Information processing device, information processing method, and program

Abstract

It provides a mechanism to complete any omissions in the content of questions entered by the user. [Solution] Obtain a first question sentence entered by the user. Obtain a second question sentence by supplementing the content omitted in the first question sentence based on one or more question sentences obtained before the first question sentence. Search for target information using the second question sentence. Generate a first instruction that outputs the answer to the second question sentence based on the search results obtained by the search means.

Description

Technical Field

[0001] This embodiment relates to an information processing apparatus, an information processing method, and a program.

Background Art

[0002] Since the emergence of ChatGPT, generative AI has continued to improve significantly in accuracy, and its utilization within companies has advanced in applications such as code generation and document summarization. Among these, the use of generative AI as a question-and-answer system has attracted a great deal of interest from many users. In order to create a generative AI that answers questions requiring local knowledge such as company rules, a technique called RAG (Retrieval-Augmented Generation) has sometimes been used.

[0003] Patent Document 1 discloses a dialogue system having a function of complementing the omission of user utterances. In this dialogue system, after complementing the omission of user utterances, the user utterance and the utterance pattern are collated, and a response utterance corresponding to the utterance pattern is returned. The correspondence between the utterance pattern and the response utterance is prepared in advance.

[0004] Also, Patent Document 2 discloses a FAQ question-and-answer system having a function of delving deeper into a question sentence. In this dialogue system, first, it is determined whether the input question sentence contains sufficient content for answer presentation. If it is determined that the content is insufficient, the question sentence is given to the generative AI to generate a probing question for concretizing the question sentence, and the generated probing question is presented to the user. After such question probing is repeated and the question sentence is sufficiently concretized, an appropriate answer sentence is obtained from the FAQ system using the question sentence.

Prior Art Documents

Patent Documents

[0005]

Patent Document 1

Patent Document 2

[0006] However, the technology disclosed in Patent Document 1 uses a dialogue system that responds based on pre-created and managed dialogue rules. Therefore, the dialogue rules must be modified every time a new utterance pattern or response utterance is added. This resulted in the problem of complicated management. In addition, because utterance patterns and response utterances are prepared in advance, there was also the problem of a limited range of responses.

[0007] Furthermore, the technology disclosed in Patent Document 2 has the problem that the user has to input the question multiple times to obtain the final answer, which places a heavy burden on the user. In addition, since the answers are pre-registered in the FAQ system, there is a problem that the range of answers is limited, similar to Patent Document 1.

[0008] This embodiment aims to provide a mechanism that complements the omission of the content of questions entered by the user. [Means for solving the problem]

[0009] The information processing device according to this embodiment is characterized by comprising: a first acquisition means for acquiring a first question sentence input by a user; a second acquisition means for acquiring a second question sentence in which the content omitted in the first question sentence is supplemented based on one or more question sentences acquired prior to the first question sentence; a search means for searching for target information using the second question sentence; and a first generation means for generating a first instruction to output an answer to the second question sentence based on the search results obtained by the search means. [Effects of the Invention]

[0010] Thus, according to this embodiment, it is possible to provide a mechanism that complements the omission of the content of questions entered by the user. [Brief explanation of the drawing]

[0011] [Figure 1] This is a block diagram showing an example of the functional configuration of the information processing device according to the embodiment. [Figure 2] This is a block diagram showing an example of a hardware configuration according to the embodiment. [Figure 3] A flowchart illustrating an example of processing in a question-answering system. [Figure 4] A flowchart illustrating an example of the detailed process for completing and concatenating omissions in question sentences. [Figure 5] A diagram illustrating prompts used for completing or omitting parts of a question. [Figure 6] A diagram illustrating prompts for obtaining answers using abbreviated and completed question sentences. [Figure 7] This figure shows an example of how answers to abbreviated and completed questions are displayed on a client device. [Modes for carrying out the invention]

[0012] Figure 1 is a block diagram showing an example of a functional unit of the information processing device 100 that constitutes the question answering system in an embodiment of the present invention.

[0013] The information processing device 100 includes a question acquisition unit 101, an omission and completion unit 102, a related document search unit 103, a prompt creation unit 104, an answer generation unit 105, a question storage unit 106, and a question storage unit update unit 107.

