Information processing device, information processing method, and program

Abstract

Provides information about the reliability of answers to questions. [Solution] An information processing device 1 has: a reception unit 131 that receives a prompt including content to be generated by an LLM (Large Language Model); a search unit 132 that searches a material database, which is a database of materials that are data including text, and extracts materials related to the content included in the prompt; a calculation unit 133 that calculates a validity score indicating the validity of each of the extracted materials as a material that serves as a basis for generating an answer to the prompt; a selection unit 134 that selects from the multiple materials a material that will serve as a basis for generating an answer based on the validity score for each of the multiple materials; and a display control unit 135 that displays the material selected by the selection unit 134 and the validity score of the material.

Description

[Technical Field]

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

[0002] Instructions are given to a large-scale language model to cause it to output a response (see, for example, Patent Document 1). [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Japanese Patent Application Publication No. 2023-158992 Summary of the Invention [Problem to be solved by the invention]

[0004] Generative AI (Artificial Intelligence) may output answers that differ from the facts (a phenomenon known as hallucination). To address this issue, it is possible to configure the generative AI so that users can check the documents that serve as the basis for generating the answer. However, since the documents may contain errors or may not be consistent with the question, it is difficult for users to determine how much they can trust the answers generated by the generative AI.

[0005] The present invention has been made in consideration of these points, and aims to provide information regarding the validity of answers given by a generation AI to questions. [Means for solving the problem]

[0006] An information processing device of a first aspect of the present invention has a reception unit that receives a prompt including content to be generated by an LLM (Large Language Model), a search unit that searches a material database, which is a database of materials that are data including text, and extracts materials related to the content included in the prompt, a calculation unit that calculates a validity score indicating the validity of each of the extracted materials as a material that serves as a basis for generating an answer to the prompt, a selection unit that selects a material from the plurality of materials that will serve as a basis for generating an answer based on the validity score for each of the plurality of materials, and a display control unit that displays the material selected by the selection unit and the validity score of the material.

[0007] The system may further include a generation unit that inputs the prompt and the material selected by the selection unit into the LLM and generates an answer to the prompt, and the display control unit may further display the answer generated by the generation unit.

[0008] The selection unit may select one or more materials from the plurality of materials that will serve as the basis for generating an answer based on the validity score for each of the plurality of materials, the display control unit may display a screen configured to allow a user to select one or more materials selected by the selection unit, and the generation unit may input the prompt and the material selected by the user from the one or more materials into the LLM and generate an answer to the prompt.

[0009] In the material database, a reliability of each of the search target materials is associated with the search target materials, and the display control unit may further display the reliability of the material selected by the selection unit.

[0010] The calculation unit may (1) calculate a degree of eligibility indicating the consistency between each of the multiple materials extracted by the search unit and the content included in the prompt, and (2) calculate the validity score for each of the multiple materials based on the reliability and eligibility for each of the multiple materials.

[0011] The search unit (1) calculates the relevance between the content contained in the prompt and the content contained in each of the materials to be searched, and (2) extracts multiple materials related to the content contained in the prompt based on the calculated relevance, and the calculation unit (1) calculates an eligibility indicating the consistency between each of the multiple materials extracted by the search unit and the content contained in the prompt, and (2) calculates the validity score for each of the multiple materials based on the reliability, relevance, and eligibility for each of the multiple materials.

[0012] An information processing method of a second aspect of the present invention includes a computer-executed reception step for receiving a prompt including content to be generated by an LLM; a search step for searching a material database, which is a database of materials that are data including text, and extracting materials related to the content included in the prompt; a calculation step for calculating a validity score indicating the validity of each of a plurality of materials extracted in the search step as a material that serves as a basis for generating an answer to the prompt; a selection step for selecting a material from the plurality of materials that will serve as a basis for generating an answer based on the validity score for each of the plurality of materials; and a display step for displaying the material selected in the selection step and the validity score of the material.

