Inquiry response device, inquiry response system, inquiry response program, and inquiry response method

Abstract

We provide technology that improves the accuracy and response speed of answers related to inquiries. [Solution] An inquiry response device comprising: a question response unit that receives the text of an inquiry; a term conversion unit that corrects the terms contained in the text using a term dictionary stored in a memory unit; a RAG data acquisition unit that extracts data related to the text of the inquiry as RAG data from a predetermined RAG database; and an answer request unit that transmits the extracted RAG data and instructs a predetermined generation AI service to provide an answer to the inquiry.

Description

[Technical Field]

[0001] The present invention relates to an inquiry response device, an inquiry response system, an inquiry response program, and an inquiry response method. [Background Art]

[0002] Technologies related to generative AI are making remarkable progress.

[0003] For example, the idea support method disclosed in Patent Document 1 is executed in a system including a generative AI server and a terminal connected to the generative AI server via a network. The terminal executes a step of transmitting an instruction to generate element groups of two or more categories for a predetermined product to the generative AI server, and the generative AI server executes a step of generating element groups of two or more categories for the predetermined product and transmitting the element groups of the two or more categories to the terminal. Further, the terminal executes a step of transmitting an instruction to generate a product concept combining element groups of two or more categories to the generative AI server, and the generative AI server executes a step of generating a product concept combining element groups of two or more categories and transmitting the product concept to the terminal. [Prior Art Documents] [Patent Documents]

[0004] [Patent Document 1] Japanese Patent No. 7520336 [Summary of the Invention] [Problems to be Solved by the Invention]

[0005] Conventional generative AI generates responses to instructions called prompts based on learned data. That is, when information including notations different from those used in the learned data is included, or when accurate information is not included for terms with similar but different notations, it is impossible to generate highly accurate responses, and therefore often generates inaccurate responses.

[0006] Therefore, a technique called RAG (Retrieval Augmented Generation), in which a generative AI generates answers based on a specific external database, is developing. To enable the generative AI to efficiently generate answers, it is desirable to enrich the database used in RAG and improve its responsiveness. Note that the technology described in Patent Document 1 does not assume the use of RAG.

[0007] The object of the present invention has been made in view of the above points, and aims to provide a technology that improves the accuracy and response speed of answers related to inquiries. [Means for solving the problem]

[0008] The present invention includes several means for solving at least some of the above problems, but an example is as follows: An inquiry response device according to one aspect of the present invention is characterized by comprising: an inquiry response unit that receives the text of an inquiry; a term conversion unit that corrects the terms contained in the text using a term dictionary stored in a memory unit; a RAG data acquisition unit that extracts data related to the text of the inquiry as RAG data from a predetermined RAG database; and an answer request unit that transmits the extracted RAG data and instructs a predetermined generation AI service to provide an answer to the inquiry.

[0009] Furthermore, the above-mentioned inquiry response device may include a notation variation correction unit that corrects any grammatical inconsistencies, general knowledge level inconsistencies, or domain-specific inconsistencies among the terms contained in the text.

[0010] Furthermore, in the above-described inquiry response device, the term conversion unit may receive the text along with the term dictionary and instruct the generation AI service to correct the terms contained in the text.

[0011] Furthermore, in the above-described inquiry response device, the terminology dictionary may define a corrected unified term for each proper noun used in a given organization, and the term conversion unit may convert the proper nouns contained in the text into the unified term.

[0012] Furthermore, in the above-described inquiry response device, the term dictionary may define a corrected unified term for each proper noun used in a given organization, and the term conversion unit may instruct the generation AI service to convert the proper nouns contained in the text into the unified term.

[0013] Furthermore, in the above-described inquiry response device, the term conversion unit may instruct the generation AI service to convert the proper nouns contained in the text into the unified phrases, and may also instruct the generation AI service to update the term dictionary if there are proper nouns that are not included in the term dictionary.

[0014] Furthermore, in the above-described inquiry response device, the terminology dictionary may define a corrected unified term and its explanatory text for each proper noun used in a given organization, and the term conversion unit may convert the proper nouns contained in the text into the unified term and add the explanatory text.

