Art Documents

Non-Patent Documents

[0007] [Non-Patent Document 1] Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, Meng Wang, and Haofen Wang, Retrieval-Augmented Generation for Large Language Models: A Survey, arXiv:2312.10997, 27 Mar 2024(https: / / arxiv.org / abs / 2312.10997) [Non-Patent Document 2] Wenhu Chen, Large Language Models are few(1)-shot Table Reasoners, arXiv:2210.06710, 23 Jan 2023(https: / / arxiv.org / abs / 2210.06710) [Disclosure of the Invention] [Problems that the invention aims to solve]

[0008] Documents registered as domain knowledge in RAG may contain structured tables. In this case, the document may include Markdown symbols or HTML tags to represent the structure.

[0009] Depending on how the search score is calculated, if extra symbols or tags are included, the search score may be lower compared to general search terms (search keywords or search queries), and it may not be possible to acquire domain knowledge suitable for generating answers.

[0010] On the other hand, as shown in Non-Patent Document 2, by including a structured table in the prompts given to the generating AI, it is possible to answer questions about tables with high accuracy. In other words, it is desirable for the tables given to the prompts to the generating AI to be structured.

[0011] Therefore, the present invention is Based on the table information, generate the appropriate answer. The purpose is to provide a system. [Means for solving the problem]

[0012] The present invention comprises: a question acquisition means for acquiring a question; a conversion means for converting a first-form table information contained in search target information into a second-form table information; and a search means for searching for information related to the question acquired by the question acquisition means from the search target information in which the first-form table information has been converted into the second-form table information by the conversion means, wherein the conversion means further converts the second-form table information contained in the information retrieved by the search means into a first-form table information, and the present invention comprises: an output means for outputting an answer to the question generated based on the information in which the second-form table information has been converted into the first-form table information by the conversion means. [Effects of the Invention]

[0013] According to the present invention, Based on the table information, generate the appropriate answer. This will make it possible to provide a system. [Brief explanation of the drawing]

[0014] [Figure 1] This figure shows an example of the system configuration of a question answering device in an embodiment of the present invention. [Figure 2] This is a block diagram showing an example of the hardware configuration of a question answering device and a user terminal in an embodiment of the present invention. [Figure 3] This figure shows an example of the functional configuration of a question answering device in an embodiment of the present invention. [Figure 4] This flowchart shows an example of a domain knowledge construction process in an embodiment of the present invention. [Figure 5] This figure shows an example of the configuration of a domain knowledge storage area in an embodiment of the present invention. [Figure 6] This flowchart shows an example of the process of separating the table structure from the text in an embodiment of the present invention. [Figure 7] It is a diagram showing an example of a document registered as a source of domain knowledge in an embodiment of the present invention. [Figure 8] It is a diagram showing an example of structured text extracted from a document registered as a source of domain knowledge in an embodiment of the present invention. [Figure 9] It is a diagram showing an example of a state in which structured text is separated into a text part and expression information in an embodiment of the present invention. [Figure 10] It is a diagram showing an example of a domain management table and an expression information management table in an embodiment of the present invention. [Figure 11] It is a flowchart showing an example of a process for generating a prompt including a table structure in an embodiment of the present invention. [Figure 12] It is a diagram showing an example of a prompt including a table structure in an embodiment of the present invention. [Figure 13] It is a diagram showing an example of a screen of a chatbot displayed on a user terminal. [Figure 14] It is a diagram showing a conceptual diagram of the process of the present invention.

Mode for Carrying Out the Invention

[0015] Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

[0016] FIG. 1 is a diagram showing an example of the system configuration of a question-and-answer device in an embodiment of the present invention.

[0017] The question-and-answer device 100 is configured to be connected via the user terminal 110 and the network 120.

[0018] The question-and-answer device 100 presents an answer sentence to the question sentence acquired from the user terminal 110.

