Basis acquisition device and basis acquisition method
The grounds acquisition device and method address the challenge of formatting grounds for answers by using a large language model to adaptively extract and present evidence in a coherent manner, improving comprehension.
Patent Information
- Application Number
- PCT/JP2025/017177
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2025-05-12
- Publication Date
- 2026-11-19
AI Technical Summary
Existing information processing devices struggle to uniformly output grounds for answers in a format that is easily understandable, especially when the reference information is in different expression formats, leading to difficulty in comprehending the basis of the answers.
A grounds acquisition device and method that utilizes a large language model to extract and output grounds descriptions based on the expression format of the reference information, ensuring clarity by adapting the extraction unit to match the format of the evidence or grounds.
Enhances the understandability of the grounds for answers by tailoring the extraction process to the specific format of the reference information, preventing partial or fragmented presentation of evidence.
Smart Images

Figure JP2025017177_19112026_PF_FP_ABST
Abstract
Description
Grounds Acquisition Device and Grounds Acquisition Method
[0001] The present disclosure relates to a technology for acquiring grounds for an answer.
[0002] Patent Document 1 discloses an information processing device that generates an answer to a question by using reference information indicating a plurality of passages. This information processing device further outputs the grounds for the answer. This information processing device selects the grounds for the answer from the plurality of passages. For this reason, this information processing device uniformly outputs the grounds for the answer in units of passages.
[0003] International Publication No. 2024 / 261982
[0004] Information indicating descriptions in various expression formats may be used as reference information indicating descriptions that serve as references for answers. For example, information indicating a description expressed in a table expression format may be used as reference information that serves as a reference for an answer. Furthermore, information indicating a description expressed in an expression format of a set of a question and an answer may be used as reference information that serves as a reference for an answer.
[0005] Like the information processing device described in Patent Document 1, if the grounds for an answer are uniformly output in units of passages regardless of the expression format of the description shown in the reference information that serves as a reference for the answer, the grounds for the answer may become difficult to understand. For example, when only a part of a table is output as the grounds for an answer, the grounds for the answer may become difficult to understand.
[0006] An objective of the present disclosure is to provide a technology capable of suppressing the grounds for an answer from becoming difficult to understand.
[0007] A grounds acquisition device according to one aspect of the present disclosure includes: a control unit that causes a large language model to extract a grounds description including the grounds from the description, in an extraction unit based on the expression format of a grounds part including the grounds for the answer among the descriptions shown in the related information, based on information related to an answer generated based on information related to a question and related information related to the question, and the related information; an acquisition unit that acquires the grounds description extracted by the large language model; and an output unit that causes a terminal to display the grounds description by outputting the grounds description to the terminal.
[0008] A method for obtaining evidence relating to another aspect of this disclosure includes: having a large-scale language model extract from a description an evidence description containing the evidence, based on information relating to a question and related information relating to the question, and the related information, in an extraction unit based on the expression format of the evidence portion of the description shown in the related information that contains the evidence for the answer; obtaining the evidence description extracted by the large-scale language model; and displaying the evidence description on a terminal by outputting the evidence description to the terminal.
[0009] This disclosure provides a technology that can mitigate the difficulty in understanding the basis of the answer.
[0010] This figure shows an example of the response system RS. This figure shows an overview of the process for extracting the basis description D1 from the specialized information B1. This figure shows an example of the relationship between the expression format of the basis part B3 and the extraction unit. This figure shows another example of the relationship between the expression format of the basis part B3 and the extraction unit. This figure shows an example of the terminal 10. This figure shows an example of the server 30. This figure shows an example of the system prompt SP1. This figure shows an example of the system prompt SP2. This figure shows an example of question A1. This figure shows an example of the specialized information B1. This figure shows an example of prompt P1. This figure shows an example of answer C1. This figure shows an example of prompt P2. This figure shows an example of the basis E1 included in the specialized information B1. This figure shows an example of the basis description D1. This figure shows an example of the basis description D1. This figure shows an example of the basis description D1. This figure shows an example of the basis description D1. This is a flowchart to explain an example of the operation of the system prompt setting unit 331. This is a flowchart to explain an example of the operation of the server 30. This figure shows the control unit 5. This figure shows an example of the system prompt SP3. This figure shows an example of the system prompt SP4. This figure shows an example of the system prompt SP5. This figure shows an example of prompt P3. This figure shows an example of the basis E1. This figure shows an example of prompt P4. This figure shows an example of extraction unit F1. This figure shows an example of prompt P5. This figure shows an example of justification description D1.
[0011] 1. Embodiments 1-1. Response System RS Figure 1 is a diagram showing an example of the response system RS. The response system RS automatically responds to question A1 from user U1. Question A1 is an example of information related to a question. Note that the information related to a question is not limited to question A1. For example, the information related to a question may be a comment that includes question A1. A comment that includes question A1 may be a comment that includes question A1 and the date on which the question was asked. A comment that includes question A1 may also be a comment that includes question A1 and the deadline for answering question A1. The response system RS includes a terminal 10, an LLM (large language model) system 20, and a server 30.
[0012] Terminal 10 is, for example, a smartphone. However, terminal 10 is not limited to a smartphone. For example, terminal 10 may be a tablet or a personal computer. Terminal 10 can communicate with server 30 via a communication network NW. Terminal 10 may also communicate with server 30 without using a communication network NW. Terminal 10 is used by user U1. Terminal 10 receives question A1 from user U1.
[0013] The LLM system 20 is composed of multiple computers. The LLM system 20 may be composed of a single computer. The LLM system 20 functions as LLM 21. LLM 21 is, for example, a generative AI (Artificial Intelligence). LLM 21 generates information such as text in response to prompts. LLM 21 can communicate with server 30 via a communication network NW. LLM 21 may also be able to communicate with server 30 without using a communication network NW. LLM 21 is used by server 30, for example. LLM 21 may be included in server 30. LLM 21 may also be able to communicate with terminal 10.
[0014] Server 30 is composed of multiple computers. Server 30 may be composed of a single computer. Server 30 can communicate with terminal 10 via a communication network NW. Server 30 may also communicate with terminal 10 without using a communication network NW. Server 30 can communicate with LLM 21 via a communication network NW. Server 30 may also communicate with LLM 21 without using a communication network NW.
[0015] Server 30 stores specialized information B1. Server 30 may also obtain specialized information B1 from other devices that store specialized information B1. Specialized information B1 is information that may not have been learned by LLM 21. For example, specialized information B1 is information about the latest knowledge. Specialized information B1 may also include information about the knowledge held by a company. The knowledge shown in specialized information B1 is an example of the description shown in specialized information B1. Specialized information B1 is information related to question A1. Specialized information B1 is an example of related information related to question A1. Specialized information B1 is used, for example, in In-Context Learning (ICL).
[0016] Server 30 receives question A1 from terminal 10. Server 30 obtains answer C1 to question A1 from LLM 21 using question A1 and specialized information B1. Answer C1 is an example of information about the answer generated based on the question information and related information. Note that the information about the answer is not limited to answer C1. For example, the information about the answer may be a comment that includes answer C1. A comment that includes answer C1 may be a comment that includes answer C1 and the date on which the answer was given. A comment that includes answer C1 may also be a comment that includes answer C1 and the name of the LLM 21 that gave the answer. Server 30 displays answer C1 on terminal 10 by sending answer C1 to terminal 10. That is, Server 30 presents answer C1 to user U1 using terminal 10.