[0014] The question acquisition unit 101 acquires the question text entered by the user. The "question text" acquired here is a string used as a question, but the method of acquisition is not particularly limited. For example, the question acquisition unit 101 may acquire the string entered by the user as the question text, or it may generate a string to be used as the question text based on the user's voice. In this embodiment, the question acquisition unit 101 acquires the string entered by the user as the question text via a UI shown in Figure 7, which will be described later, on a client terminal (not shown), but it may also acquire the string entered by the user as the question text via an input controller 205, which will be described later.

[0015] The ellipsis completion unit 102 processes the question text acquired by the question acquisition unit 101 and completes any omitted elements. For example, the ellipsis completion unit 102 can determine whether or not there are omitted elements in the question text being processed by performing coreference analysis (for example, using previously acquired question texts). Here, the omitted elements may be, for example, case elements determined to be omitted by coreference analysis. For example, in a sentence where a so-called zero pronoun occurs, the case elements omitted in the sentence are considered omitted elements and are targeted for completion. In the following explanation, we will assume that completion of ellipsis in the sentence is performed by coreference analysis, but the process is not limited to this method as long as the location of ellipsis in the sentence can be estimated in a similar manner, and any known natural language processing technique may be used. For example, the ellipsis completion unit 102 may complete the omitted elements using a machine learning model that has been trained to take a sentence as input and output the location where ellipsis occurs and the omitted word in the sentence. In the following, when simply referred to as "ellipsis," it refers to the ellipsis of elements in the question text.

[0016] The ellipsis completion unit 102 according to this embodiment generates a prompt including an instruction to generate a corrected question by supplementing the content omitted in the question text to be processed based on the question texts acquired in the past, and can acquire the corrected question by inputting such an instruction into a large language model (generative AI). Since the prompt generated by such an ellipsis completion unit 102 will be described later with reference to FIG. 5, the description here will be omitted.

[0017] The related document search unit 103 performs a search for documents from the document database (DB) 110 outside the information processing apparatus 100 based on the question text on which the completion by the ellipsis completion unit 102 has been performed (or without completion if there is no completion part). In the following, the question text that has undergone such a completion process by the ellipsis completion unit 102 may be referred to as a "completed question text". The search by the related document search unit 103 according to this embodiment may be a process of performing a keyword search using the completed question text as a query, or may be a method using a vector search based on the semantic similarity with each document in the document DB, and any search method using the completed question text can be arbitrarily adopted. Here, for example, the related document search unit 103 can search for information specified by a URL that satisfies a predetermined condition. The document DB 110 for performing such a search will be described below.

[0018] Document DB 110 is a database that stores information serving as a basis for answers to questions from users. The Document DB 110 according to the present embodiment is, for example, a database that stores information posted on the employee-oriented site of a company, such as information on web pages or posted document files, when used in a search in a search system for an employee-oriented site of a company (including a question-and-answer system in an interactive (chat) format). Here, as described above, the Document DB 110 registers information specified by, for example, a URL that satisfies a predetermined condition, and it is possible to execute a search by the related document search unit 103 using such a Document DB 110. For example, the Document DB 110 registers information on a site obtained by searching for a URL by forward matching, and updates it at regular intervals, or registers data in advance by an administrator, etc., and can perform registration and management of documents. In this way, by performing a search from information based on a predetermined condition of a URL, such as information on a site searched by forward matching of a URL, it is possible to collect search information corresponding to local rules, such as using only information obtained from in-house documents.

[0019] The prompt creation unit 104 generates an instruction (a prompt including such an instruction) to generate an answer to the completed question text based on the search results by the related document search unit 103. The prompt creation unit 104 according to the present embodiment creates a prompt for answering with a generation AI using the completed question text and the search results by the related document search unit 103. Also, the prompt creation unit 104 issues an instruction to generate an output using the generated prompt. Here, the prompt creation unit 104 creates a prompt to be input to the generation AI so as to create an answer to the completed question text based on the content of the document that is the search result by the related document search unit 103.