[0019] The user terminal 110 sends the question entered by the user to the question answering device 100 and displays the answer returned by the question answering device 100. Specifically, the user terminal 110 can be a personal computer (such as a notebook PC or desktop PC), a tablet terminal, or a smartphone, but is not limited to these. Furthermore, the configuration of various terminals or devices connected on the network 120 in Figure 1 is just one example, and it goes without saying that there are various configuration examples depending on the application and purpose.

[0020] Figure 2 is a block diagram showing an example of the hardware configuration of an information processing device applicable as a server device 101 or client terminal 102 according to the present invention.

[0021] As shown in Figure 2, the information processing device is connected via a system bus 204 to a CPU (Central Processing Unit) 201, RAM (Random Access Memory) 202, ROM (Read Only Memory) 203, input controller 205, video controller 206, memory controller 207, and communication I / F controller 208.

[0022] CPU201 provides comprehensive control over all devices and controllers connected to the system bus 204.

[0023] RAM202 functions as the main memory, work area, etc., of the CPU201. The CPU201 loads the necessary programs, etc., from ROM203 or external memory 211 into RAM202, and then executes the loaded programs to perform various operations.

[0024] ROM203 or external memory211 holds the BIOS (Basic Input / Output System) and OS (Operating System), which are control programs executed by the CPU201, as well as computer-readable and executable programs and various necessary data (including data tables) for realizing this information processing method.

[0025] The input controller 205 controls input from input devices such as a keyboard 209 or a pointing device such as a mouse (not shown). If the input device is a touch panel, the user can give various instructions by pressing (touching with a finger, etc.) icons, cursors, or buttons displayed on the touch panel.

[0026] Furthermore, the touch panel may be a multi-touch screen or other type of touch panel capable of detecting the positions of multiple fingers touching it.

[0027] The video controller 206 controls the display to an external output device such as the display 210. The display includes the display of a notebook computer integrated with the main unit. The external output device is not limited to a display; for example, it may be a projector. Furthermore, for the aforementioned touch-enabled device, an input device is also provided.

[0028] The video controller 206 can control the video memory (VRAM) used for display control. It can use a portion of RAM 202 as the video memory area, or it can provide a separate, dedicated video memory.

[0029] The memory controller 207 controls access to the external memory 211. The external memory can include an external storage device (hard disk), a flexible disk (FD), or a CompactFlash® memory connected to a PCMCIA card slot via an adapter, which stores boot programs, various applications, font data, user files, editing files, and other data.

[0030] The communication interface controller 208 connects to and communicates with external devices via a network and performs communication control processing over the network. For example, it can communicate using TCP / IP, telephone lines such as ISDN, and 3G mobile phone lines.

[0031] Furthermore, the CPU 201 enables display on the display 210 by, for example, performing the process of expanding (rasterizing) outline fonts into the display information area in RAM 202. The CPU 201 also enables user input via a mouse cursor (not shown) on the display 210.

[0032] Next, Figure 3 shows an example of the functional configuration of the question answering device 100 and the user terminal 110. The functions of each functional unit will be explained in the flowchart shown in Figure 4 and other references.

[0033] (Overview) Figure 14 illustrates the outline of the present invention. In this embodiment, the present invention will be explained assuming a scenario in which, for example, an employee on an employee portal site asks a chatbot a question about information (documents) on the site and obtains an answer. In this embodiment, an example using a large language model (LLM) as the answer generation processing unit 307 will be described, but the invention is not limited to this, and generative AI, machine learning models, and other methods capable of generating answers may be used.

[0034] Documents registered as domain knowledge may include structured tables ("structured tables" is an example of the first form of table information in this invention). However, when searching, the inclusion of extraneous symbols and tags may result in a lower search score for common search terms, making unstructured tables (for example, a table in a text-based format represented by characters, spaces, and line breaks, as shown in 901 of Figure 9; "unstructured tables" is an example of the second form of table information in this invention) preferable. On the other hand, when providing information to an LLM, it is preferable for tables, etc., to be in a structured state (for example, HTML format as shown in 803 and 804 of Figure 8). Therefore, it is desirable to use text-based documents when searching and structured documents when generating answers with an LLM, and by using both appropriately, highly accurate searches and answer generation become possible.