[0017] Based on the answer C1 and the expert information B1, the server 30 causes the LLM 21 to extract the supporting description D1 from the expert information B1.
[0018] Figure 2 is a schematic diagram illustrating the process of extracting the supporting description D1 from the specialized information B1. The description B2 shown in the specialized information B1 has a supporting portion B3 that includes the supporting evidence E1 for the answer C1.
[0019] For example, if evidence E1 is included in "Table 1" within description B2, then the evidence section B3 containing evidence E1 is "Table 1". If the evidence section B3 is "Table 1", then the format of the evidence section B3 is "table".
[0020] If evidence E1 is included in the "first sentence" within description B2, then the evidence section B3 containing evidence E1 is the "first sentence." If the evidence section B3 is the "first sentence," then the expression format of the evidence section B3 containing evidence E1 is a "sentence."
[0021] If evidence E1 is included in the "question and answer set QA1" within description B2, then the evidence section B3 containing evidence E1 is the "question and answer set QA1". If the evidence section B3 is the "question and answer set QA1", then the expression format of the evidence section B3 is the "question and answer set". The "question and answer set" is also called a QA (Question and Answer) set.
[0022] If evidence E1 is included in both the "second sentence" and the "first diagram" within description B2, then the evidence section B3 containing evidence E1 is a "combination of the second sentence and the first diagram." Note that "diagram" means a figure or a table. If the evidence section B3 is a "combination of the second sentence and the first diagram," then the expression format of the evidence section B3 is a "combination of sentence and diagram."
[0023] Server 30 causes LLM 21 to extract the basis description D1, which includes basis E1, from description B2, based on the expression format of basis part B3, which includes basis E1, as an extraction unit.
[0024] Figure 3 shows an example of the relationship between the expression format of evidence section B3 and the extraction unit. Figure 4 shows another example of the relationship between the expression format of evidence section B3 and the extraction unit. In Figures 3 and 4, when the expression format of evidence section B3 is "table", the extraction unit based on the expression format of evidence section B3 is "table". Therefore, when the expression format of evidence section B3 is "table", evidence description D1 is a table containing evidence E1. When the expression format of evidence section B3 is "table", evidence description D1 is consistent with evidence section B3.
[0025] In Figure 3, if the expression format of evidence portion B3 is "text," the extraction unit based on the expression format of evidence portion B3 is "paragraph." Therefore, if the expression format of evidence portion B3 is "text," then the evidence description D1 is a paragraph. If evidence portion B3 is a sentence within a paragraph composed of multiple sentences, then the evidence description D1, which is a paragraph of text, does not coincide with evidence portion B3, which is a sentence within a paragraph. Note that if the expression format of evidence portion B3 is "text," the extraction unit based on the expression format of evidence portion B3 may also be "single sentence," as shown in Figure 4.
[0026] The "table" format of the basis section B3 is an example of the first format. The "text" format of the basis section B3 is an example of the second format, which is different from the first format. The extraction unit (table) when the format of the basis section B3 is a "table" is different from the extraction unit (paragraph or sentence) when the format of the basis section B3 is a "text". Thus, the extraction unit when the format of the basis section B3 is the first format is different from the extraction unit when the format of the basis section B3 is the second format. In other words, the extraction unit based on the first format is different from the extraction unit based on the second format.
[0027] The first format is not limited to a "table." For example, the first format could be "text," "a set of questions and answers," or "a combination of text and a chart." The second format can be any format different from the first format, and is not limited to "text."
[0028] The following describes examples of terminal 10 and server 30 included in the response system RS.
[0029] 1-2. Terminal 10 Figure 5 shows an example of terminal 10. Terminal 10 includes an input device 11, a display device 12, a communication device 13, a storage device 14, a processing device 15, and a bus 16.
[0030] Bus 16 is a wiring configuration for communicating information. Bus 16 connects the input device 11, the display device 12, the communication device 13, the storage device 14, and the processing device 15 to each other. Bus 16 may consist of, for example, a single bus. Bus 16 may also consist of buses provided between each device.
[0031] The input device 11 is the user interface of the terminal 10. The input device 11 includes a touch panel. In addition to the touch panel, the input device 11 may include a plurality of operation keys. The input device 11 may include a plurality of operation keys without including a touch panel. The input device 11 may include a voice input device. The input device 11 receives various inputs from the user U1. For example, the input device 11 receives a question A1 from the user U1. The input device 11 may also receive a comment from the user U1 that includes the question A1.
[0032] The display device 12 displays various information. For example, the display device 12 displays question A1, answer C1, and supporting statement D1. For example, the display device 12 displays question A1, answer C1, and supporting statement D1 together. The display device 12 may also display question A1 separately from answer C1 and supporting statement D1. In this case, the display device 12 displays answer C1 and supporting statement D1 together. The display device 12 may also display question A1, answer C1, and supporting statement D1 separately.
[0033] The communication device 13 can communicate with the server 30 via the communication network NW. The communication device 13 may also communicate with the server 30 without using the communication network NW. The communication device 13 sends question A1 to the server 30. If the input device 11 receives a comment containing question A1 from user U1, the communication device 13 may send the comment containing question A1 to the server 30. The communication device 13 receives answer C1 and justification description D1 from the server 30.
[0034] The storage device 14 is a recording medium that can be read by the processing device 15. The storage device 14 includes at least one memory. The storage device 14 includes, for example, non-volatile memory and volatile memory. Non-volatile memory is, for example, ROM (Read Only Memory), EPROM (Erasable Programmable Read Only Memory), and EEPROM (Electrically Erasable Programmable Read Only Memory). Volatile memory is, for example, RAM (Random Access Memory) and VRAM (Video Random Access Memory). The storage device 14 stores program PG1. Program PG1 includes at least one instruction.
[0035] The processing unit 15 includes at least one CPU (Central Processing Unit). The at least one CPU is an example of at least one processor. The at least one processor is an example of at least one computer. The processing unit 15 reads the program PG1 from the storage device 14. The processing unit 15 functions as an operation control unit 151 by executing the program PG1. The operation control unit 151 may be composed of circuits such as a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), or an FPGA (Field Programmable Gate Array).
[0036] The operation control unit 151 controls the operation of the terminal 10. For example, the operation control unit 151 causes the communication device 13 to send question A1 to the server 30. The operation control unit 151 causes the display device 12 to display question A1, answer C1, and justification description D1.
[0037] 1-3. Server 30 Figure 6 shows an example of a server 30. Server 30 is an example of a data acquisition device. Server 30 includes a communication device 31, a storage device 32, a processing device 33, and a bus 34.
[0038] Bus 34 is a wiring system for communicating information. Bus 34 connects the communication device 31, the storage device 32, and the processing device 33 to each other. Bus 34 may consist of, for example, a single bus. Bus 34 may also consist of buses provided between each device.