[0013] In a third aspect of the program of the present invention, a computer is caused to execute the following steps: a receiving step of receiving a prompt containing content to be generated by an LLM; a search step of searching a material database, which is a database of materials that are data containing text, and extracting materials related to the content contained in the prompt; a calculation step of calculating a validity score indicating the validity of each of the multiple materials extracted in the search step as a material that serves as a basis for generating an answer to the prompt; a selection step of selecting a material from the multiple materials that will serve as a basis for generating an answer based on the validity score for each of the multiple materials; and a display step of displaying the material selected in the selection step and the validity score of the material. [Effects of the Invention]

[0014] According to the present invention, it is possible to provide information regarding the reliability of answers to questions. [Brief explanation of the drawings]

[0015] [Figure 1] 1 is a diagram for explaining an overview of an information processing system S according to an embodiment. [Figure 2] 1 is a block diagram showing a configuration of an information processing device 1. FIG. [Figure 3] FIG. 2 is a diagram illustrating an example of a data structure of a document database. [Figure 4] 10 is a diagram showing an example of a screen displayed by a display control unit 135. FIG. [Figure 5] 10 is a diagram showing an example of a screen displayed by a display control unit 135. FIG. [Figure 6] 3 is a flowchart showing the flow of processing in the information processing device 1. DETAILED DESCRIPTION OF THE INVENTION

[0016] First Embodiment 1 is a diagram illustrating an overview of an information processing system S according to an embodiment. The information processing system S is a system for providing a question and answer system using a chatbot. The information processing system S includes an information processing device 1, an information terminal 2, and a document database 3.

[0017] The information processing device 1 is a device for receiving questions from an information terminal 2 used by a user and outputting answers to the received questions. The information processing device 1 stores an answer generation model. The answer generation model is an LLM (Large Language Model) that receives as input a prompt including a question and materials selected as the basis for the answer, and generates and outputs text that is an answer to the input question. The prompt is an instruction to the answer generation model. The prompt is text that includes a question that the answer generation model is desired to answer.

[0018] The information terminal 2 is a terminal used by a user of the information processing system S. The information terminal 2 is, for example, a smartphone, a tablet, or a personal computer.

[0019] The document database 3 is a searchable database of multiple documents that the answer generation model references as evidence for generating an answer based on a question received from a user. As will be described in detail later, the document database records the content of each of the multiple documents and information related to the reliability of the document (e.g., the source of publication, the time of publication, the number of citations, reliability, etc.). The document database 3 may be configured integrally with the information processing device 1.

[0020] The processing in the information processing system S will be described. The information processing device 1 receives a prompt from the information terminal 2 ((1) in FIG. 1). The information processing device 1 searches the material database 3 based on the received prompt and extracts materials for answering the prompt ((2) in FIG. 1). Specifically, the information processing device 1 generates a search query for searching the material database 3 based on the received prompt. The information processing device 1 searches the material database 3 based on the search query and extracts materials that match the conditions indicated by the search query.

[0021] The information processing device 1 calculates a validity score for each of the extracted materials ((3) in FIG. 1). The validity score is a score for quantitatively evaluating the validity of the material as a basis for generating an answer to a prompt. The calculation of the validity score will be described later.

[0022] The information processing device 1 selects, from the extracted materials, materials to be used as the basis for generating an answer based on the calculated validity score ((4) in FIG. 1). The information processing device 1 displays the selected materials and the validity scores of the selected materials on the information terminal 2 ((5) in FIG. 1).

[0023] The information processing device 1 may input the prompt and the selected material into the answer generation model, and cause the information terminal 2 to display the answer generated by the answer generation model.

[0024] By displaying information regarding the validity of the supporting documents for the answer, the information processing system S can provide information regarding the reliability of the answer to the question.

[0025] [Configuration of information processing device 1] 2 is a block diagram showing the configuration of the information processing device 1. The information processing device 1 has a communication unit 11, a storage unit 12, and a control unit 13. The control unit 13 has a reception unit 131, a search unit 132, a calculation unit 133, a selection unit 134, a display control unit 135, and a generation unit 136.

[0026] The communication unit 11 is a communication interface for transmitting and receiving data to and from other devices via a network. The storage unit 12 is a storage medium including a ROM (Read Only Memory), a RAM (Random Access Memory), an SSD (Solid State Drive), a hard disk drive, etc. The storage unit 12 stores in advance a program to be executed by the control unit 13.

[0027] The control unit 13 is a processor such as a CPU (Central Processing Unit), etc. The control unit 13 executes a program stored in the storage unit 12, thereby functioning as a reception unit 131, a search unit 132, a calculation unit 133, a selection unit 134, a display control unit 135, and a generation unit 136.

[0028] The receiving unit 131 receives a prompt including content to be generated by the answer generation model. As an example, the display control unit 135 causes the information terminal 2 to display a screen for inputting a prompt. The user operates the information terminal 2 and inputs the prompt on the displayed screen. The information terminal 2 transmits the input prompt to the information processing device 1. The receiving unit 131 acquires the transmitted prompt.