[0035] Therefore, the present invention performs the following processing: (1) Extract structured tables contained in the document, which is domain knowledge, section by section and convert them into text. (2) Manage the original structured tables in association with the tables converted into text. In other words, for each section, the text tables and structured tables are associated (details will be described later in Figures 4 and 6). Then, (3) search the domain knowledge storage area for information related to the question entered by the user. Since the domain knowledge is searched in text form, it is possible to obtain domain knowledge suitable for the user's question. (4) Convert the text tables contained in the retrieved domain knowledge into structured tables corresponding to those tables (details will be described later in Figure 11). (5) Provide the LLM with a prompt that includes the question and the domain knowledge converted into structured tables, and (6) obtain the answer to the user's question. By providing the LLM with structured tables, it becomes possible to generate answers with high accuracy.

[0036] As mentioned above, even for domain knowledge that includes tabular information, managing both text-based tables for searching and structured tables for generating answers, and using them appropriately depending on the situation, enables highly accurate searching and question answering.

[0037] In this invention, domain knowledge refers to the information that forms the basis for answering user questions. For example, when applying this invention as a search system for a company's employee website (including a question-answering system in a conversational (chat) format), the information posted on the employee website (such as web page information and posted document files) constitutes the domain knowledge. Domain knowledge is registered and managed by methods such as registering site information obtained by searching URLs using prefix matching and updating it at regular intervals, or by pre-registering data by an administrator. Domain knowledge is an example of searchable information.

[0038] (First Embodiment) Next, the process executed by the domain knowledge construction processing unit 301 in the first embodiment of the present invention will be described using the flowchart in Figure 4.

[0039] (Domain knowledge building process) The flowchart in Figure 4 shows the process by which the CPU 201 of the question answering device 100 reads and executes a predetermined control program, and the process by which the domain knowledge construction processing unit 301 constructs the information necessary for performing question answering from a set of documents (domain knowledge) given in advance as knowledge in the domain knowledge storage area 302. In this embodiment, a structured table is referred to as first-form table information, and an unstructured table is referred to as second-form table information. In addition, for example, if the domain knowledge includes images, a table in image state may be extracted as the first form, and a table converted to text by OCR (Optical Character Recognition) or the like may be referred to as the second form. In that case, the second-form text table may be used when searching for domain knowledge, and it may be converted to the first-form image when having the LLM generate an answer.

[0040] In step S401, the domain knowledge building processing unit 301 obtains a list of documents that have been given as knowledge in advance (for example, information within the employee portal site).

[0041] In step S402, the domain knowledge construction processing unit 301 initializes the domain knowledge storage area 302.

[0042] In step S403, the domain knowledge building processing unit 301 starts the iterative processing up to step S410 for each document in the document list.

[0043] In step S404, the domain knowledge construction processing unit 301 extracts structured text (for example, information organized according to a certain format or style, such as HTML format as shown in 803 and 804 in Figure 8) from the document to be processed.

[0044] In embodiments of the present invention, the data format of the document is not limited. Furthermore, the representation format of the extracted structured text is not particularly limited as long as the table structure can be identified, and it does not have to be in HTML format.

[0045] In embodiments of the present invention, structured documents are represented in HTML format to facilitate understanding.

[0046] In step S405, the domain knowledge construction processing unit 301 divides the structured text extracted in step S404 into sections.

[0047] While embodiments of the present invention describe a method of dividing into sections, the entire document may be treated as a single section, or it may be configured to be divided by page or character count. Furthermore, if the generation AI used for answer generation has a character limit for prompts, it may be configured to divide the document according to that limit.

[0048] In step S406, the domain knowledge building processing unit 301 starts the iterative processing up to step S409 for each section.