[0039] The communication device 31 can communicate with the terminal 10 via the communication network NW. The communication device 31 may also communicate with the terminal 10 without using the communication network NW. The communication device 31 can communicate with the LLM 21 via the communication network NW. The communication device 31 may also communicate with the LLM 21 without using the communication network NW. The communication device 31 may also communicate with an administrator terminal (not shown) operated by the administrator of the server 30.
[0040] The storage device 32 is a recording medium readable by the processing device 33. The storage device 32 includes at least one memory. The storage device 32 includes, for example, non-volatile memory and volatile memory. The storage device 32 stores a program PG2 and specialized information B1. The storage device 32 stores, for example, specialized information B1 obtained from another device. The program PG2 includes at least one instruction.
[0041] The processing unit 33 is another example of a data acquisition device. The processing unit 33 includes at least one CPU. The processing unit 33 reads the program PG2 from the storage device 32. The processing unit 33 functions as a plurality of elements by executing the program PG2. The plurality of elements include a system prompt setting unit 331, a question receiving unit 332, an answer acquisition unit 333, a control unit 334, an acquisition unit 335, and an output unit 336. Each of the plurality of elements may be composed of a circuit such as a DSP, ASIC, or FPGA.
[0042] Furthermore, the processing unit 33 does not necessarily have to function as any of the system prompt setting unit 331, question receiving unit 332, or answer acquisition unit 333. In this case, the system prompt setting unit 331, question receiving unit 332, and answer acquisition unit 333 may be mounted on an information processing device different from the server 30, for example.
[0043] The system prompt setting unit 331 sets system prompts. For example, the system prompt setting unit 331 sets a system prompt SP1 related to the answer C1 and a system prompt SP2 related to the ground statement D1 in response to an instruction from the administrator terminal.
[0044] FIG. 7 is a diagram illustrating an example of the system prompt SP1. The system prompt SP1 includes a request statement SP11, a question field SP12, and a specialized information field SP13.
[0045] The request statement SP11 is a statement that requests generation of the answer C1 to the question A1 using the specialized information B1. In FIG. 7, the request statement SP11 is the statement "Please output the answer to the following question with reference to the following specialized information". The request statement SP11 is not limited to the statement shown in FIG. 7, as long as it is a statement that requests generation of the answer C1 to the question A1 using the specialized information B1.
[0046] The question field SP12 is a field in which the question A1 is described. In the system prompt SP1, the question field SP12 is a blank field.
[0047] The specialized information field SP13 is a field in which the specialized information B1 is described. In the system prompt SP1, the specialized information field SP13 is a blank field.
[0048] FIG. 8 is a diagram illustrating an example of the system prompt SP2. The system prompt SP2 includes a request statement SP21, an answer field SP22, and a specialized information field SP23.
[0049] The request statement SP21 is a statement that requests extraction of the ground statement D1 as grounds. More specifically, the request statement SP21 is a statement that requests extracting, as grounds, the ground statement D1 including the ground E1 from the description B2 indicated in the specialized information B1, in extraction units based on the expression format of the ground portion B3 including the ground E1 for the answer C1.
[0050] In Figure 8, the request text SP21 reads: "The following answer was generated by LLM with reference to the following specialized information. Please extract the sentences used as justification from the specialized information when generating this answer. There may be multiple sentences that serve as justification. Please determine the unit of extraction according to the following rules and extract the justification. If the justification is written in the text of the specialized information, extract the justification at the paragraph level. If the justification is written in the table of the specialized information, extract the justification at the table level. If the justification is written in the QA set of the specialized information, extract the justification at the QA set level. If the justification is written in both the text and figures / tables of the specialized information, extract the justification at the combination level of paragraphs and figures / tables." The request text SP21 is not limited to the text shown in Figure 8, as long as it is a request to extract justification description D1 from description B2 shown in specialized information B1 as justification, based on the expression format of the justification section B3.
[0051] The answer field SP22 is where answer C1 is entered. In the system prompt SP2, the answer field SP22 is blank.
[0052] The Specialized Information field SP23 is where Specialized Information B1 is written. In the System Prompt SP2, the Specialized Information field SP23 is blank.
[0053] In Figure 6, the question receiving unit 332 receives question A1 from the terminal 10. The question receiving unit 332 receives question A1 from the terminal 10 via the communication device 31. For the sake of simplicity, the explanation of the communication device 31 will be omitted below. The question receiving unit 332 may also receive comments including question A1 from the terminal 10.
[0054] The answer acquisition unit 333 acquires the answer C1 to question A1 from the LLM 21. First, the answer acquisition unit 333 generates a prompt P1 related to answer C1. The answer acquisition unit 333 generates prompt P1 by writing question A1 and expert information B1 to the system prompt SP1.
[0055] The response acquisition unit 333 provides prompt P1 to the LLM 21 by sending prompt P1 to the LLM 21. The LLM 21 generates a response C1 in response to prompt P1. The response acquisition unit 333 retrieves the response C1 from the LLM 21. The LLM 21 may also generate a comment containing the response C1 in response to prompt P1. In this case, the response acquisition unit 333 retrieves the comment containing the response C1 from the LLM 21.
[0056] Figure 9 shows an example of Question A1. Question A1 is "What is the price for Service A?". Question A1 is not limited to the question shown in Figure 9 and can be changed as appropriate.
[0057] Figure 10 shows an example of specialized information B1. Specialized information B1 shows a description B2 of a service. Description B2 includes a table B21 titled "List of Various Services" and text B22 that describes the various services. Description B2 is not limited to the description shown in Figure 10 and can be changed as appropriate.
[0058] Figure 11 shows an example of prompt P1 generated by writing question A1 and expert information B1 in system prompt SP1. In prompt P1, question A1 is written in the question field SP12 of system prompt SP1. Also, in prompt P1, expert information B1 is written in the expert information field SP13 of system prompt SP1. Note that in Figure 11, the details of description B2 indicated by expert information B1 are omitted. Prompt P1 is an example of an instruction to generate answer C1 to question A1 using expert information B1. Note that the instruction to generate answer C1 to question A1 using expert information B1 is not limited to prompt P1 shown in Figure 11 and can be changed as appropriate.
[0059] Figure 12 shows an example of answer C1. Answer C1 is "The fee for service A is 1000 yen." Answer C1 is not limited to the answer shown in Figure 12 and can be changed as appropriate.
[0060] In Figure 6, the control unit 334 causes the LLM 21 to extract the justification description D1 from the specialized information B1 based on the answer C1 and the specialized information B1. As described above, the answer C1 is an example of information related to the answer. That is, the control unit 334 causes the LLM 21 to extract the justification description D1 from the specialized information B1 based on the information related to the answer and the specialized information B1. In the following, for the sake of simplicity, an example in which the information related to the answer is the answer C1 will be described. Note that the information related to the answer is not limited to the answer C1, as described above.
[0061] The control unit 334 first generates a prompt P2 related to the basis description D1. The control unit 334 generates prompt P2 by describing the answer C1 and the specialized information B1 in the system prompt SP2.
[0062] The control unit 334 provides prompt P2 to the LLM21 by transmitting prompt P2 to the LLM21. In response to prompt P2, the LLM21 extracts the basis description D1 from the specialized information B1.