[0029] The search unit 132 searches a material database, which is a database of materials that are data including text, and extracts materials related to the content included in the prompt. The search unit 132 generates a search query based on the prompt accepted by the acceptance unit 131. As an example, the search unit 132 generates a search query configured to extract materials whose semantic vector similarity to the content included in the prompt is equal to or greater than a predetermined threshold. The search unit 132 searches the material database 3 based on the generated search query, and extracts materials that are highly related to the content included in the prompt.

[0030] FIG. 3 is a diagram showing an example of the data structure of a material database. In the material database, "material ID," "material name," "content," "material attribute," and "reliability" are associated with each other. "Material ID" is an ID (Identification) assigned to a material. "Material name" is the name of the material. "Content" is the text contained in the material. "Content" may also include images, audio, etc. "Material attribute" is information related to determining the reliability of the material. As an example, "material attribute" includes the source of publication such as the author or publisher, the time of publication, the number of citations, etc. "Reliability" is information indicating the degree to which the materials contained in the material database are reliable as a source of information.

[0031] The reliability is determined in advance based on, for example, the reliability of the information source, the expected degree of outdatedness, and the evaluations of others who use the material. For example, the reliability is set high when the publisher is a national or local government or a specific organization, and set low when the publisher is anonymous. The reliability may also be set higher the more recently published the material is, and lower the older the material is. The reliability may be set higher the more frequently cited the material is, and lower the more frequently cited the material is. The reliability may be expressed as a scale such as "A," "B," or "C," or as a quantitative number.

[0032] The search unit 132 calculates a relevance score between the content included in the prompt and the content included in each of the search target materials. The relevance score is a score indicating the similarity between the content included in the prompt and the content included in the search materials. For example, the similarity is the cosine similarity between the content included in the prompt and the content included in the search materials. The relevance score may also be calculated based on the distance between the semantic vectors of the content included in the prompt and the content included in the search materials. The search unit 132 extracts multiple materials related to the content included in the prompt based on the calculated relevance score. For example, the search unit 132 extracts a predetermined number of materials in descending order of the calculated relevance score. The predetermined number is determined based on the processing load for calculating the validity score.

[0033] The calculation unit 133 calculates a validity score indicating the validity of each of the extracted materials as a basis for generating an answer to the prompt. Specifically, the calculation unit 133 first calculates an eligibility score for each of the extracted materials and the content included in the prompt. The eligibility score is an evaluation value indicating the level of logical consistency between two texts. For example, if a sentence is composed of words with the same meaning, but the meanings indicated by the two texts as a whole are different, the eligibility score is evaluated as low. Also, if the meanings indicated by the two texts as a whole are consistent, the eligibility score is evaluated as high. For example, if two sentences are composed of words with similar meanings and the subject and object are interchangeable, the eligibility score is evaluated as low. As another example, if two sentences contain modifiers with similar meanings but the words modified by the modifiers are different, the eligibility score is evaluated as low.

[0034] The calculation unit 133 calculates a validity score for each of the multiple documents based on the reliability, relevance score, and eligibility score for each of the multiple documents. As an example, the calculation unit 133 calculates the validity score such that the higher the reliability, relevance score, and eligibility score, the higher the validity score. More specifically, the calculation unit 133 multiplies each of the reliability, relevance score, and eligibility score by a predetermined coefficient, and calculates the sum of the reliability, relevance score, and eligibility score multiplied by the coefficient as the validity score. Furthermore, the calculation unit 133 may calculate the sum of the reliability, relevance score, and eligibility score multiplied by the coefficient as the validity score.

[0035] The selection unit 134 selects, from the plurality of materials, a material that will be the basis for generating an answer, based on the validity score for each of the plurality of materials. As an example, the selection unit 134 selects, from the extracted materials, the material with the highest calculated validity score.

[0036] The display control unit 135 displays the material selected by the selection unit 134 and the relevance score of the material. Fig. 4 is a diagram showing an example of a screen displayed by the display control unit 135. The screen shown in Fig. 4 displays a prompt entered by the user and information about the material selected by the selection unit 134 based on the prompt entered by the user. The information about the selected material includes the title, author, publisher, publication date, number of citations, reliability, and relevance score of the material.

[0037] By configuring the information processing device 1 to display a score indicating the validity of the material selected as the basis for generating an answer, it is possible to provide information regarding the validity of the answer to the question.

