Reply generation support system, reply generation support method, and reply generation support program
By dividing instruction data and using a large-scale language model to generate outputs from multiple pieces of data, the method addresses token limitations and enhances the accuracy and generation of long answers in natural language processing systems.
Patent Information
- Application Number
- JP2023075714
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2023-05-01
- Publication Date
- 2024-11-14
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Large-scale language models face challenges in generating accurate long answers due to token limitations and the need for users to devise appropriate input methods, which can delay processing and reduce accuracy.
A method is introduced to divide instruction data into multiple parts and transmit it to an answer generation system, allowing for the acquisition of answers without token restrictions, using a large-scale language model to generate outputs based on multiple pieces of data.
This approach supports the generation of desired answers, particularly long answers, by overcoming token limitations and improving accuracy through divided data transmission.
Smart Images

Figure 2024160571000001_ABST
Abstract
Description
[Technical field]
[0001] The present invention relates to an answer generation support system, an answer generation support method, and an answer generation support program that support the generation of a desired answer. [Background technology]
[0002] With the development of natural language processing technology, the accuracy of dialogue via human-machine interfaces is improving.
[0003] Non-Patent Document 1 discloses an AI chatbot built on the basis of a large-scale language model (LLM). The AI chatbot described in Non-Patent Document 1 is characterized by generating answers that humans find natural in response to questions, requests, calls, etc. input by humans. In particular, because of its high accuracy in responding to all inputs, it is increasingly being used in information gathering, text summarization, text creation including articles and scripts, and automatic response systems via websites, etc.
[0004] Non-Patent Document 2 discloses a generation support service that supports the generation of articles, scripts, catchphrases, etc., constructed based on a large-scale language model. The generation support service described in Non-Patent Document 2 is capable of acquiring a generation result from a large-scale language model corresponding to the input of necessary items related to the text to be generated. [Prior art documents] [Non-patent literature]
[0005] [Non-Patent Document 1] "Introducing ChatGPT", [Retrieved April 20, 2023], Internet<URL:https: / / openai.com / blog / chatgpt> [Non-Patent Document 2] "Catchy Writing Assistant," [Retrieved April 20, 2023], Internet<URL:https: / / lp.ai-copywriter.jp / > Summary of the Invention [Problem to be solved by the invention]
[0006] While large-scale language models can generate highly accurate answers, the type of input they provide has a large impact on the accuracy of the answer. Therefore, in order to obtain the answer they want, users need to understand the characteristics of the large-scale language model to some extent and then consider what type of input to provide.
[0007] Furthermore, large-scale language models usually have restrictions on tokens. A token is a unit for processing natural language. The token restriction means that the amount of data that can be input at one time is limited by the model structure (e.g., neural network structure) of the large-scale language model. Furthermore, from the viewpoint of an increase in processing load due to excessive access to the large-scale language model and a delay in processing speed, it is also conceivable that a restriction is set on the amount of data that can be communicated at one time. Therefore, a challenge has been that users are required to be creative with the content and method of input, particularly when trying to obtain a long answer.
[0008] The present invention aims to provide a technique for supporting the generation of a desired answer, and in particular, to provide a technique for favorably supporting the generation of a long answer. [Means for solving the problem]
[0009] In order to solve the above-mentioned problems, the present invention is a method executed by a computer, which includes the steps of generating instruction data based on request data, dividing the instruction data a predetermined number of times and transmitting it to an answer generation means, and obtaining multiple answer data from the answer generation means.
[0010] With this configuration, the instruction data can be divided and the answers can be obtained from the answer generating means. In particular, the result of the answer generating means can be obtained without being subject to token restrictions.
[0011] The present invention is also a method further comprising the steps of: receiving an input of the request data; and generating an output based on the plurality of response data in the obtaining step, the method being executed by a computer.
[0012] The present invention also provides a system comprising: a request acquisition means for acquiring request data; a first generation means for generating instruction data based on the request data; an instruction means for dividing the instruction data a predetermined number of times and transmitting the divided data to an answer generation means; and a second generation means for generating an output based on multiple answer data acquired by the answer generation means.
[0013] The present invention is also a program that causes a computer to function as a request acquisition means for acquiring request data, a first generation means for generating instruction data based on the request data, an instruction means for dividing the instruction data a predetermined number of times and transmitting it to an answer generation means, and a second generation means for generating an output based on multiple answer data acquired by the answer generation means.