[0049] In step S407, the domain knowledge construction processing unit 301 uses the table information processing unit 303 to obtain the table structure and the text-formatted section from the section. The process of separating the table structure and text will be described later.

[0050] In step S408, the domain knowledge construction processing unit 301 registers the second-format text section obtained in step S407 as domain knowledge in the domain knowledge management table 501, and saves the first-format table structure in the table information management table 502 in association with the domain knowledge.

[0051] In step S409, the domain knowledge construction processing unit 301 repeats the process from step S406 if there are still sections to be processed. If there are no sections to be processed, the process moves to step S410.

[0052] In step S410, the domain knowledge building processing unit 301 repeats the process from step S403 if there are still documents to be processed. If there are no documents to be processed, the process ends.

[0053] (Separation process of table structure and main text) Next, using the flowchart in Figure 6, we will explain the process of separating the table structure and the main text from the section in step S407.

[0054] In step S601, the table information processing unit 303 initializes the temporary area.

[0055] In step S602, the table information processing unit 303 extracts a structured table of the first form from the section to be processed. The method of extracting the table is not particularly limited, but for example, a tag that indicates the start of the table (for example) (etc.) and tags that indicate the end (for example) You can register tags (such as those mentioned above) and extract data by detecting the registered tags (a technique similar to web scraping). You can also use other techniques that identify and extract data based on the characteristics of the table.

[0056] In step S603, the table information processing unit 303 starts the iterative processing up to step S607 for each structured table.

[0057] In step S604, the table information processing unit 303 extracts only the text from the structured table to be processed. Table delimiters (for example, tags such as , , etc. in HTML format) are replaced with spaces or line breaks to obtain the table information in a second format.

[0058] In step S605, the table information processing unit 303 converts the portion of the section occupied by the structured table to be processed into the text extracted in step S604. This results in the entire document being in text format. The location of the table may be identified using techniques such as those used in step S602.

[0059] In step S606, the table information processing unit 303 stores the start and end positions of the table in the section of the table portion converted in step S605 in a temporary area, associating them with the structured table being processed. The start and end positions of the table refer to the information of which character the text portion 901 of the table starts and ends in section 802.

[0060] In step S607, the table information processing unit 303 repeats the process from step S603 if there are still structured tables to be processed. If there are no tables to be processed, the process moves to step S608.

[0061] In step S608, the table information processing unit 303 returns the text-converted section and the table information stored in the temporary area to the caller.

[0062] (Specific example of domain knowledge building process) Next, as a concrete example of the flowchart in Figure 4, we will explain using Figures 7 to 10 the case where a domain knowledge building process is performed on the document shown in Figure 7.

[0063] In step S401, the domain knowledge construction processing unit 301 obtains a list of documents related to document 701.

[0064] In step S402, the domain knowledge construction processing unit 301 clears the domain knowledge management table 501 and the table information management table 502 in the domain knowledge storage area 302.

[0065] In step S403, the domain knowledge construction processing unit 301 starts the iterative processing up to step S410 for document 701 in the document list.

[0066] In step S404, the domain knowledge construction processing unit 301 extracts the structured text 801 shown in Figure 8 from the document 701 to be processed.

[0067] In step S405, the domain knowledge construction processing unit 301 divides the structured text 801 extracted in step S404 into sections.

[0068] In step S406, the domain knowledge construction processing unit 301 starts iterative processing on section 802 up to step S409.

[0069] In step S407, the domain knowledge construction processing unit 301 uses the table information processing unit 303 to call the process of separating the table structure and text for section 802.

[0070] In step S601, the table information processing unit 303 initializes the temporary area 900 shown in Figure 9.

[0071] In step S602, the table information processing unit 303 extracts structured tables 803 and 804 from the section 802 to be processed.

[0072] In step S603, the table information processing unit 303 starts iterative processing on the structured table 803 up to step S607.