[0063] Figure 13 shows an example of prompt P2. In prompt P2, answer C1 is written in the answer field SP22 of system prompt SP2. Also, in prompt P2, expert information B1 is written in the expert information field SP23 of system prompt SP2. Note that in Figure 13, the details of description B2 indicated by expert information B1 are omitted. Prompt P2 is an example of an instruction to generate the basis for answer C1 using expert information B1. The instruction to generate the basis for answer C1 using expert information B1 is not limited to prompt P2 shown in Figure 13 and can be changed as appropriate.
[0064] Figure 14 shows an example of evidence E1 included in specialized information B1. The description B2 shown by specialized information B1 includes the description "Service A 1000 yen 100GB" in Table B21 as evidence E1 for the answer C1, "The fee for Service A is 1000 yen." Therefore, in description B2, the evidence part B3 that includes evidence E1 becomes Table B21. Thus, the expression format of the evidence part B3 that includes evidence E1 is a "table". Since evidence E1 is written in Table 21 of specialized information B1, the evidence (evidence description D1) is extracted on a table basis.
[0065] Figure 15 shows an example of a justification statement D1. In Figure 15, Table B21 is extracted as the justification statement D1.
[0066] In this way, the control unit 334 transmits prompt P2 to the LLM 21, causing the LLM 21 to extract the basis description D1, which includes the basis E1, from the description B2, based on the expression format of the basis portion B3 that includes the basis E1.
[0067] Examples of the relationship between the expression format of the basis section B3 and the extraction unit are shown in Figures 3 and 4 above. If the expression format of the basis section B3 is a "table," the extraction unit is a "table." If the expression format of the basis section B3 is a "text," the extraction unit is a "paragraph." If the expression format of the basis section B3 is a "text," the extraction unit may be a "single sentence" instead of a "paragraph." If the expression format of the basis section B3 is a "question and answer set" (QA set), the extraction unit is a "question and answer set" (QA set). If the expression format of the basis section B3 is a "combination of text and a figure / table," the extraction unit is a "combination of a paragraph and a figure / table." If the expression format of the basis section B3 is a "combination of text and a figure / table," the extraction unit may be a "combination of a single sentence and a figure / table" instead of a "combination of a paragraph and a figure / table."
[0068] Figure 16 shows an example of a justification description D1 when the expression format of the justification section B3, which includes justification E1, is "text". In the example shown in Figure 16, the answer C1 is "Increase the value of parameter B." The description B2 shown by the expert information B1 includes text B23, text B24, and text B25. Texts B23, B24, and B25 are each paragraphs. As justification E1 for answer C1, description B2 includes the statement in text B24: "The speed increases as the value of parameter B increases, and the speed decreases as the value of parameter B decreases." Therefore, in description B2, the justification section B3, which includes justification E1, is text B24. Thus, the expression format of the justification section B3, which includes justification E1, is "text". Since justification E1 is written in text B24 of the expert information B1, the justification (justification description D1) is extracted on a paragraph basis. In this case, paragraph B24 is extracted as the supporting statement D1.
[0069] Figure 17 shows an example of a justification description D1 when the expression format of the justification section B3, which includes justification E1, is a "set of answer and question". In the example shown in Figure 17, the answer C1 is "Increase the value of parameter B." The description B2 shown by the expert information B1 includes a QA set (set of answer and question) B26 and a QA set (set of answer and question) B27. As justification E1 for answer C1, description B2 includes the description in QA set B26: "Q1: How can I increase the speed? A1: Increase the value of parameter B." Therefore, in description B2, the justification section B3, which includes justification E1, becomes QA set B26. Thus, the expression format of the justification section B3, which includes justification E1, is a "set of answer and question". Since justification E1 is described in QA set B26 of expert information B1, the justification (justification description D1) is extracted in units of a set of answer and question (QA set). In this case, QA set B26 is extracted as the justification statement D1.
[0070] Figure 18 shows an example of a justification description D1 when the expression format of the justification section B3, which includes justification E1, is a "combination of text and diagrams." In the example shown in Figure 18, the answer C1 is "Service C is recommended for heavy users." The description B2 shown by the expert information B1 includes Table B2a, text B2b, text B2c, and text B2d. As justification E1 for answer C1, description B2 includes the description "Service C 3000 yen 300GB" in Table B2a and the description "The fees for various services are as shown in Table 1, with Service C being best for high-capacity users, Service A for low-capacity users, and Service B for the middle tier" in text B2d. Therefore, in description B2, the justification section B3, which includes justification E1, consists of Table B2a and text B2d, respectively. Thus, the expression format of the justification section B3, which includes justification E1, is a "combination of text and diagrams." Since the basis E1 is described in both Table B2a and Text B2d of the specialized information B1, the basis (basis description D1) is extracted as a combination of paragraph and figure / table. In this case, Table B2a and Text B2d are extracted as basis description D1. Therefore, even if there are other descriptions between Table B2a and Text B2d in Description B2, only Table B2a and Text B2d can be extracted as basis description D1.
[0071] In Figure 6, the control unit 334 makes the extraction unit different when the representation format of the basis portion B3 is the first representation format compared to the extraction unit when the representation format of the basis portion B3 is a second representation format different from the first representation format. That is, the extraction unit based on the first representation format is different from the extraction unit based on the second representation format. "Table," which is the representation format of the basis portion B3, is an example of the first representation format. "Text," which is the representation format of the basis portion B3, is an example of the second representation format. The first representation format is not limited to "table." For example, the first representation format may be "text," "a set of questions and answers," or "a combination of text and a diagram." The second representation format can be changed as appropriate, not limited to "text," as long as it is a different representation format from the first representation format.
[0072] The acquisition unit 335 acquires the basis description D1 extracted by the LLM 21.
[0073] The output unit 336 outputs the justification description D1 to the terminal 10, thereby displaying the justification description D1 on the terminal 10. For example, the output unit 336 transmits the answer C1 and the justification description D1 to the terminal 10, thereby displaying the answer C1 and the justification description D1 on the display device 12 of the terminal 10. In other words, the output unit 336 presents the answer C1 and the justification description D1 to the user U1 using the display device 12. Note that the output unit 336 does not have to transmit the answer C1 to the terminal 10. In this case, the acquisition unit 335 may transmit the answer C1 to the terminal 10.
[0074] 1-4. Operational Explanation Figure 19 is a flowchart illustrating an example of the operation of the system prompt setting unit 331. The operation shown in Figure 19 is performed before question A1 is received.
[0075] In step S101, the system prompt setting unit 331 sets a system prompt SP1 related to answer C1 and a system prompt SP2 related to justification description D1, in response to instructions from the administrator terminal.
[0076] Figure 20 is a flowchart illustrating an example of the operation of server 30.
[0077] In step S201, the question receiving unit 332 receives question A1 from terminal 10.
[0078] Next, in step S202, the answer acquisition unit 333 generates prompt P1 by writing question A1 and expert information B1 to system prompt SP1.
[0079] Next, in step S203, the response acquisition unit 333 provides prompt P1 to the LLM21 by transmitting prompt P1 to the LLM21. The LLM21 generates response C1 in response to prompt P1.