[0020] In this embodiment, a large-scale language model using a method called RAG (Retrieval-Augmented Generation) can be used as a generative AI that outputs answers to question sentences. In a generative AI using RAG, for example, a search is performed in a document search system using the question sentence as a query, and a prompt is generated to be input to the generative AI based on the search results and the question sentence. Then, when this prompt is input to the generative AI, an answer to the question sentence is obtained. With a technology like RAG, it is possible to improve the accuracy of searching for related documents from a question sentence entered by a user by supplementing abbreviated parts of the question sentence. RAG is a well-known technology, and the details of the process of generating prompts are omitted here.

[0021] Figure 6 shows an example of a prompt created by the prompt creation unit 104 according to this embodiment and the output when the prompt is input to the generation AI. In the example prompt shown in Figure 6, in addition to the question, five search results are presented as information labeled DOC1 to DOC5, and the user is instructed to select one (or more) of them to use as the answer to the question, and to output the labels of the information used in the answer after "R:" so that they can be identified. In this example, in response to the question "Q: Can I take time off for a home visit?", the answer "R:DOC1 A: Yes, you can take time off for a home visit." is output.

[0022] In this way, the prompt generation unit 104 can generate prompts that, in addition to providing an answer to the completed question sentence, also output information that identifies the search result used from among the (multiple) search results obtained by the related document search unit 103 in that answer. By generating prompts that cause the generating AI to output the search result used in the answer in an identifiable way, the user can confirm what search results the answer is based on, making it easier for the user to understand whether the generating AI has referenced appropriate information (whether an appropriate answer has been provided).

[0023] The answer generation unit 105 generates an answer to the completed question sentence. In this embodiment, the answer generation unit 105 takes the prompt created by the prompt creation unit 104 as input to the generation AI and returns the obtained output as an answer (in this case, to the client terminal).

[0024] The question storage unit 106 stores question sentences for which answers have been obtained, and the question storage unit update unit 107 stores these question sentences as previously answered question sentences. Here, the question storage unit update unit 107 stores only the most recent question sentences and deletes previously stored question sentences, but the process is not limited to this as long as past question sentences are stored in a way that allows them to be referenced. For example, the question storage unit update unit 107 may store question sentences along with a timeline, allowing the most recent question sentence among those stored to be referenced.

[0025] With this configuration, it is possible to obtain a third question sentence by supplementing the content omitted in the first question sentence entered by the user based on the second question sentence obtained previously, search for the target information using the third question sentence, and then generate instructions to output the answer to the third question sentence based on the search results. Therefore, it is possible to supplement the omitted content of the question sentence without requiring the user to go through the trouble of multiple exchanges, and to obtain an answer from a question answering system using RAG as a method using the supplemented question sentence. Thus, it is possible to provide a mechanism for supplementing the content of questions entered by users in a question answering system using RAG as a method.

[0026] In Figure 1, the information processing device 100 is described as a single device, but the information processing device 100 may be composed of multiple devices as long as similar processing can be performed, and is not limited to this configuration. For example, the processing by the generating AI may be performed by a separate device, and a part of the processing described as being performed by the information processing device 100 may be performed by a client device.

[0027] Figure 2 is a block diagram showing an example of the hardware configuration of the information processing device 100 in an embodiment of the present invention. As shown in Figure 2, the information processing device 100 includes a CPU (Central Processing Unit) 201, ROM (Read Only Memory) 202, RAM (Random Access Memory) 203, input controller 205, video controller 206, memory controller 207, and communication I / F controller 208, all connected via a system bus 204.

[0028] The CPU 201 comprehensively controls each device and controller connected to the system bus 204. The ROM 202 or external memory 211 stores the control programs executed by the CPU 201, such as the BIOS (Basic Input / Output System) and OS (Operating System), as well as computer-readable and executable programs and various necessary data (including data tables) for realizing this information processing method.

[0029] RAM203 functions as the main memory or work area of ​​the CPU201. The CPU201 loads programs necessary for processing from ROM202 or external memory211 into RAM203, and then executes the loaded programs to perform various operations.