[0073] In step S604, the table information processing unit 303 extracts the text portion 901 shown in Figure 9 from the structured table 803 that is to be processed.

[0074] In step S605, the table information processing unit 303 converts the portion of section 802 occupied by the structured table 803 to be processed into the text 901 extracted in step S604 (Figure 9).

[0075] In step S606, the table information processing unit 303 stores the start position 902 (33rd character) and end position 903 (94th character) of section 802 of the text portion 901 of the table converted in step S605 as table information 904 in the temporary area 900, associating them with the structured table 803 to be processed.

[0076] In step S607, the table information processing unit 303 performs the repetitive processing from step S603 because there are still structured tables 804 to be processed.

[0077] In step S603, the table information processing unit 303 starts iterative processing up to step S607 on the structured table 804.

[0078] In step S604, the table information processing unit 303 extracts the text portion 905 from the structured table 804 that is to be processed.

[0079] In step S605, the table information processing unit 303 converts the portion of section 802 occupied by the structured table 804 to be processed into the text 905 extracted in step S604.

[0080] In step S606, the table information processing unit 303 stores the start position 906 (97th character) and end position 907 (129th character) of the text portion 905 of the table converted in step S605 in section 802 as table information 908 in the temporary area 900, associating them with the structured table 804 to be processed.

[0081] In step S607, the table information processing unit 303 determines that there are no more structured tables to process and proceeds to step S608.

[0082] In step S608, the table information processing unit 303 returns the text-formatted section 802 and the table information stored in the temporary area 904 and 908 to the domain knowledge construction processing unit 301.

[0083] In step S408, the domain knowledge construction processing unit 301 registers the text-formatted section 802 obtained in step S407 as domain knowledge 1001 (id=1) in the domain knowledge management table 501, and saves table structures 904 and 908 in the table information management table 502, associating them with domain knowledge 1001 (domain knowledge id=1) (Figure 10).

[0084] In step S409, the domain knowledge construction processing unit 301 moves to step S406 because there is still section 805 to be processed.

[0085] The same process is repeated below to build domain knowledge.

[0086] (Prompt generation process) Next, we will explain the process of generating prompts using the constructed domain knowledge, using the flowchart in Figure 11.

[0087] The flowchart in Figure 11 shows the process by which the CPU 201 of the question answering device 100 reads and executes a predetermined control program, and the process by which the question answering processing unit 305 generates a prompt for the answer generation processing unit 307 to answer the question in response to the question received by the dialogue management unit 304.

[0088] In the embodiments of the present invention, an example using an LLM as the answer generation processing unit 307 has been described, but the invention is not particularly limited, and any generative AI, machine learning model, or other method capable of generating answers may be used. A general-purpose large-scale language model that has been pre-trained unsupervised using a large amount of text data to learn grammar, word meanings, etc., may be used. Furthermore, a fine-tuned large-scale language model that has been supervised learning specific domain knowledge may also be used. In addition, in this embodiment, the answer generation processing unit 307 is configured to be inside the question answer processing device 100, but these may be located externally, for example, by obtaining answers via an API.

[0089] In step S1101, the question answering processing unit 305 initializes the prompt.

[0090] In step S1102, the question answering processing unit 305 searches the domain knowledge management table 501 for the question text obtained from the dialogue management unit 304 and retrieves a predetermined number of sections that are strongly related to the question text. The search method is not particularly limited, but strongly related sections may be extracted using a full-text search based on the question text or a search query generated from the question text. A search query is a sentence, phrase, or combination thereof used to search for information suitable for answering the question text. The method for creating a search query from the question text is not particularly limited, but it may be determined by processing such as extracting only independent words from the results of morphological analysis of the question text.

[0091] In the embodiments of the present invention, a predetermined number of items are acquired, but the system may be configured to acquire items according to limitations such as the number of characters in the prompt. Alternatively, the relationship between the question and domain knowledge may be evaluated using a score or the like, and only the top-ranking domain knowledge items may be acquired.