[0080] Next, in step S204, the response acquisition unit 333 acquires the response C1 from the LLM 21. The response acquisition unit 333 may further transmit the response C1 to the terminal 10.
[0081] Next, in step S205, the control unit 334 generates prompt P2 by writing the answer C1 and specialized information B1 to the system prompt SP2.
[0082] Next, in step S206, the control unit 334 provides prompt P2 to the LLM21 by transmitting prompt P2 to the LLM21. That is, by transmitting prompt P2 to the LLM21, the control unit 334 causes the LLM21 to extract a basis description D1 containing basis E1 from description B2, based on the expression format of the basis portion B3 containing basis E1. The LLM21 generates the basis description D1 in response to prompt P2.
[0083] Next, in step S207, the acquisition unit 335 acquires the basis description D1 from the LLM 21.
[0084] Next, in step S208, the output unit 336 transmits the justification description D1 to the terminal 10, thereby causing the justification description D1 to be displayed on the display device 12 of the terminal 10. In other words, the output unit 336 presents the justification description D1 to the user U1 using the display device 12.
[0085] 1-5. Summary of Embodiments The server 30 includes a control unit 334, an acquisition unit 335, and an output unit 336. The control unit 334 causes the LLM 21 to extract the basis description D1 from the description B2 based on the expression format of the basis portion B3 of the description B2 shown in the expert information B1, based on the answer C1 generated based on the question A1 and the expert information B1. The basis portion B3 includes the basis E1 for the answer C1. The basis description D1 includes the basis E1. Therefore, the extraction unit of the basis description D1 can be set according to the expression format of the basis portion B3. Thus, it is possible to suppress the basis description D1 from becoming difficult to understand.
[0086] The control unit 334 makes the extraction unit when the representation format of the basis portion B3 is the first representation format different from the extraction unit when the representation format of the basis portion B3 is a second representation format different from the first representation format. In this case, the extraction unit of the basis description D1 can be changed according to the representation format of the basis portion B3. Therefore, it is possible to suppress the basis description D1 from becoming difficult to understand.
[0087] If the format of the supporting evidence B3 is a table, the extraction unit is a table. In this case, it is possible to prevent only a portion of the table from being extracted as supporting evidence D1.
[0088] If the expression format of evidence section B3 is a sentence, the extraction unit is a paragraph. In this case, descriptions related to evidence E1 can be included in evidence description D1. This makes evidence description D1 easier to understand.
[0089] If the expression format of the supporting evidence B3 is a sentence, the extraction unit may be just one sentence. In this case, the supporting evidence description D1 can be made more concise.
[0090] If the format of the evidence section B3 is a combination of text and figures / tables, the extraction unit is a combination of a paragraph and a figure / table. In this case, the evidence description D1 can include both a paragraph and a figure / table. Therefore, the evidence description D1 can be made easier to understand. Note that the paragraph and figure / table included in the evidence description D1 may be consecutive in description B2, or they may be separated from each other in description B2.
[0091] If the format of the evidence section B3 is a combination of text and a figure / table, the extraction unit may be a combination of one sentence and a figure / table. In this case, the evidence description D1 can include one sentence and a figure / table. Therefore, the evidence description D1 can be made easier to understand. The sentence and figure / table included in the evidence description D1 may be consecutive in description B2, or they may be separated from each other in description B2.
[0092] If the expression format of the basis section B3 is a set of question and answer, then the extraction unit is a set of question and answer. In this case, the basis description D1 can include a set of question and answer. Therefore, it is possible to prevent the basis description D1 from being only either the question or the answer within the set.
[0093] 2. Modifications The following are examples of modifications to the above-described embodiment. Two or more modifications can be arbitrarily selected from the following examples and combined as appropriate, within the bounds of mutual non-contradictory relationships.
[0094] 2-1. First Modification In the above embodiment, the control unit 334 causes the LLM 21 to extract the justification description D1 by using prompt P2. However, the control unit 334 may cause the LLM 21 to extract the justification description D1 without using prompt P2. For example, the control unit 334 may cause the LLM 21 to extract the justification description D1 by using multiple prompts.
[0095] Figure 21 shows a control unit 5, which is another example of the control unit 334. When the control unit 5 is used as the control unit 334, the system prompt setting unit 331 sets the system prompts SP1, SP3, SP4, and SP2 described above.
[0096] Figure 22 shows an example of system prompt SP3. System prompt SP3 is a system prompt related to basis E1. System prompt SP3 includes a request statement SP31, a response field SP32, and a specialized information field SP33.
[0097] Request statement SP31 is a sentence that requests the extraction of evidence E1 from description B2 based on answer C1 and expert information B1. In Figure 22, request statement SP31 is a sentence that reads, "The answer below was generated by LLM with reference to the expert information below. Please extract the sentences used as evidence when generating this answer from the expert information. It is acceptable to have multiple sentences that serve as evidence." Request statement SP31 is not limited to the sentence shown in Figure 22, as long as it is a sentence that requests the extraction of evidence E1 from description B2 based on answer C1 and expert information B1.
[0098] The answer field SP32 is where answer C1 is entered. In the system prompt SP3, the answer field SP32 is blank.
[0099] The Specialized Information field SP33 is where Specialized Information B1 is written. In the System Prompt SP3, the Specialized Information field SP33 is blank.
[0100] Figure 23 shows an example of system prompt SP4. System prompt SP4 is a system prompt relating to extraction unit F1. System prompt SP4 includes a request statement SP41, a basis field SP42, and a specialized information field SP43.
[0101] Request statement SP41 is a statement that requests the generation of extraction unit F1 based on expert information B1 and evidence E1 extracted by LLM21. In Figure 23, request statement SP41 is a statement that reads: "Please determine the extraction unit according to the following rules. If the evidence is described in the text of the expert information, extract the evidence at the paragraph level. If the evidence is described in the table of the expert information, extract the evidence at the table level. If the evidence is described in the QA set of the expert information, extract the evidence at the QA set level. If the evidence is described in the text and figures / tables of the expert information, extract the evidence at the combination level of paragraphs and figures / tables." Request statement SP41 is not limited to the statement shown in Figure 23, as long as it is a statement that requests the generation of extraction unit F1 based on expert information B1 and evidence E1 extracted by LLM21.
[0102] The "Reason" field SP42 is where Reason E1 is written. In the system prompt SP4, the "Reason" field SP42 is blank.
[0103] The SP43 field in the Specialized Information section is where Specialized Information B1 is entered. In the System Prompt SP4, the SP43 field in the Specialized Information section is blank.
[0104] Figure 24 shows an example of system prompt SP5. System prompt SP5 is a system prompt related to the basis description D1. System prompt SP5 includes a request statement SP51, a response field SP52, a specialized information field SP53, and an extraction unit field SP54.
[0105] Request statement SP51 is a statement requesting that the supporting statement D1 be extracted from description B2 using the extraction unit F1 generated by LLM21. In Figure 24, request statement SP51 is a statement that reads, "The answer below was generated by LLM with reference to the following specialized information. When generating this answer, please extract the sentences used as the basis from the specialized information. When doing so, please extract the basis using the following extraction unit. It is acceptable to have multiple sentences that serve as the basis." Request statement SP51 is not limited to the statement shown in Figure 24, as long as it is a statement requesting that the supporting statement D1 be extracted from description B2 using the extraction unit F1 generated by LLM21.