[0038] Although the above description illustrates an example in which the search unit 132 extracts materials based on the relevance score, the search unit 132 may also be configured to extract materials based on the degree of match between words included in the prompt and words included in the search materials. In one example, the material database may associate an inverted index with words included in the materials, and materials whose order of appearance matches the order of words included in the prompt may be extracted based on the inverted index. As another example, the search unit 132 may extract materials related to the prompt based on BM25. For example, the search unit 132 may extract a predetermined number of materials based on the relevance score, and may also extract a predetermined number of materials based on the degree of match between words included in the materials and the prompt, and calculate the relevance score of the extracted materials based on the degree of match between words included in the materials and the prompt. Configuring the search unit 132 in this way enables the extraction of a variety of materials.

[0039] The information processing device 1 may be configured to further display the answer generated by the answer generation model based on the selected material.

[0040] The generation unit 136 inputs the prompt and the material selected by the selection unit 134 into the answer generation model, and generates an answer to the prompt. The display control unit 135 further displays the answer generated by the generation unit 136. The screen shown in FIG. 4 further includes an area showing the answer generated by the generation unit 136.

[0041] By configuring the information processing device 1 to display the answer generated based on the question and the validity of the supporting materials for the answer, it becomes possible to provide information for the user to determine whether the answer to the question is valid or not.

[0042] The information processing device 1 may be configured so that the user can select materials to be used to generate an answer from the materials extracted by the search unit 132.

[0043] The selection unit 134 selects one or more documents from the plurality of documents that will serve as the basis for generating an answer based on the validity score for each of the plurality of documents. As an example, the selection unit 134 selects a predetermined number of documents in descending order of validity score. The predetermined number is a number that is determined in advance as a number that can be confirmed without imposing a burden on the user selecting the documents. In this case, the predetermined number is, for example, 3, 5, 10, etc.

[0044] The display control unit 135 displays a screen configured to allow the user to select one or more materials selected by the selection unit 134. Fig. 5 shows an example of the screen displayed by the display control unit 135 in this case. The screen shown in Fig. 5 displays information on the multiple materials selected by the selection unit 134. The screen shown in Fig. 5 has buttons arranged to associate with each material and allow selection.

[0045] When the user operates the information terminal 2 to select a material, the reception unit 131 acquires information indicating the material selected by the user from the information terminal 2. The generation unit 136 inputs the prompt and the material selected by the user from one or more materials into the answer generation model, and generates an answer to the prompt.

[0046] By configuring the information processing device 1 to generate an answer based on materials selected by the user, the information processing device 1 can generate an answer that is more relevant to the user.

[0047] In the above example, the calculation unit 133 calculates the validity score based on the reliability, the relevance score, and the eligibility score. However, the calculation unit 133 may be configured to calculate the validity score based on the reliability and the eligibility score. That is, the calculation unit 133 calculates the validity score for each of the multiple documents based on the reliability and the eligibility score for each of the multiple documents. As an example, the calculation unit 133 calculates the validity score such that the higher the reliability and the eligibility score, the higher the validity score. More specifically, the calculation unit 133 multiplies each of the reliability and the eligibility score by a predetermined coefficient and calculates the sum of the reliability and the eligibility scores multiplied by the coefficient as the validity score. Alternatively, the calculation unit 133 may calculate the sum of the reliability and the eligibility scores multiplied by the coefficient as the validity score.

[0048] By calculating the validity score in this way, it becomes possible to evaluate materials with a simpler structure.

[0049] [Processing flow in information processing device 1] Fig. 6 is a flowchart showing the flow of processing in the information processing device 1. The flowchart shown in Fig. 6 starts from the point at which a prompt is received.

[0050] The receiving unit 131 receives a prompt from the information terminal 2 (S01). The searching unit 132 searches the material database for materials related to the prompt based on the prompt received by the receiving unit 131 (S02).

[0051] The calculation unit 133 calculates the validity score of each of the extracted materials (S03). The calculation unit 133 determines whether or not the validity score of any of the extracted materials is equal to or greater than a predetermined threshold (S04).

[0052] If the validity score of any of the extracted materials is not equal to or greater than the predetermined threshold (NO in S04), the display control unit 135 causes the information terminal 2 to display a message urging the user to re-enter the prompt (S05). Then, the information processing device 1 ends the process.

[0053] If the validity score of any of the extracted materials is equal to or greater than a predetermined threshold (YES in S04), the selection unit 134 selects a material to be used as the basis for the answer from the extracted materials based on the validity score (S06).