[0092] In step S1103, the question answering processing unit 305 starts the iterative processing up to step S1109 for each of the domain knowledge acquired in step S1102.

[0093] In step S1104, the question answering processing unit 305 retrieves table information associated with the domain knowledge to be processed from the table information table 502.

[0094] In step S1105, the question answering processing unit 305 starts the iterative processing up to step S1107 for each of the table information obtained in step S1104.

[0095] In step S1106, the question answering processing unit 305 converts the relevant portion (second-form table information) in the domain knowledge text from the start position to the end position of the table information to be processed into linked structured first-form table information. In this embodiment, the conversion process includes a replacement process that replaces the table. That is, it may be a process that replaces it with linked structured first-form table information, or it may be a process that extracts text from the second-form table information and reassembles it into a first-form table. Furthermore, although an embodiment has been described in which the region specified by the start position to the end position of the second-form table information in a section is converted into first-form table information when converting second-form table information to first-form table information, it is sufficient if the region to be converted can be identified by some method without using the start position or end position.

[0096] In step S1107, the question answering processing unit 305 repeats the process from step S1105 if there is still table information to be processed. If there is no table to be processed, the process moves to step S1108.

[0097] In step S1108, the question answering processing unit 305 adds the domain knowledge to be processed to the prompt.

[0098] In step S1109, the question answering processing unit 305 repeats the process from step S1103 if it still has domain knowledge to process. If there is no table to process, it moves to step S1110.

[0099] In step S1110, the question answering processing unit 305 adds the question to the prompt. At this time, if necessary, it may also add a string that prompts the LLM to output.

[0100] Figure 12 shows an example of a prompt including a table structure. 1202 is an empty prompt. 1203 is an instruction to cause the LLM to create an answer, added in step S1101. 802 is the structured body text of the table, added in step S1108. 803 and 804 are structured tables, which are converted into structured tables in step S1106. 1201 is the question entered by the user, which the question answering processing unit 305 obtains from the dialogue management unit 304 and adds to the prompt in step S1110.

[0101] (Specific example of prompt generation process) Next, as a concrete example of the prompt generation process shown in the flowchart of Figure 11, we will explain the case where question 1201 ("How much is the daily allowance for a manager who travels to Osaka on a business trip?") is entered into the dialogue management unit 304, using Figure 12 as well.

[0102] In step S1101, the question answering processing unit 305 initializes the empty prompt 1202 shown in Figure 12 by adding an instruction 1203 to the LLM.

[0103] In step S1102, the question answering processing unit 305 retrieved the question sentence 1201 from the dialogue management unit 304, searched the domain knowledge management table 501, and retrieved only the domain knowledge 1001 as a section strongly related to the question sentence 1201.

[0104] In step S1103, the question answering processing unit 305 starts iterative processing up to step S1109 with respect to the domain knowledge 1001 acquired in step S1102.

[0105] In step S1104, the question answering processing unit 305 retrieves table information 904 and table information 908 associated with the domain knowledge 1001 (domain id=1) to be processed from the table information table 502.

[0106] In step S1105, the question answering processing unit 305 starts the iterative processing up to step S1107 with respect to the table knowledge 904 acquired in step S1104.

[0107] In step S1106, the question answering processing unit 305 converts the relevant portion 901 in the domain knowledge text 802, from the 23rd character at the start position to the 94th character at the end position of the table information 904, into the structured table 803 of the table information 904.

[0108] In step S1107, the question answering processing unit 305 performs repeated processing from step S1105 because there is still table information 908 to be processed.

[0109] In step S1105, the question answering processing unit 305 starts the iterative processing up to step S1107 with respect to the table knowledge 908 acquired in step S1104.

[0110] In step S1106, the question answering processing unit 305 converts the relevant portion 905 in the domain knowledge text 802 from the 97th character at the start position to the 129th character at the end position of the table information 908 into the structured table 804 of the table information 908.