[0014] With this configuration, it is possible to obtain output based on a plurality of response data from a simple request of the user without being subject to token restrictions.
[0015] In a preferred embodiment of the invention, the answer generating means is a large scale language model. With this configuration, it is possible to obtain an answer based on highly accurate natural language processing.
[0016] In a preferred embodiment of the present invention, the desired data includes a category of an output to be generated, and the first generating means obtains a template corresponding to the category and generates the instruction data based on the template. With this configuration, it is possible to easily generate instruction data according to a category and obtain a desired answer.
[0017] In a preferred embodiment of the present invention, the request data further includes example sentence data corresponding to a category, and the first generating means generates the instruction data by inputting the example sentence data into a predetermined portion of the template. With this configuration, instruction data corresponding to example sentence data can be easily generated, and an answer that is in line with the tendency of the example sentence can be obtained.
[0018] In a preferred embodiment of the present invention, a division rule corresponding to the category is defined in the template, and the instruction means transmits an answer generation request to the answer generating means in accordance with the number of divisions defined in the division rule. In a preferred embodiment of the present invention, the division rule is a rule that limits the number of characters included in the answer generation request to a predetermined number or less. With this configuration, it is possible to generate an answer without being subject to token restrictions. Also, by dividing the answer appropriately, it is possible to improve the accuracy of the answer.
[0019] In a preferred form of the present invention, the instruction data has a common area and divided areas obtained by dividing the common area a predetermined number of times, and the instruction means divides the instruction data into different divided areas and transmits them to the answer generation means. With this configuration, answers can be generated for divided areas in the common area, improving the accuracy of the answers.
[0020] In a preferred embodiment of the present invention, the second generating means generates the output based on a plurality of pieces of answer data included in a plurality of pieces of dialogue data each consisting of one piece of the instruction data and one piece of the answer data. With this configuration, multiple answers can be extracted to generate a single output, and even when a simple request is input, a relatively long output can be obtained. Effect of the Invention
[0021] According to the present invention, it is possible to assist in generating a desired answer, and in particular, according to the present invention, it is possible to suitably assist in generating a long answer. [Brief description of the drawings]
[0022] [Figure 1] FIG. 1 is a block diagram of a system according to an embodiment of the present invention. [Diagram 2] FIG. 2 is a diagram showing the hardware configuration of the present embodiment. [Diagram 3] FIG. 4 is a process sequence diagram of the present embodiment. [Figure 4] 4 shows an example of request data and an input screen thereof according to the present embodiment. [Diagram 5] 4 is a process flowchart of the generation support device according to the present embodiment. [Figure 6] 4 shows an example of template data and its input screen according to the present embodiment. [Figure 7] 4 shows an example of instruction data according to the present embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0023] Hereinafter, an answer generation support system according to an embodiment of the present invention will be described with reference to the drawings. Note that the embodiment described below is an example of the present invention, and the present invention is not limited to the embodiment described below, and various configurations can be adopted.
[0024] In this embodiment, the configuration, operation, etc. of an answer generation support system, a generation support device, and a generation support program are described, but a method with a similar configuration, a computer program, and a program recording medium on which the program is recorded also have the same effects. By using a program recording medium, for example, the program can be installed on a computer. A series of processes according to this embodiment described below is provided as a computer-executable program, and can be provided via a non-transitory computer-readable recording medium such as a CD-ROM or a flexible disk, or even via a communication line.
[0025] The answer generation support system is configured by a computer device. The computer device has an arithmetic device such as a CPU (Central Processing Unit) and a storage device. The computer device can function as a generation support device by executing a generation support program stored in the storage device with the arithmetic device. The generation support method is realized by processing of the computer device including the generation support device.
[0026] In this embodiment, the answer generation support system provides an answer as an output in response to an input by a user. Examples of the input by a user include, but are not limited to, a question, a request, or a call. In other words, the input by a user can be defined into a plurality of categories. The answer generation support system can provide an appropriate answer according to the category.
[0027] 1 shows a block diagram of an answer generation support system 1. The answer generation support system 1 includes a generation support device 2, a user terminal 3, and an answer generation device 4, and these components are communicatively connected via a communication network NW.
[0028] The generation support device 2 has, as functional components, a request acquisition means 21 for acquiring request data, a first generation means 22 for generating instruction data, an instruction means 23 for sending an answer generation request, and a second generation means 24 for generating an output based on the result of the answer generation request.