[0015] Furthermore, in the above-described inquiry response device, the term dictionary defines a corrected unified term and its explanatory text for each proper noun used in a given organization, and the term conversion unit may instruct the generation AI service to convert the proper nouns contained in the text into the unified term and to add the explanatory text.

[0016] Furthermore, another aspect of the present invention relates to an inquiry response system comprising an operator terminal and an inquiry response device, wherein the operator terminal performs a question processing step of receiving the input of an inquiry text and passing the text to the inquiry response device, and the inquiry response device performs a question response step of receiving the inquiry text, a term conversion step of correcting terms contained in the text using a term dictionary stored in a memory unit, a RAG data acquisition step of extracting data related to the inquiry text as RAG data from a predetermined RAG database, and a response request step of transmitting the extracted RAG data and instructing a predetermined generation AI service to provide an answer to the inquiry.

[0017] Furthermore, another aspect of the present invention relates to an inquiry response program, which provides an information processing device with an inquiry response function, wherein the information processing device comprises a storage unit for storing a term dictionary and a processing unit, and the processing unit is instructed to perform a question response step of receiving an inquiry text, a term conversion step of correcting terms contained in the text using the term dictionary stored in the storage unit, a RAG data acquisition step of extracting data related to the inquiry text as RAG data from a predetermined RAG database, and a response request step of transmitting the extracted RAG data and instructing a predetermined generation AI service to provide an answer to the inquiry.

[0018] Also, an inquiry response method according to another aspect of the present invention is an inquiry response method for causing an information processing apparatus to realize an inquiry response function. The information processing apparatus includes a storage unit that stores a term dictionary and a processing unit. The processing unit performs a question and answer step of receiving an inquiry text, a term conversion step of correcting terms included in the text using the term dictionary stored in the storage unit, a RAG data acquisition step of extracting data related to the inquiry text as RAG data from a predetermined RAG database, and a response request step of transmitting the extracted RAG data and instructing a predetermined generation AI service to generate an answer to the inquiry.

Advantages of the Invention

[0019] According to the present invention, the accuracy and response speed of answers related to inquiries can be improved.

[0020] Problems, configurations, and effects other than those described above will be clarified by the description of the following embodiments.

Brief Description of the Drawings

[0021] [Figure 1] It is an example of a block diagram of an inquiry response system according to an embodiment. [Figure 2] It is a diagram showing an example of the data structure of a term dictionary. [Figure 3] It is a diagram showing an example of the hardware configuration of an inquiry response apparatus. ​​​​​​​​​​​​​​​​[Modes for carrying out the invention]

[0022] Below, an inquiry response system 1 to which an embodiment according to one aspect of the present invention is applied will be described with reference to the drawings.

[0023] Generally, in RAG (Retrieval Augmented Generation), it is desirable to enrich the database used and improve responsiveness. However, when the amount of information in the database becomes excessive, these two aspects tend to be contradictory. For example, if all irrelevant information is included in the search, unnecessary matching processes increase, slowing down the response speed and potentially leading to the acquisition of incorrect information that constitutes noise. Furthermore, if the database contains information with different notation than that used in the trained data, or if accurate information on similar but distinct terms is not included, the response speed may also decrease.

[0024] By using the inquiry response system 1 according to this embodiment, it becomes possible to unify the database used as the RAG and the terminology used in queries, enabling accurate and rapid narrowing of the information to be searched. As a result, it is possible to improve response speed and suppress the occurrence of hallucination by reducing noise information, which is beneficial.

[0025] Figure 1 is an example of a block diagram of the inquiry response system according to this embodiment. The inquiry response system 1 includes an inquiry response device 100, a generation AI service 200, an operator terminal 300, and a RAG database 400.

[0026] The operator terminal 300 is a terminal owned and used by users of the inquiry response system 1, such as a company's customer service representative. The operator terminal 300 is connected to the inquiry response device 100 via the network 50 so as to be able to communicate.

[0027] Network 50 is a communication network such as a LAN (Local Area Network), WAN (Wide Area Network), the Internet, or a mobile phone network. Network 50 may also be a VPN (Virtual Private Network) on a wireless communication network such as a mobile phone network.