[0111] In step S1107, the question answering processing unit 305 determines that there is no more table information to process and proceeds to step S1108.

[0112] In step S1108, the question answering processing unit 305 adds the title of the domain knowledge 1001 to be processed and the table-structured body text 802 to the prompt.

[0113] In step S1109, the question answering processing unit 305 does not have domain knowledge to process, so it moves the process to step S1110.

[0114] In step S1110, the question answering processing unit 305 adds the question sentence 1201 and the output prompt string 1204 to the prompt.

[0115] The question answering processing unit 303 obtains the correct answer by instructing the answer generation processing unit 307 using the prompt 1202.

[0116] In the embodiments of the present invention, for the sake of simplicity, the correction of the shift in the conversion position of the table has been omitted. The table may be configured to be converted from the end of the text to prevent the conversion position from shifting.

[0117] Figure 13 shows an example of a screen that displays an answer to the user. The user's question, "What is the daily allowance for a manager traveling to Osaka?", is entered in question form 1301. The generated answer is displayed in answer form 1302. In answer form 1302 in Figure 13, only the answer is displayed, but if the table information itself can serve as the answer, the table itself may be output as the answer. Furthermore, along with the answer, the domain knowledge obtained in S1102, which served as the basis for generating the answer, is displayed as background information 1303. If table information is referenced as background information 1303, the referenced table may be displayed as is. In Figure 13, background information 1303 is a link, allowing the content of the background information to be displayed by clicking or other instructions. The link may display the contents of the domain knowledge management table 501, or it may be configured to reference external information. In addition, there may be a function that allows the user to evaluate the answer generated for the question. Additional question form 1304 allows the user to input questions they wish to ask on an ongoing basis.

[0118] In the embodiment of the present invention, the text information is stored in the domain knowledge storage area and linked to the structured table (i.e., only the table portion of the structured information and the text information overlap), which also makes it possible to optimize the storage area. If there is no limit to the capacity of the domain storage area, the text document and the structured document may be associated and stored in a way that completely overlaps, and when searching, the text document is used and the corresponding structured document is entered into the LLM.

[0119] As mentioned above, by separately managing the searchable text and table structure of knowledge in RAG, domain knowledge registered in tabular format can also be searched appropriately, and by providing LLMs with structured tables as prompts, it becomes possible to generate more accurate answers.

[0120] Although embodiments have been described above, the present invention can take the form of, for example, a system, apparatus, method, program, or recording medium. Specifically, it may be applied to a system consisting of multiple devices, or to an apparatus consisting of a single device.

[0121] Furthermore, the program in this invention is a program that allows a computer to execute the processing method shown in the flowchart in Figure 3, and the storage medium of this invention stores a program that allows a computer to execute the processing method shown in Figure 3. The program in this invention may also be a program for each processing method of each device shown in Figure 3.

[0122] As described above, it goes without saying that the object of the present invention can also be achieved by supplying a recording medium containing a program that realizes the functions of the embodiments described above to a system or device, and by having the computer (or CPU or MPU) of that system or device read and execute the program stored on the recording medium.

[0123] In this case, the program read from the recording medium itself realizes the novel function of the present invention, and the recording medium on which that program is recorded constitutes the present invention.

[0124] For recording media used to supply programs, examples include flexible disks, hard disks, optical disks, magneto-optical disks, CD-ROMs, CD-Rs, DVD-ROMs, magnetic tapes, non-volatile memory cards, ROMs, EEPROMs, silicon disks, and the like.

[0125] Furthermore, it goes without saying that the functions of the aforementioned embodiments are realized not only by the computer executing the program it has read, but also by the operating system (OS) running on the computer performing some or all of the actual processing based on the instructions of that program, thereby realizing the functions of the aforementioned embodiments.

[0126] Furthermore, it goes without saying that this also includes cases where, after a program read from a recording medium is written to the memory of a function expansion board inserted into a computer or a function expansion unit connected to a computer, the CPU or other components of the function expansion board or function expansion unit perform some or all of the actual processing based on the instructions of the program code, and the functions of the aforementioned embodiments are realized through that processing.