[0029] The generation support device 2 receives data from the user terminal 3. Based on the received data, the generation support device 2 transmits an answer generation request to the answer generation device 4. The generation support device 2 receives a result of the answer generation request from the answer generation device 4. The generation support device 2 generates output based on the result. The generation support device 2 transmits the generated output to the user terminal 3.
[0030] The user terminal 3 can transmit request data to the generation support device 2. The user terminal 3 can accept input of request data from a user, and transmit an answer generation request based on the request data to the generation support device 2. The user terminal 3 can receive answer data from the generation support device 2. The user terminal 3 can display a screen based on the answer data, which is the result of the answer generation request, on a display unit such as a display. Although only one user terminal 3 is shown in FIG. 1, there may be multiple user terminals.
[0031] The answer generation device 4 includes an answer generation means 40 as a functional component. The answer generation means 40 executes a process of generating answer data for instruction data included in the answer generation request, and transmits the generated answer data to the sender of the answer generation request. In this embodiment, the sender of the answer generation request is the generation assistance device 2, but is not limited to this, and a mode in which the sender is the user terminal 3 can also be adopted.
[0032] In this embodiment, the answer generating means 40 is a natural language processing model (NLP: Natural Language Processing). The answer generating means 40 is preferably a large-scale language model (LLM). The natural language processing model realizes computer processing of data input as a natural language. In this embodiment, the type of natural language processing model adopted is not limited.
[0033] The natural language processing model can output text classification, sentiment analysis, text summarization, question answering, etc., as a result of natural language processing of data input as a natural language. In this embodiment, the answer generating means 40 functions as a chatbot and outputs a processing result for data input as a natural language. Note that data input as a natural language may be in a natural language at the input stage, but may not be in a natural language before or during processing in the natural language processing model. For example, data input as a natural language may be converted into data in a format that can be processed by the natural language processing model and then processed.
[0034] The answer generating means 40 only needs to be able to output an answer as a processing result for an input in natural language, and may be implemented in a form other than a natural language processing model.
[0035] The answer generating means 40 may include image data in the input data. The answer generating means 40 may further have a classification model, for example, and may output a classification result based on the feature amount of the input image data. The answer generating means 40 may output, for example, a subject included in the image data as a classification result based on the feature amount included in the image data.
[0036] The answer generating means 40 can output the processing results for the input of data including natural language and image data. For example, the answer generating means 40 can accept input of questions and processing related to image data in natural language and output the processing results for the same.
[0037] Fig. 2(a) shows a hardware configuration diagram of the generation support device 2. The generation support device 2 includes, as its hardware configuration, a control unit 201, a storage unit 202, and a communication unit 203. In this embodiment, the generation support device 2 may use a computer device such as a server device or a personal computer. Note that the generation support device 2 may be configured by a plurality of computer devices, and is not limited to the configuration shown in Fig. 2(a) as long as it can realize the above-mentioned functional components (21-24) as a whole.
[0038] The control unit 201 is configured with one or more processors such as a CPU, and executes a generation assistance program, an OS (Operating System), and other applications to control the overall processing in the generation assistance device 2. The storage unit 202 is a hard disk drive (HDD), a solid state drive (SSD), a flash memory, a random access memory (RAM), etc., and stores the generation assistance program and various data. The communication unit 203 controls communication with the communication network NW, and realizes data communication with the user terminal 3 and the answer generation device 4.
[0039] The answer generation device 4 may have the same hardware configuration as the generation support device 2. The answer generation device 4 may be implemented in the same computer device as the generation support device 2. Furthermore, the answer generation means 40 may be realized as a functional component of the generation support device 2.
[0040] 2(b) shows a hardware configuration diagram of the user terminal 3. The user terminal 3 includes, as its hardware configuration, a control unit 301, a storage unit 302, a communication unit 303, an input unit 304, and a display unit 305. In this embodiment, the user terminal 3 may be a smartphone, a personal computer, a tablet terminal, or the like.
[0041] The control unit 301 is composed of one or more processors such as a CPU, and controls the overall processing in the user terminal 3 by executing an OS and other applications. The storage unit 302 is an HDD, SSD, flash memory, RAM, etc., and stores a browser application and various data. The communication unit 303 controls communication with the communication network NW, and realizes data communication at least with the generation support device 2. The input unit 304 is an input interface that accepts input operations by the user, and is composed of a touch panel, a mouse, a keyboard, etc. The display unit 305 is composed of a display that displays the processing results by the control unit 301, etc.