[0054] The generation unit 136 inputs the prompt and the extracted materials into the answer generation model, causing it to generate an answer to the prompt (S07). The display control unit 135 associates the answer generated by the answer generation model with information on the selected materials and displays them on the information terminal 2 (S08). Then, the information processing device 1 ends the process.

[0055] [Effects of information processing device 1] As described above, the information processing device 1 is configured to display a score indicating the validity of the material selected as the basis for generating an answer, thereby making it possible to provide information regarding the reliability of the answer to the question.

[0056] The present invention has been described above using embodiments, but the technical scope of the present invention is not limited to the scope described in the above embodiments, and various modifications and changes are possible within the scope of the gist of the present invention. For example, all or part of the device can be configured by functionally or physically distributing or integrating any unit. Furthermore, new embodiments resulting from any combination of multiple embodiments are also included in the embodiments of the present invention. The effects of the new embodiments resulting from the combination also have the effects of the original embodiments. [Explanation of symbols]

[0057] 1. Information processing equipment 2. Information terminal 3. Document database 11 Communications Department 12 Storage section 13 Control Unit 131 Reception 132 Search Section 133 Calculation Unit 134 Selection Section 135 Display control unit 136 Generation part

Claims

1. a reception unit that receives a prompt including content to be generated by an LLM (Large Language Model); a search unit that searches a material database, which is a database of materials that are data including text, and extracts materials related to the content included in the prompt; a calculation unit that calculates a validity score indicating the validity of each of the extracted materials as a basis for generating an answer to the prompt; a selection unit that selects a document that serves as a basis for generating an answer from the plurality of documents based on the validity score for each of the plurality of documents; a display control unit that displays the material selected by the selection unit and the validity score of the material; An information processing device having the above.

2. a generating unit that inputs the prompt and the material selected by the selecting unit into the LLM and generates an answer to the prompt; the display control unit further displays the answer generated by the generation unit. The information processing device according to claim 1 .

3. the selection unit selects one or more documents from the plurality of documents that will serve as a basis for generating an answer based on the validity score for each of the plurality of documents; the display control unit displays a screen configured to allow a user to select one or more materials selected by the selection unit, the generation unit inputs the prompt and a material selected by the user from the one or more materials into the LLM, and generates an answer to the prompt; The information processing device according to claim 2 .

4. In the document database, a reliability of each of the search target documents is associated with each of the search target documents, The display control unit further displays the reliability of the material selected by the selection unit. The information processing device according to claim 1 .

5. The search unit extracts a plurality of materials related to the content included in the prompt based on a similarity between the content included in the prompt and the materials; The calculation unit (1) calculating a degree of eligibility indicating the logical consistency between each of the plurality of documents extracted by the search unit and the content included in the prompt; (2) calculating the validity score for each of the plurality of documents based on the reliability and eligibility of each of the plurality of documents; The information processing device according to claim 4 .

6. The search unit (1) calculating a relevance indicating the degree of similarity between the content included in the prompt and the content included in each of the search target materials; (2) extracting a plurality of materials related to the content included in the prompt based on the calculated relevance; The calculation unit (1) calculating a degree of eligibility indicating the logical consistency between each of the plurality of documents extracted by the search unit and the content included in the prompt; (2) calculating the relevance score for each of the plurality of documents based on the reliability, relevance, and eligibility of each of the plurality of documents; The information processing device according to claim 4 .

7. The computer executes a receiving step of receiving a prompt including content to be generated by the LLM; a search step of searching a material database, which is a database of materials that are data including text, and extracting materials related to the content included in the prompt; a calculation step of calculating a validity score indicating the validity of each of the plurality of materials extracted in the search step as a material that serves as a basis for generating an answer to the prompt; a selection step of selecting a document from the plurality of documents that will be used as a basis for generating an answer based on the validity score for each of the plurality of documents; a display step of displaying the materials selected in the selection step and the appropriateness scores of the materials; An information processing method comprising:

8. On the computer, a receiving step of receiving a prompt including content to be generated by the LLM; a search step of searching a material database, which is a database of materials that are data including text, and extracting materials related to the content included in the prompt; a calculation step of calculating a validity score indicating the validity of each of the plurality of materials extracted in the search step as a material that serves as a basis for generating an answer to the prompt; a selection step of selecting a document from the plurality of documents that will be used as a basis for generating an answer based on the validity score for each of the plurality of documents; a display step of displaying the materials selected in the selection step and the appropriateness scores of the materials; A program that executes the following.

Images