[0028] The inquiry response device 100 is a so-called server device. However, it is not limited to this, and the inquiry response device 100 may be any kind of information processing device such as a personal computer, smartphone, workstation, PDA (Personal Data Assistant), or tablet device.

[0029] The inquiry response device 100 receives inquiry questions from the operator terminal 300 via the network 50, searches and extracts information from the RAG database 400, which is a database used as RAG appropriately according to the content of the question, and passes it to the generation AI service 200 to instruct it to provide an answer, thereby realizing the inquiry response function. Alternatively, the inquiry response device 100 may obtain an answer from the generation AI service 200 and send the answer to the operator terminal 300 via the network 50.

[0030] When the operator terminal 300 receives a predetermined response from the inquiry response device 100 or the generation AI service 200 via the network 50, it performs output processing. The operator terminal 300 also transmits input information received from input devices, etc., provided in its own device (described later), to the inquiry response device 100 via the network 50 as inquiry question text (query).

[0031] The operator terminal 300 comprises a processing unit 310, an input receiving unit 320, a display unit 330, and a communication unit 340. The processing unit 310 includes a question processing unit 311.

[0032] The question processing unit 311 receives the text of the inquiry via the input receiving unit 320 and passes the text to the inquiry response device 100. When the question processing unit 311 receives a predetermined answer from the inquiry response device 100 or the generation AI service 200, it outputs it via the display unit 330.

[0033] The input reception unit 320 receives input instructions from the operator of the operator terminal 300, for example, a counter staff member, and transmits them to the inquiry response device 100. The display unit 330 receives screen information in a predetermined format instructed to be displayed by the inquiry response device 100 or the generation AI service 200 from the communication unit 340 and draws it on the display device. The communication unit 340 communicates with the inquiry response device 100 and the generation AI service 200 via the network 50.

[0034] The Generative AI Service 200 is a service that provides the functions of so-called generative AI, such as GPT and Gemini, via an API (Application Programming Interface). The Generative AI Service 200 provides commands (prompts) in natural language to the Generative AI and causes it to generate the desired result. The Generative AI Service 200 comprises an AI processing unit 210 and a communication unit 220.

[0035] In this embodiment, when the generation AI service 200 receives instructions, for example via an API, it causes the generation AI to generate response information and sends the result to the source of the instructions as a return value of the API. At that time, the generation AI receives the relevant information necessary for the response as RAG information, reads the RAG information as part of the information to be added and considered, and reflects it in the generation of the response.

[0036] For example, the generation AI service 200 may refer to a log of questions and answers about a particular product and generate answers to similar questions. Alternatively, the generation AI service 200 may obtain similar information at runtime by searching other devices via the internet or the like.

[0037] The inquiry response device 100 includes a processing unit 110, a storage unit 120, and a communication unit 130. The processing unit 110 includes a question answering unit 111, a notation variation correction unit 112, a term conversion unit 113, a RAG data acquisition unit 114, and a response request unit 115. The storage unit 120 includes a term dictionary 121.

[0038] Figure 2 shows an example of the data structure of a terminology dictionary. The terminology dictionary 121 stores the registration date and time 121a, the dictionary term (Terminology) 121b, and the uniform name 121c in association with each other. The terminology dictionary 121 defines a corrected uniform name for each proper noun used in a given organization, but is not limited to this; it may define uniform names for all parts of speech, such as common nouns, verbs, adjectives, and adjectival nouns.

[0039] The registration date and time 121a is information that identifies the date and time when the correspondence between the dictionary term 121b and the unified term 121c was registered. The dictionary term 121b is information that shows one form of a term with variations in spelling. The unified term 121c is the spelling of the term that will be standardized for terms with variations in spelling. For example, suppose "HISYS" is registered in the dictionary term 121b and is registered in the unified term 121c as "Hitachi Systems". In that case, if the query text contains the community-specific term "HISYS", the term "HISYS" will be converted to the unified term "Hitachi Systems" for processing, and the degree of match with the unified term "Hitachi Systems" included in the RAG database 400 will increase, making it more likely that appropriate information will be obtained.