[0042] In this embodiment, the user terminal 3 accesses the answer generation support system 1 via a web browser, but is not limited to this. For example, the user terminal 3 may store a terminal generation support program in the storage unit 302 and execute the program by the control unit 301 to start an application and access the answer generation support system 1. The user terminal 3 may also be configured to realize a part of the functional configuration (21-24) of the generation support device 2.
[0043] FIG. 3 illustrates a processing sequence in the answer generation support system 1.
[0044] First, the user terminal 3 transmits request data to the generation support device 2 (S110). The generation support device 2 generates instruction data based on the request data, and transmits an answer generation request including the instruction data to the answer generation device 4 (S121). The answer generation device 4 generates answer data according to the instruction data included in the received answer generation request, and transmits the answer data to the generation support device 2 (S122). The generation support device 2 generates an output based on the received answer data, and transmits the output to the user terminal 3 (S130).
[0045] Step S120, which is an exchange between the generation support device 2 and the answer generation device 4, can be executed multiple times with step S121 and step S122 as one set. In step S120, the generation support device 2 transmits multiple answer generation requests, and the answer generation device 4 generates multiple answer data corresponding to the multiple answer generation requests and transmits it to the generation support device 2. In step S130, the generation support device 2 generates one output based on the multiple answer data and transmits it to the user terminal 3.
[0046] The requested data is input data for obtaining an answer desired by a user. In this embodiment, the requested data is provided with at least one or more predetermined items according to the category. The category indicates the type of answer desired by the user, and includes, for example, article writing, script writing, story writing, answer writing, name writing such as catchphrase, song lyrics writing, email writing, posting writing, and rules and regulations writing. The categories are not limited to these.
[0047] For example, in the case of article creation, categories can be set according to the specific type of article, such as blog article, newspaper article, news article, etc. Also, in the case of blog articles, categories can be set according to the specific content of the blog, such as blogs about health, blogs about food, blogs about animals, etc. The granularity of categories can be set in a wide range, from broad to narrow.
[0048] 4(a) shows an example of the configuration of the request data. The request data corresponds to the data set by the administrator and is composed of data input by the user.
[0049] The data set by the administrator includes a category ID, which is identification information of the category of the output to be generated, and categorical items corresponding to the category. In Fig. 4(a), the category of the request data is "clinic review response", the user is a clinic staff member, etc., and the data is used to create a response message to a review posted on the homepage of the clinic, a review site, etc.
[0050] The categorized items include at least one or more basic items that are arbitrarily set for each category. In the example of Fig. 4(a), the categorized items include reviews, clinic names, clinic regions, past reviews, etc. The categorized items can be arbitrarily set by the system administrator.
[0051] The category items may further include example sentence data. The example sentence data presents sentences created by the user in the past as example sentences. By inputting the example sentence data, the answer generating means 40 can generate an answer that reflects the characteristics and trends of the example sentences. In the example of FIG. 4(a), "past word-of-mouth responses" corresponds to the example sentence data.
[0052] FIG. 4(b) shows an example of a screen display of a request input screen W10 displayed on the user terminal 3 and for inputting request data by a user. The request input screen W10 includes an icon W11, a basic item input section W12, an example sentence input section W13, and a send button W14. The icon W11 displays an icon such as a character set corresponding to a category. It is preferable that the icon W11 is displayed differently for each category so that the category can be distinguished at a glance. The basic item input section W12 and the example sentence input section W13 are provided according to the number and contents of items for each category.
[0053] The basic item input section W12 accepts input of basic items as text data, pictograms, image data, etc. The example sentence input section W13 accepts input of example sentence data as text data, pictograms, image data, etc. When the send button W14 is pressed, it transmits the data input in the basic item input section W12 and the example sentence input section W13 to the generation support device 2 as request data. The transmitted request data may be accompanied by the ID of the user who sent it, etc.
[0054] FIG. 5 illustrates a process flowchart of the generation support device 2.
[0055] The request obtaining means 21 obtains request data transmitted from the user terminal 3 (S201). The request obtaining means 21 stores the obtained request data in the storage unit 202.