[0106] The answer field SP52 is where answer C1 is entered. In the system prompt SP5, the answer field SP52 is blank.
[0107] The Specialized Information field SP53 is where Specialized Information B1 is written. In the System Prompt SP5, the Specialized Information field SP53 is blank.
[0108] The extraction unit field SP54 is where the extraction unit F1 is entered. In the system prompt SP5, the extraction unit field SP54 is blank.
[0109] In Figure 21, the control unit 5 causes the LLM 21 to extract the justification description D1 from the specialized information B1 based on the answer C1 and the specialized information B1. The control unit 5 includes a first control unit 51, a second control unit 52, and a third control unit 53.
[0110] The first control unit 51 provides the LLM 21 with a first instruction to extract the basis E1 from the description B2 based on the answer C1 and the expert information B1, thereby causing the LLM 21 to extract the basis E1.
[0111] The first control unit 51 first generates a prompt P3 related to the basis E1. The first control unit 51 generates prompt P3 by writing the answer C1 and the specialized information B1 to the system prompt SP3.
[0112] The first control unit 51 provides prompt P3 to the LLM 21 by transmitting prompt P3 to the LLM 21. The LLM 21 generates a basis E1 in response to prompt P3. The first control unit 51 obtains the basis E1 from the LLM 21.
[0113] Figure 25 shows an example of prompt P3. In prompt P3, answer C1 is written in the answer field SP32 of system prompt SP3. Also, in prompt P3, expert information B1 is written in the expert information field SP33 of system prompt SP3. Note that in Figure 25, the details of description B2 indicated by expert information B1 are omitted. Prompt P3 is an example of a first instruction that extracts basis E1 from description B2 based on answer C1 and expert information B1. Note that the first instruction is not limited to prompt P3 shown in Figure 25 and can be changed as appropriate.
[0114] Figure 26 shows an example of evidence E1. In the example shown in Figure 26, the description "Service A 1000 yen 100GB" from Table B21 is extracted as evidence E1 for answer C1, "The fee for Service A is 1000 yen."
[0115] The second control unit 52 causes the LLM 21 to generate an extraction unit F1 by providing the LLM 21 with a second instruction to generate an extraction unit F1 based on the specialized information B1 and the basis E1 extracted by the LLM 21.
[0116] The second control unit 52 first generates a prompt P4 related to the extraction unit F1. The second control unit 52 generates prompt P4 by describing the basis E1 and specialized information B1 in the system prompt SP4.
[0117] The second control unit 52 provides prompt P4 to the LLM 21 by sending prompt P4 to the LLM 21. The LLM 21 generates an extraction unit F1 in response to prompt P4. The second control unit 52 obtains the extraction unit F1 from the LLM 21.
[0118] Figure 27 shows an example of prompt P4. In prompt P4, the basis E1 is described in the basis column SP42 of system prompt SP4. Also, in prompt P4, the expert information B1 is described in the expert information column SP43 of system prompt SP4. Note that in Figure 27, the details of description B2 indicated by expert information B1 are omitted. Prompt P4 is an example of a second instruction that generates an extraction unit F1 based on expert information B1 and basis E1 extracted by LLM21. Note that the second instruction is not limited to prompt P4 shown in Figure 27 and can be changed as appropriate.
[0119] Figure 28 shows an example of the extraction unit F1. In the example shown in Figure 28, since the basis E1 is listed in Table B21, the extraction unit F1 "Please output by table" is generated.
[0120] The third control unit 53 provides the LLM 21 with a third instruction to extract the basis description D1 from the description B2 using the extraction unit F1 generated by the LLM 21, thereby causing the LLM 21 to extract the basis description D1.
[0121] The third control unit 53 first generates a prompt P5 related to the basis description D1. The third control unit 53 generates prompt P5 by describing the answer C1, the specialized information B1, and the extraction unit F1 in the system prompt SP5.
[0122] The third control unit 53 provides prompt P5 to the LLM21 by transmitting prompt P5 to the LLM21. The LLM21 extracts the basis description D1 in response to prompt P5.
[0123] Figure 29 shows an example of prompt P5. In prompt P5, answer C1 is written in the answer field SP52 of system prompt SP5. Also in prompt P5, expert information B1 is written in the expert information field SP53 of system prompt SP5. Also in prompt P5, extraction unit F1 is written in the extraction unit field SP54 of system prompt SP5. Note that in Figure 29, the details of description B2 indicated by expert information B1 are omitted. Prompt P5 is an example of a third instruction that extracts the basis description D1 from description B2 using extraction unit F1 generated by LLM21. Note that the third instruction is not limited to prompt P5 shown in Figure 29 and can be changed as appropriate.
[0124] Figure 30 shows an example of a justification description D1. In the example shown in Figure 30, since justification E1 is listed in Table B21, Table B21 is extracted as the justification description D1. The justification description D1 is acquired by the acquisition unit 335.
[0125] In the first modified example, the first control unit 51 provides prompt P3 to the LLM 21, causing the LLM 21 to extract the basis E1. The second control unit 52 provides prompt P4 to the LLM 21, causing the LLM 21 to generate an extraction unit F1. The third control unit 53 provides prompt P5 to the LLM 21, causing the LLM 21 to extract the basis description D1. Therefore, by using multiple prompts, the LLM 21 can be made to extract the basis description D1.
[0126] 2-2. Second Modification In the above-described embodiment and the first modification, the answer C1 does not have to be generated by the LLM21. For example, the answer acquisition unit 333 may receive the answer C1 from the terminal 10. The second modification can also achieve the same effects as those achieved by the first embodiment.
[0127] 3. Other (1) Each function illustrated in Figure 5, Figure 6, or Figure 21 can be implemented by any combination of hardware and software. The method of implementing each function is not particularly limited. Each function may be implemented using one physically or logically coupled device, or it may be implemented using a device configured by directly or indirectly connecting two or more physically or logically separated devices (for example, using wired, wireless, etc.). Each function may be implemented by combining the above one device or the above multiple devices with software.
[0128] (2) In this disclosure, the term “apparatus” may be replaced with other terms such as circuit, device or unit.
[0129] (3) In each of the embodiments and the first to second modifications, each of the storage devices 14 and 32 may consist of at least one of the following: an optical disc such as a CD-ROM (Compact Disc ROM), a hard disk drive, a flexible disk, a magneto-optical disk (e.g., compact disc, digital multipurpose disc, Blu-ray® disc), a smart card, flash memory (e.g., card, stick, key drive), a floppy® disk, a magnetic strip, etc. The program may also be transmitted from a network via a telecommunications line.