[0030] The input controller 205 controls input from an input device 209, such as a keyboard or a pointing device like a mouse (not shown). If the input device 209 is a touch panel, the user can give various instructions by pressing (touching with a finger, etc.) at the location of icons, cursors, or buttons displayed on the touch panel. This touch panel may be a multi-touch screen or other touch panel capable of detecting the positions of multiple fingers touching it.

[0031] The video controller 206 controls the display to an external output device such as the display 210. The display 210 in this embodiment may be the display of a notebook-type personal computer that is integrated with the main unit of the information processing device 100 (i.e., has the information processing device 100 built-in). The external output device used here is not limited to a display; for example, a projector may be used as the external output device. Furthermore, for the device that can accept touch operations as described above, an input device is also provided.

[0032] Furthermore, the video controller 206 according to this embodiment is capable of controlling the video memory (VRAM) for display control. In this case, the video controller 206 may use a portion of the RAM 203 as the video memory area, or it may provide a separate dedicated video memory.

[0033] The memory controller 207 controls access to the external memory 211. The external memory 211 may be an external storage device (hard disk), a flexible disk (FD), or a CompactFlash® memory connected to a PCMCIA card slot via an adapter, which stores boot programs, various applications, font data, user files, editing files, and various other data.

[0034] The communication interface controller 208 connects to and communicates with external devices via a network and performs communication control processing over the network. The communication interface controller 208 can, for example, communicate using TCP / IP, telephone lines such as ISDN, or 3G mobile phone lines.

[0035] The CPU 201 enables display on the display 210 by, for example, performing the process of expanding (rasterizing) outline fonts into the display information area in RAM 203. The CPU 201 also enables user input via a mouse cursor (not shown) on the display 210.

[0036] Next, the response generation process performed by the question answering system, including the information processing device 100 according to this embodiment, will be described using the flowcharts in Figures 3 and 4. The flowchart in Figure 3 is an example of the process by which the question answering system generates a response after receiving a user question. The process shown in Figure 3 is assumed to start when the information processing device 100 receives a question.

[0037] In S301, the question acquisition unit 101 acquires the question text entered by the user and sets the question text as the target for processing. In S302, the omission completion unit 102 completes the acquired question text and generates a completed question text. Details of S302 will be explained with reference to Figure 4.

[0038] Figure 4 is a flowchart showing an example of the details of the question completion process performed in S302. S401-S402 are processes that determine whether or not to perform omission completion on the acquired question.

[0039] In S401, the omission completion unit 102 determines whether or not to perform omission completion for the question sentence to be processed. Here, the omission completion unit 102 determines to perform omission completion if there is a question sentence already stored by the question storage unit 106 (a question sentence entered before the question sentence to be processed). If such a question sentence exists, the process proceeds to S402; otherwise, the completion process is not performed and the process proceeds to S403.

[0040] In S402, the ellipsis completion unit 102 determines whether the question sentence to be processed contains omitted elements, and if it does, it completes them. Here, the ellipsis completion unit 102 can use question sentences that were input before the question sentence to be processed, whose existence was confirmed in S401, to determine whether the question sentence to be processed contains omitted elements using natural language processing techniques as described above. Alternatively, for example, the ellipsis completion unit 102 may determine that the question sentence to be processed contains omitted elements if the total number of characters in the question sentence is below a predetermined threshold (for example, 10 characters or less). Furthermore, for example, the ellipsis completion unit 102 may use a machine learning model trained to infer and complete omitted elements contained in a sentence to determine whether the question sentence to be processed contains omitted elements. In S403, the ellipsis completion unit 102 completes the process shown in Figure 4 and proceeds to S303.

[0041] The omission completion unit 102 may also use a generation AI that simultaneously or separately performs the following processes: determining whether or not to complete the omissions in the question sentence in S401, and completing the omissions in the question sentence in S402 if the question sentence contains omissions. In this case, the prompts input to the generation AI may be different prompts in S401 and S402, or they may be the same prompt in S401 and S402.