[0056] The first generating means 22 obtains a template corresponding to a category included in the obtained request data (S202). A template is data from which instruction data is generated, and is configured so that basic items and example sentence data of the request data can be input in predetermined positions. The templates are stored in the storage unit 202 by category.
[0057] The first generating means 22 generates instruction data based on the acquired desired data and the template (S203). Specifically, the first generating means 22 generates the instruction data by inputting the desired data into a predetermined portion of the template.
[0058] An example of a template in the "Clinic review response" category is shown in the following paragraph. Here, the template is composed of text data in natural language. Variables in the template such as "%clinic%", "%area%", "%sankou%", and "%kutikomi%" correspond to "Clinic name", "Clinic area", "Past review responses", and "Reviews" in the request data entered by the user, respectively. These variables will be substituted with each item of the request data entered by the user.
[0059] A review has been posted for the clinic. Write a response thanking them for their review and encouraging them to visit again. Please output text in natural language. #Review information Reviewer name: 〇〇 #Clinic information Clinic name: %clinic% Clinic area:%area% Review reply reference:%sankou% The first request is: Posted reviews:%kutikomi% <Replacement rules> Please enter your review: %kutikomi% Clinic Name: %clinic% Clinic area name: %area% Enter a review reply that you have actually replied to: %sankou%
[0060] 6(a) shows an example of the data structure of template data. The template data has a category ID and a template, and a division rule corresponding to the category is defined. The template data may further have a template ID to enable identification of the template.
[0061] The division rule includes a rule regarding the number of times an answer generation request is divided. The amount of data and the number of characters that can be transmitted at one time to the answer generation means 40 is limited. In this embodiment, a predetermined number of divisions is defined in advance, since it is difficult to obtain all answers by one answer generation request, particularly for categories that require long sentence generation. In other words, the division rule is a rule that limits the number of characters included in the answer generation request to a predetermined number or less. The instruction means 23 transmits a plurality of answer generation requests divided according to the number of divisions to the answer generation means 40. Note that, in the case of short sentence generation, the number of divisions may be 0, and the desired answer can be obtained from the answer generation means 40 by transmitting a single answer generation request without dividing the answer generation request.
[0062] The division rules include rules regarding the common area and divided area of the instruction data to be generated. The common area is an area common to multiple pieces of instruction data to be divided and transmitted. Here, it is preferable that the common area shows the whole picture of the answer to be generated. The common area has the role of inputting the whole picture to the answer generating means 40, thereby making it possible to recognize the steps and stages of answer generation, and improving the accuracy of the answer.
[0063] The divided areas are different areas in the multiple pieces of instruction data that are divided and transmitted. The divided areas have the role of controlling the answers given by the answer generating means 40 to the respective answer generation requests.
[0064] The divided regions can be regions obtained by dividing the common region a predetermined number of times. Specifically, the divided regions are preferably regions obtained by dividing the overall image included in the common region into partial images. For example, the generation of a "story" is an example of long text generation. In the common region, the titles of chapters 1 to 12 are set as the overall image of the story. The divided regions are divided three times, for example, into these partial images such as "Chapters 1 to 4," "Chapters 5 to 8," and "Chapters 9 to 12."
[0065] 6(b) shows an example of a template setting screen W20 on which template data is registered, edited, etc. The template setting screen W20 is displayed on an administrator's terminal device (not shown). The template setting screen W20 includes a category setting section W21, a desired data setting section W22, a template setting section W23, a divided area setting section W24, a restriction setting section W25, and a send button W26.
[0066] The category setting section W21 receives setting inputs such as a category name, an icon, a description of the category, and a major category indicating a major classification of the category.
[0067] The desired data setting unit W22 accepts setting inputs related to replacement rules for desired data. The replacement rules accept setting inputs of definitions of variables and the like that reflect desired data input by a user in a template.
[0068] The template setting unit W23 receives setting input related to the body of the template. The body of the template corresponds to a common area in the instruction data. In this embodiment, the template setting unit W23 is set so that the body includes an overall configuration showing an overall picture and constraint conditions.
[0069] The divided area setting unit W24 receives setting input related to the divided areas. In the illustrated example, two divided areas, a first divided area A and a second divided area B, are set and input, but the number of divided areas (number of divisions) can be added or deleted as desired. In this embodiment, the number of divisions is determined by the number of divided areas, but a divided area may be added by setting and inputting the number of divisions.