[0130] (4) Each of the embodiments and the first to second modified examples is based on LTE (Long Term Evolution), LTE-A (LTA-Advanced), SUPER 3G, IMT-Advanced, 4G (4th generation mobile communication system), 5G (5th generation mobile communication system), 6th generation mobile communication system (6G), xth generation mobile communication system (xG) (where x is, for example, an integer or decimal), FRA (Future Radio Access), NR (new Radio), New radio access (NX), Future generation radio access (FX), W-CDMA (registered trademark), GSM (registered trademark), CDMA2000, UMB (Ultra Mobile Broadband), IEEE 802.11 (Wi-Fi (registered trademark)), IEEE 802.16 (WiMAX (registered trademark)), IEEE 802.20 may apply to at least one system utilizing UWB (Ultra-WideBand), Bluetooth®, or other appropriate systems, and to next-generation systems extended, modified, created, or defined based thereon. Alternatively, multiple systems may be applied in combination (e.g., a combination of at least one of LTE and LTE-A with 5G).
[0131] (5) The processing procedures, sequences, flowcharts, etc., exemplified in each of the embodiments and the first to second modifications may be rearranged in order, as long as they do not contradict each other. For example, the methods described in this disclosure present various step elements in an exemplary order and are not limited to the specific order presented.
[0132] (6) In each of the embodiments and the first to second modifications, the input and output information may be stored in a specific location (e.g., memory) or managed using a management table. The input and output information may be overwritten, updated, or appended to. The output information may be deleted. The input information may be transmitted to other devices.
[0133] (7) In each of the embodiments and the first to second modifications, the determination may be based on a value represented by one bit (0 or 1), on a Boolean value (true or false), or on a numerical comparison (for example, a comparison with a predetermined value).
[0134] (8) The programs illustrated in each of the embodiments and the first to second modifications should be broadly interpreted to mean instructions, instruction sets, code, code segments, program code, subprograms, software modules, applications, software applications, software packages, routines, subroutines, objects, executable files, execution threads, procedures, or functions, whether they are called software, firmware, middleware, microcode, or hardware description languages or by other names. Furthermore, software, or instructions, etc., may be transmitted or received via a transmission medium. For example, if software is transmitted from a website, server, or other remote source using at least one of wired technology (such as coaxial cable, fiber optic cable, twisted pair, and digital subscriber line (DSL)) and wireless technology (such as infrared, microwave, etc.), at least one of these wired and wireless technologies is included in the definition of a transmission medium.
[0135] (9) The information described in each of the embodiments and the first to second modifications may be represented using any of the following different technologies. For example, the data and information that may be referred to throughout the above description may be represented by voltage, current, electromagnetic waves, magnetic fields, magnetic particles, optical fields, photons, or any combination thereof. Notwithstanding the terms described herein and the terms necessary for understanding this disclosure, terms may be replaced with terms having the same or similar meanings.
[0136] (10) In each of the embodiments and the first to second modifications, the terms “system” and “network” are used interchangeably.
[0137] (11) In each of the embodiments and the first to second modifications, the server 30 may be, for example, a mobile station. A mobile station may also be referred to by those skilled in the art as a subscriber station, mobile unit, subscriber unit, wireless unit, remote unit, mobile device, wireless device, wireless communication device, remote device, mobile subscriber station, access terminal, mobile terminal, wireless terminal, remote terminal, handset, user agent, mobile client, client, or several other appropriate terms.
[0138] (12) A mobile station may also be called a transmitting device, receiving device, or communication device. A mobile station may also be a device mounted on a mobile body, or the mobile body itself. A mobile body means a movable object. The speed of movement of a mobile body is arbitrary. A mobile body can be stopped. A mobile body includes, but is not limited to, vehicles, transport vehicles, automobiles, motorcycles, bicycles, connected cars, excavators, bulldozers, wheel loaders, dump trucks, forklifts, trains, buses, handcarts, rickshaws, ships and other watercraft, airplanes, rockets, satellites, drones (registered trademark), multicopters, quadcopters, balloons, and things mounted on them. A mobile body may be a mobile body that moves autonomously based on operational commands. A mobile body may be a vehicle (e.g., a car, an airplane, etc.), an unmanned mobile body (e.g., a drone, an autonomous vehicle, etc.), or a robot (manned or unmanned). A mobile station also includes devices that do not necessarily move during communication operations. For example, the mobile station could be an IoT (Internet of Things) device such as a sensor.
[0139] (13) In each of the embodiments and the first to second modifications, the term “decision” may encompass a wide variety of actions. “Decision” may include, for example, judging, calculating, computing, processing, deriving, investigating, looking up, searching, inquiry (e.g., searching in a table, database or another data structure), and ascertaining. “Decision” may also include, for example, receiving (e.g., receiving information), transmitting (e.g., sending information), input, output, and accessing (e.g., accessing data in memory), and ascertaining. “Decision” may also include, for example, resolving, selecting, choosing, establishing, and comparing. In other words, "decision" can include considering that some action has been "decided." Furthermore, "decision" can be reinterpreted as "assuming," "expecting," or "considering."
[0140] (14) In each of the embodiments and the first to second modifications, the term “connected,” or any variation thereof, means any direct or indirect connection or coupling between two or more elements, and may include the presence of one or more intermediate elements between two elements that are “connected” or “coupled” to each other. The coupling or connection between elements may be physical, logical, or a combination thereof. For example, “connection” may be read as “access.” As used in the present disclosure, two elements may be considered to be “connected” or “coupled” to each other using at least one of one or more wires, cables, and printed electrical connections, and, in some non-limiting and non-exclusive examples, electromagnetic energy having wavelengths in the radio frequency domain, microwave domain, and optical (both visible and invisible) domain.
[0141] (15) In the first embodiment and each of the first to second modifications, the phrase "based on" does not mean "based solely on" unless otherwise specified. In other words, the phrase "based on" means both "based solely on" and "based at least on".
[0142] (16) Any reference to elements using the designations “first” and “second” as used in this disclosure does not limit the quantity or order of those elements in general. These designations may be used in this disclosure as a convenient way to distinguish between two or more elements. Accordingly, references to the first and second elements do not imply that only two elements may be adopted or that the first element must precede the second element in any way.
[0143] (17) Where “include,” “including,” and variations thereof are used in the Embodiments and the First to Second Modifications, these terms are intended to be inclusive, as are the terms “comprising.” Furthermore, where the terms “or” are used in the Embodiments and the First to Second Modifications, they are not intended to be exclusive OR.
[0144] (18) Where articles are added in the translation, for example, a, an, and the in English, the disclosure may include the fact that the noun following these articles is plural.
[0145] (19) The information, parameters, etc. described in this disclosure may be expressed using absolute values, relative values from a given value, or other corresponding information.
[0146] (20) In this disclosure, the term “A and B are different” may mean “A and B are different from each other.” The term may also mean “A and B are each different from C.” Terms such as “separate” and “combined” may be interpreted in the same way as “different.”
[0147] (21) Each aspect / embodiment described herein may be used individually, in combination, or switched between as needed during implementation. Furthermore, notification of certain information (e.g., notification that "it is X") is not limited to explicit notification, but may also be implicit (e.g., by not providing such notification).
[0148] (22) It will be obvious to those skilled in the art that the present invention is not limited to the embodiments described herein. The present invention can be implemented in modified and altered forms without departing from the spirit and scope of the invention as defined by the claims. Accordingly, the descriptions herein are for illustrative purposes only and are not intended to be restrictive in any way to the present invention. Furthermore, multiple embodiments selected from those illustrated herein may be combined.