[0042] Figure 5 shows an example of a prompt that is input when S401 and S402 are performed simultaneously by the generation AI, and the completed question sentence that is generated by that prompt. In Figure 5, the sentence "What about home visits?" is input as the question sentence to be processed (Q2), and the sentence "Can I take time off in units of hours for a class observation?" is input in the prompt as a previously input question sentence (Q1). In this case, the prompt is instructed to generate a question sentence Q3 that completes the parts omitted in Q2 based on Q1. As a result of completing the omissions, the omission completion unit 102 outputs the sentence "Can I take time off in units of hours for a home visit?" as Q3.

[0043] In Figure 5, if the topic changes between Q1 and Q2, the prompt is configured to use Q2 as is for Q3. Thus, the ellipsis completion unit 102 can generate prompts that instruct the system to complete the ellipsis if the content of the current question is related to the content of the previous question (the topic has not changed), and not to complete the ellipsis if there is no relationship (the current question is used as the final question input to the AI). Alternatively, the ellipsis completion unit 102 may determine whether the content of the current question is related to the content of the previous question, generate a prompt to complete the ellipsis if a relationship is found, and not generate such a prompt if no relationship is found, using Q2 as the completed question. Alternatively, the ellipsis completion unit 102 may be configured to use not only the previously input question (Q1), but also multiple past questions from further back. For example, it may be configured to use a predetermined number of questions, or it may use all questions from the start of the question-answer session until the dialogue is cleared (updated).

[0044] In S303, the related document search unit 103 searches the document database 110 for documents related to the completed query generated in S302. In S304, the related document search unit 103 determines whether one or more related documents were found in S303. If one or more documents are found, the process proceeds to S305; otherwise, the process proceeds to S307. In S307, the related document search unit 103 uses the search failure message as the response and proceeds to S308. The search failure message is assumed to be a predetermined message.

[0045] In S305, the prompt creation unit 104 uses the completed question text and the search results from the related document search unit 103 to create a prompt for the generation AI to generate an answer, and outputs it to the generation AI. In S306, the answer generation unit 105 inputs the created prompt to the generation AI, generates an answer to the completed question text, and proceeds to S308.

[0046] In S308, the question storage update unit 107 stores the completed question as a past question. Here, the question storage update unit 107 replaces the currently processed single past question that was previously stored. In S309, the information processing device 100 outputs the generated answer (and possibly a search failure message) to the client terminal and terminates processing. Note that, in addition to the answer, information on related documents may also be returned at the same time.

[0047] Figure 7 shows an example of how the answers to the completed question sentences output in S309 are displayed on the client device. In the example in Figure 7, the previous question and its answer are displayed side by side with the question and its answer that was processed this time. In Figure 7, information on related documents used as search results is also displayed alongside the answers.

[0048] In Figure 7, a field for the user to input the next question and a button to send the entered question to the information processing device 100 are displayed at the bottom of the UI. The acquisition of the question in S301 can be performed, for example, by acquiring user input through such a UI. Alternatively, for example, a button to enable "conversation update" could be provided on the screen, allowing the user to clear the display of previous question answers each time the theme or context of the question changes.

[0049] Although one embodiment of the information processing device 100 has been described above, the present invention can take the form of, for example, a system, device, method, program, or recording medium. Specifically, the information processing device 100 may be applied as a system composed of multiple devices, or as a device consisting of a single device.

[0050] Furthermore, the program in this invention is a program that allows a computer to execute the processing method shown in the flowchart in Figure 3, and the storage medium of this invention stores a program that allows a computer to execute the processing method shown in Figure 3. Note that the program in this invention may also be a program for each processing method of each device shown in Figure 3.

[0051] As described above, it goes without saying that the object of the present invention can also be achieved by supplying a recording medium containing a program that realizes the functions of the embodiments described above to a system or device, and by having the computer (or CPU or MPU) of the system or device read and execute the program stored on the recording medium.

[0052] In this case, the program read from the recording medium itself realizes the novel function of the present invention, and the recording medium on which that program is recorded constitutes the present invention.

[0053] For recording media used to supply programs, examples include flexible disks, hard disks, optical disks, magneto-optical disks, CD-ROMs, CD-Rs, DVD-ROMs, magnetic tapes, non-volatile memory cards, ROMs, EEPROMs, or silicon disks.