[0070] The restriction setting unit W25 accepts the setting input of restriction information including a disclosure status indicating whether the data is disclosed to the user or not, a maximum token number which is an upper limit on the number of tokens, and a temperature which is a parameter for adjusting words used in the answer by the answer generating means 40. The maximum token number is used to send a notification to the user terminal 3 to reduce the number of characters, etc., when the request data input by the user exceeds the upper limit on the number of tokens.
[0071] Fig. 7 shows an example of the command data generated in step S203. Fig. 7 shows examples of command data I1 to I7 in the category of "novel creation." The command data I1 to I7 indicate six divided pieces of command data.
[0072] The instruction data I1 has a common area I11 and a divided area I12. Data is substituted into the variable I13 based on the request data. The variable is not limited to the common area I11, and may be set in the divided area I12.
[0073] The common area I11 has a premise I111, a constraint condition I112, and an overall configuration I113. The premise I111 is an area that instructs the answer generating means 40 on assumptions such as a position, an answer format, and a processing method. The constraint condition I112 is an area that instructs constraints to be given to the answer of the answer generating means 40. Here, the constraints can be constraints that must be legally observed in the answer, constraints for improving the accuracy of the answer, constraints for making the content of the answer interesting, constraints on the character of the answer generating means 40 (tone of voice, personality, person, occupation, etc.), constraints on parameter settings for the answer generating means 40 (degree of emotion, strength of character), and the like. The overall configuration I113 is an area that instructs the overall picture of a scenario, etc., particularly in generating long sentences.
[0074] The divided area I12 is an area indicating a partial configuration of the overall configuration I113. The divided area I12 is an area to which one answer data generated by the answer generating means 40 corresponds. The instruction data I1 to I7 are generated so that the contents of the divided area I12 are different partial configurations of the overall configuration I113.
[0075] 5 again, the instruction means 23 transmits an answer generation request to the answer generating means 40 (S204). The instruction means 23 transmits an answer generation request for the number of divisions to the answer generating means 40 according to the number of divisions defined in the division rule. Here, when the number of divisions is one or more, the instruction means 23 divides the instruction data into different division areas and transmits the answer generation request to the answer generating means 40.
[0076] The instruction means 23 acquires answer data corresponding to the answer generation request from the answer generating means 40 (step S205). The instruction means 23 stores the acquired answer data in the memory unit 202 in association with the answer generation request. The answer generation request includes one piece of instruction data. In this embodiment, the instruction data refers to data sent to the answer generating means 40 regardless of whether it is divided or not. When the instruction data is divided, it becomes a plurality of data with different divided areas.
[0077] The instruction means 23 refers to the division rule and judges whether all answer generation requests for the instruction data divided a predetermined number of times have been completed (step S206). If it is judged that the requests have been completed (YES in S206), the instruction means 23 outputs a processing command to execute the following S207. If it is judged that the requests have not been completed (NO in S206), the instruction means 23 returns to S204 in accordance with the division rule and transmits an answer generation request including instruction data having a divided area following the answer generation request transmitted in the previous S204. Note that when transmitting an answer generation request including a plurality of instruction data, the order in which each of the instruction data having a different divided area is transmitted is not limited.
[0078] The second generating means 24 generates an output based on one or more pieces of answer data acquired by the answer generating means 40 (step S207). When multiple pieces of answer data have been acquired, the second generating means 24 generates an output based on multiple pieces of answer data included in multiple dialogue data consisting of one piece of instruction data and answer data corresponding to the instruction data. Here, the output may be one piece of answer data, or may be generated by extracting one or more predetermined areas included in the answer data. Also, if the output is multiple pieces of answer data, the output may be a combination of the multiple pieces of answer data, or may be generated by extracting and combining one or more predetermined areas included in each of the multiple pieces of answer data.
[0079] The second generating means 24 completes the process by transmitting the generated output to the user terminal 3 which is the sender of the request data. The user terminal 3 can display the answer result to the input request data on the display unit 305 based on the received output. The user terminal 3 may output the answer result as voice based on the received output. The output is not particularly limited as long as the user can recognize the answer result. This allows the user to obtain an accurate and desired answer result by relatively easily inputting the request data.