[0149] 4. Aspects that can be understood from the above-described forms or modifications The following aspects can be understood from at least one of the above-described forms or modifications.
[0150] 4-1. First Embodiment The evidence acquisition device according to the first embodiment includes a control unit and an acquisition unit. The control unit causes a large-scale language model to extract evidence descriptions containing the evidence from the descriptions, based on information about a question and information about an answer generated based on related information related to the question, and the related information, in extraction units based on the expression format of the evidence portion of the description shown in the related information that contains the evidence for the answer. The acquisition unit acquires the evidence descriptions extracted by the large-scale language model. The output unit causes the evidence descriptions to be displayed on the terminal by outputting the evidence descriptions to the terminal.
[0151] According to this embodiment, the extraction unit of the supporting description can be set according to the expression format of the supporting part. Therefore, it is possible to suppress the possibility of the supporting description becoming difficult to understand.
[0152] 4-2. Second Embodiment In the example of the first embodiment (second embodiment), the control unit makes the extraction unit when the expression format of the basis portion is the first expression format different from the extraction unit when the expression format of the basis portion is a second expression format different from the first expression format. According to this embodiment, the extraction unit of the basis description can be changed according to the expression format of the basis portion. Therefore, it is possible to suppress the basis description from becoming difficult to understand.
[0153] 4-3. Third Embodiment In the example of the first embodiment (third embodiment), if the expression format of the basis portion is a table, the extraction unit is a table. According to this embodiment, it is possible to suppress the extraction of only a part of the table as the basis description.
[0154] 4-4. Fourth Embodiment In the example of the first embodiment (fourth embodiment), if the expression format of the basis portion is a sentence, the extraction unit is a paragraph. According to this embodiment, descriptions related to the basis can be included in the basis description. Therefore, the basis description can be made easier to understand.
[0155] 4-5. Fifth Embodiment In the example of the first embodiment (fifth embodiment), if the expression format of the basis part is a sentence, the extraction unit is one sentence. According to this embodiment, the basis description can be made concise.
[0156] 4-6. Sixth Embodiment In the example of the first embodiment (sixth embodiment), if the expression format of the basis portion is a combination of text and figures / tables, the extraction unit is a combination of a paragraph and a figure / table. According to this embodiment, paragraphs and figures / tables can be included in the basis description. Therefore, the basis description can be made easier to understand.
[0157] 4-7. Seventh Embodiment In the example of the first embodiment (seventh embodiment), when the expression format of the basis portion is a combination of text and a figure / table, the extraction unit is a combination of one sentence and a figure / table. According to this embodiment, the basis description can include one sentence and a figure / table. Therefore, the basis description can be made easier to understand.
[0158] 4-8. Eighth Embodiment In the example of the first embodiment (eighth embodiment), if the expression format of the basis portion is a set of question and answer, the extraction unit is a set of question and answer. According to this embodiment, the basis description can include a set of question and answer. Therefore, it is possible to prevent the basis description from consisting of only one of the question or answer within the set.
[0159] 4-9. 9th Embodiment In the example of the 9th embodiment, the control unit includes a first control unit, a second control unit, and a third control unit. The first control unit causes the large language model to extract the basis by providing the large language model with a first instruction to extract the basis from the description based on the information relating to the answer and the related information. The second control unit causes the large language model to generate the extraction unit by providing the large language model with a second instruction to generate the extraction unit based on the related information and the basis extracted by the large language model. The third control unit causes the large language model to extract the basis description by providing the large language model with a third instruction to extract the basis description from the description using the extraction unit generated by the large language model. According to this embodiment, the large language model can be made to extract the basis description by using a plurality of instructions.
[0160] 4-10. Tenth Embodiment The method for obtaining evidence according to the tenth embodiment includes: having a large-scale language model extract evidence descriptions containing the evidence from a description, based on information relating to a question and related information relating to the question, and the related information, in extraction units based on the expression format of the evidence portion of the description shown in the related information that contains the evidence for the answer; obtaining the evidence descriptions extracted by the large-scale language model; and displaying the evidence descriptions on a terminal by outputting the evidence descriptions to the terminal. According to this embodiment, the same effects as in the first embodiment can be achieved.
[0161] 5...Control unit, 10...Terminal, 11...Input device, 12...Display device, 13...Communication device, 14...Storage device, 15...Processing device, 16...Bus, 20...LLM system, 21...LLM, 30...Server, 31...Communication device, 32...Storage device, 33...Processing device, 34...Bus, 51...First control unit, 52...Second control unit, 53...Third control unit, 151...Operation control unit, 331...System prompt setting unit, 332...Question receiving unit, 333...Answer acquisition unit, 334...Control unit, 335...Acquisition unit, 336...Output unit.
Claims
1. A basis acquisition device comprising: a control unit that causes a large-scale language model to extract basis descriptions containing the basis for the answer from a description, based on information about a question and related information related to the question, and the related information, in an extraction unit based on the expression format of the basis portion of the description shown in the related information that contains the basis for the answer; an acquisition unit that acquires the basis descriptions extracted by the large-scale language model; and an output unit that displays the basis descriptions on a terminal by outputting the basis descriptions to the terminal.
2. The evidence acquisition apparatus according to claim 1, wherein the control unit makes the extraction unit different when the representation format of the evidence portion is a first representation format from the extraction unit when the representation format of the evidence portion is a second representation format different from the first representation format.
3. If the representation format of the basis portion is a table, the extraction unit is a table, as described in claim 1, the basis acquisition device.
4. The evidence acquisition device according to claim 1, wherein, if the expression format of the evidence portion is a sentence, the extraction unit is a paragraph.
5. The evidence acquisition device according to claim 1, wherein, when the expression format of the evidence portion is a sentence, the extraction unit is one sentence.
6. The evidence acquisition device according to claim 1, wherein, when the expression format of the evidence portion is a combination of text and a figure or table, the extraction unit is a combination of a paragraph and a figure or table.
7. The evidence acquisition device according to claim 1, wherein, when the expression format of the evidence portion is a combination of text and a figure or table, the extraction unit is a combination of one sentence and a figure or table.
8. The evidence acquisition device according to claim 1, wherein, if the expression format of the evidence portion is a set of questions and answers, the extraction unit is a set of questions and answers.
9. The evidence acquisition device according to claim 1, comprising: a first control unit that causes the large-scale language model to extract the evidence by providing the large-scale language model with a first instruction to extract the evidence from the description based on the information relating to the answer and the related information; a second control unit that causes the large-scale language model to generate the extraction unit by providing the large-scale language model with a second instruction to generate the extraction unit based on the related information and the evidence extracted by the large-scale language model; and a third control unit that causes the large-scale language model to extract the evidence description by providing the large-scale language model with a third instruction to extract the evidence description from the description using the extraction unit generated by the large-scale language model.
10. A method for obtaining evidence, comprising: causing a large-scale language model to extract evidence descriptions containing the evidence from a description, based on information about a question and related information related to the question, and the related information, in an extraction unit based on the expression format of the evidence portion of the description shown in the related information that contains the evidence for the answer; obtaining the evidence descriptions extracted by the large-scale language model; and displaying the evidence descriptions on a terminal by outputting the evidence descriptions to the terminal.