[0127] Furthermore, the present invention may be applied to a system consisting of multiple devices or to a device consisting of a single device. It goes without saying that the present invention can also be applied when the results are achieved by supplying a program to a system or device. In this case, by reading a recording medium containing a program for achieving the present invention into the system or device, the system or device can enjoy the effects of the present invention.

[0128] Furthermore, by downloading and reading the program for achieving the present invention from a server, database, etc. on a network using a communication program, the system or device can enjoy the effects of the present invention. It should be noted that configurations combining the above-described embodiments and their variations are all included in the present invention. [Explanation of Symbols]

[0129] 100 Question Answering Device 110 User terminals 120 Networks

Claims

1. A means of obtaining a question, A conversion means that converts structured table information contained in search target information, which is document data, into unstructured table information by replacing the information used for structuring contained in the document data, A search means for searching information related to a question obtained by the question acquisition means from search target information that has been converted into unstructured tabular information by the conversion means, Equipped with, The conversion means further features converting unstructured table information contained in the information retrieved by the search means into structured table information. Output means for outputting instructions to the generating AI to output an answer to the question based on the information converted into structured table information by the conversion means, An information processing device characterized by comprising:

2. The output means outputs the question acquired by the question acquisition means and the information retrieved by the search means, thereby causing the generating AI to generate an answer. The information processing apparatus according to claim 1, characterized by the following:

3. The output means outputs a response generated by outputting to the generating AI a prompt that includes the information obtained by the conversion means, which has been converted from unstructured table information to structured table information, and the question obtained by the question acquisition means. The information processing apparatus according to claim 1, characterized by the following:

4. The system includes an extraction means for extracting structured table information from the search target information based on a specific string contained in the search target information, The conversion means converts the structured table information extracted by the extraction means into the unstructured table information. The information processing apparatus according to claim 1, characterized by the following:

5. Equipped with a storage means for storing information to be searched, The storage means stores structured table information contained in the search target information in association with unstructured table information obtained by converting the structured table information. The information processing apparatus according to claim 1, characterized by the following:

6. The storage means is characterized by storing information indicating the location of unstructured table information in the information to be retrieved, and the structured table information corresponding to said unstructured table information, in association with each other. The conversion means converts unstructured table information identified by location information into structured table information stored in association with the unstructured table information. The information processing apparatus according to claim 5, characterized by the following:

7. The process by which the conversion means converts unstructured table information into structured table information is a process of replacing the unstructured table information with structured table information stored in association with the unstructured table information. The information processing apparatus according to claim 6, characterized by the following:

8. The search means searches for the target information using words extracted from the questions obtained by the question acquisition means. The information processing apparatus according to claim 1, characterized by the following:

9. The output means outputs information retrieved by the search means, including the table used to generate the answer, in an identifiable manner. The information processing apparatus according to claim 1, characterized by the following:

10. The information subject to search by the aforementioned search means is information identified by a URL that satisfies predetermined conditions. The information processing apparatus according to claim 1, characterized by the following:

11. The question acquisition means of the information processing device includes a question acquisition step for acquiring a question, The conversion means of the information processing device converts structured table information contained in search target information, which is document data, into unstructured table information by replacing the information used for structuring contained in the document data. The search means of the information processing device searches for information related to the question obtained in the question acquisition step from the search target information which has been converted into unstructured tabular information in the conversion step, The conversion means of the information processing device includes a second conversion step in which unstructured table information contained in the information retrieved by the search step is converted into structured table information, The output means of the information processing device outputs an instruction to the generating AI to output an answer to the question based on the information that the generating AI has converted into structured tabular information by the second conversion step. A control method for an information processing device, characterized by comprising the following:

12. A program for causing at least one computer to function as one of the means of an information processing apparatus described in any one of claims 1 to 10.

Images