[0054] Furthermore, it goes without saying that the functions of the aforementioned embodiments are realized not only by the computer executing the program it has read, but also by the operating system (OS) running on the computer performing some or all of the actual processing based on the instructions of that program, thereby realizing the functions of the aforementioned embodiments.

[0055] Furthermore, it goes without saying that this also includes cases where, after a program read from a recording medium is written to the memory of a function expansion board inserted into a computer or a function expansion unit connected to a computer, the CPU or other components of the function expansion board or function expansion unit perform some or all of the actual processing based on the instructions of the program code, and the functions of the aforementioned embodiments are realized through that processing.

[0056] Furthermore, the present invention may be applied to a system composed of multiple devices or to a device consisting of a single device. It goes without saying that the present invention can also be applied when the results are achieved by supplying a program to a system or device. In this case, by reading a recording medium containing a program for achieving the present invention into the system or device, the system or device can enjoy the effects of the present invention.

[0057] Furthermore, by downloading and reading the program for achieving the present invention from a server, database, or the like on a network using a communication program, the system or device can enjoy the effects of the present invention. It should be noted that all configurations combining the above-described embodiments and their modified forms are also included in the present invention. [Explanation of Symbols]

[0058] 100: Information processing device, 101: Question acquisition unit, 102: Omission and completion unit, 103: Related document search unit, 104: Prompt creation unit, 105: Answer generation unit, 106: Question storage unit, 107: Question storage unit update unit, 110: Document DB

Claims

1. A first acquisition means for obtaining a first question sentence entered by the user, A second acquisition means for acquiring a second question sentence in which the content omitted in the first question sentence is supplemented based on one or more question sentences acquired prior to the first question sentence, A search means for searching for target information using the second question sentence described above, A first generation means generates a first instruction that outputs an answer to the second question based on the search results obtained by the search means, An information processing device characterized by comprising:

2. The system further comprises a second generation means for generating a second instruction to output a second question sentence in which the content omitted in the first question sentence is supplemented based on one or more question sentences obtained prior to the first question sentence, The information processing apparatus according to claim 1, characterized in that the second acquisition means acquires the second question sentence based on the second instruction.

3. The information processing apparatus according to claim 2, characterized in that the second instruction is an instruction input to a large-scale language model.

4. The information processing apparatus according to claim 2, characterized in that, when it is determined that the first question contains elements that have been omitted, the apparatus outputs a second question in which the content omitted in the first question has been supplemented based on one or more question sentences obtained prior to the first question.

5. The information processing apparatus according to claim 1, characterized in that if it is determined that the first question does not contain any omitted elements, the first question is used as the second question.

6. The information processing apparatus according to claim 2, characterized in that, if it is determined that the first question statement does not contain any omitted elements, the second instruction is generated such that the first question statement is used as the second question statement.

7. The information processing apparatus according to claim 1, characterized in that if it is determined that the content of the first question is unrelated to the content of one or more question sentences obtained prior to the first question, the first question is used as the second question.

8. The information processing apparatus according to claim 2, characterized in that, if it is determined that the content of the first question is unrelated to the content of one or more question sentences obtained prior to the first question, the second instruction is generated such that the first question is used as the second question.

9. The information processing apparatus according to claim 1, characterized in that the first generation means generates the first instruction to output information that identifies the first search result among the search results obtained by the search means used when generating the answer to the second question.

10. The information processing apparatus according to claim 1, characterized in that the first instruction is an instruction input to a large-scale language model.

11. The information processing apparatus according to claim 1, characterized in that the search target information is information identified by a URL that satisfies predetermined conditions.

12. The information processing apparatus according to claim 11, characterized in that the search target information is site information obtained by searching for URLs using prefix matching.

13. A first acquisition step involves obtaining a first question sentence entered by the user, A second acquisition step of acquiring a second question sentence by supplementing the content omitted in the first question sentence based on one or more question sentences acquired before the first question sentence, A search process to search for target information using the second question statement described above, A generation step that generates a first instruction to output an answer to the second question based on the search results in the search step, An information processing method characterized by comprising:

14. A program for causing a computer to function as one of the means of an information processing device according to any one of claims 1 to 12.

Images