[0080] The answer generation means 40 can generate answer data in response to an input of an answer generation request and respond to the requester, but the input data and answer data are limited in the number of characters by tokens. Such a character limit is set due to the structure of the natural language processing model of the answer generation means 40. It is also assumed that the character limit is set to alleviate an increase in processing load and a delay in response caused by a large number of requests being concentrated at the same time on the answer generation means 40. It is also considered that setting a character limit for one input is effective in terms of improving the accuracy of the answer by limiting the scope of questions, requests, appeals, etc. included in one answer generation request. Such a character limit has been a factor in preventing a user from obtaining a desired answer, since it requires ingenuity in the input contents and transmission procedure. The present invention can provide answer generation support for such a factor by providing a means for dividing and transmitting an answer generation request, particularly for a long answer.
[0081] <Embodiment 2> In the second embodiment, the generation support device 2 further includes an answer generating means 40. In the second embodiment, the instruction means 23 inputs instruction data to the answer generating means 40, and obtains answer data from the answer generating means 40 as an output.
[0082] <Embodiment 3> In the third embodiment, the user terminal 3 includes a first generating means 22, an instruction means 23, and a second generating means 24. The user generates instruction data based on request data via the user terminal 3. The user divides the instruction data a predetermined number of times and transmits it to the answer generating means 40 via the user terminal 3. The user obtains answer data for each divided piece of instruction data from the answer generating means 40 via the user terminal 3.
[0083] In the third embodiment, the user terminal 3 may further include a request acquisition means 21. The request acquisition means 21 acquires request data input via the user terminal 3. The first generation means 22 generates instruction data based on the request data. The instruction means 23 divides the instruction data a predetermined number of times and transmits it to the answer generation means 40. The second generation means 24 acquires multiple pieces of answer data from the answer generation means 40 and generates an output based on the answer data. The user terminal 3 outputs the generated output via an output interface (such as a display unit 305 or a speaker). [Explanation of symbols]
[0084] 1 Answer Generation Support System 2 Generation support device 21. Means of obtaining requests 22 First generation means 23 Instruction means 24 Second generation means 3. User terminal 4 Answer generator 40 Answer generation means NW Communication network
Claims
1. A request acquisition means for acquiring request data; a first generating means for generating instruction data based on the request data; instruction means for dividing the instruction data into a predetermined number of parts and transmitting the divided parts to the answer generating means; and a second generating means for generating an output based on the plurality of answer data acquired by the answer generating means.
2. The answer generation support system according to claim 1 , wherein the answer generation means is a large-scale language model.
3. the desire data includes a category of output to be generated; 3. The answer generation support system according to claim 1, wherein the first generating means obtains a template corresponding to the category, and generates the instruction data based on the template.
4. The request data further includes example sentence data according to a category, The answer generation support system according to claim 3 , wherein the first generation means generates the instruction data by inputting the example sentence data into a predetermined portion of the template.
5. The template defines a division rule corresponding to the category, 4. The answer generation support system according to claim 3, wherein the instruction means transmits an answer generation request to the answer generating means in accordance with a number of divisions defined in the division rule.
6. 6. The answer generation support system according to claim 5, wherein the division rule is a rule that limits the number of characters included in the answer generation request to a predetermined number or less.
7. the instruction data has a common area and divided areas obtained by dividing the common area a predetermined number of times; 3. The answer generation support system according to claim 1, wherein the instruction means divides the instruction data into different divided areas and transmits the divided areas to the answer generating means.
8. 3. The answer generation support system according to claim 1, wherein the second generation means generates the output based on a plurality of answer data included in a plurality of dialogue data each including one of the instruction data and one of the answer data.
9. generating instruction data based on the request data; a step of dividing the instruction data into a predetermined number of parts and transmitting the divided parts to a response generating means; and acquiring a plurality of pieces of answer data from the answer generating means.
10. accepting input of the request data; The answer generation support method according to claim 9 , further comprising the step of: generating an output based on the plurality of answer data in the obtaining step, the output being executed by the computer.
11. A request acquisition means for acquiring request data; a first generating means for generating instruction data based on the request data; instruction means for dividing the instruction data into a predetermined number of parts and transmitting the divided parts to the answer generating means; a second generating means for generating an output based on the plurality of answer data acquired from the answer generating means;
Citation Information
Cited By
Dialogue systems, dialogue methods, and dialogue programs
JP2026161580A
Message processing method and apparatus, device, and storage medium
WO2026102728A1