[0040] Returning to the explanation of Figure 1, the question answering unit 111 receives the text of the inquiry.

[0041] The notation variation correction unit 112 corrects any grammatical inconsistencies, general knowledge level inconsistencies, or inconsistencies in proprietary domain names, or any combination thereof, among the terms included in the query text.

[0042] In this embodiment, grammatical variations in spelling refer to variations that occur regardless of the context in which they are used. For example, variations such as (speed up, speedup, speed-up) are included.

[0043] Furthermore, in this embodiment, variations in notation at the general knowledge level refer to variations in notation that occur in general terms and their abbreviations. For example, variations such as (Ministry of Economy, Trade and Industry, METI) are examples of such variations.

[0044] Furthermore, in this embodiment, grammatical inconsistencies refer to inconsistencies in terminology and abbreviations used within a given domain (e.g., industry). For example, inconsistencies such as (GBDT, Gradient Boosting Decision Tree).

[0045] The term conversion unit 113 corrects the terms included in the query text using the term dictionary 121 stored in the memory unit 120.

[0046] In this embodiment, the dictionary terms and unified terms defined in the terminology dictionary 121 are variations in spelling that occur within a given community, organization, etc. For example, these include variations in proper nouns such as organization names, product names, and project names, as well as variations in internal company terminology. For example, it is intended to unify variations in spelling such as HISYS and Hitachi Systems into one of the terms. The term conversion unit 113 then converts terms such as proper nouns contained in the inquiry text into unified terms defined in the terminology dictionary.

[0047] Furthermore, if the term dictionary 121 has additional explanatory text associated with each unified term, the term conversion unit 113 may convert proper nouns included in the inquiry text into unified terms and add explanatory text for the unified terms.

[0048] The RAG data acquisition unit 114 extracts data related to the query text as RAG data from a predetermined RAG database 400.

[0049] The response request unit 115 transmits the extracted RAG data to the generation AI service 200 and instructs the generation AI service 200 to provide a response to the inquiry.

[0050] The communication unit 130 communicates with other devices, namely the generation AI service 200, the operator terminal 300, and the RAG database 400, via the network 50.

[0051] Figure 3 shows an example of the hardware configuration of an inquiry response device. The inquiry response device 100 comprises a processor 101, memory 102, storage 103, communication device 104, and a bus 105 connecting the devices. In addition, the inquiry response device 100 may also include an input device.

[0052] The processor 101 is a computing device such as a CPU (Central Processing Unit) or GPU (Graphics Processing Unit), and it performs processing according to a program recorded in memory 102 or storage 103. In the inquiry response device 100, processing is performed by the processor 101, which operates according to a program read from memory 102 or storage 103. The processing unit 110, the question answering unit 111, the notation variation correction unit 112, the term conversion unit 113, the RAG data acquisition unit 114, and the answer request unit 115 each realize their respective functions by having the processor 101 execute a program.

[0053] Memory 102 is a storage device such as RAM (Random Access Memory) or flash memory, and functions as a storage area where programs and data are temporarily read. Storage 103 is a writable and readable storage device. The storage unit 120's functions are realized by either memory 102 or storage 103. Alternatively, the storage unit 120 may be realized by a storage device connected via communication device 104.

[0054] The communication device 104 is an interface for connecting the inquiry response device 100 to an external device. For example, the communication device 104 uses an antenna that can utilize predetermined radio waves (e.g., 5GHz band, 2.4GHz band, etc.) to establish connections with the generated AI service 200, the operator terminal 300, and the RAG database 400 according to the Wi-Fi standard, and performs wireless communication.

[0055] Furthermore, the processing of each component of the inquiry response device 100 may be performed on one piece of hardware or on multiple pieces of hardware. Also, the processing of each component of the inquiry response device 100 may be implemented by one program or by multiple programs.

[0056] The communication unit 130 of the inquiry response device 100 described above is implemented by the communication device 104. The above is an example of the hardware configuration of the inquiry response device 100.

[0057] Each component of the inquiry response device 100 can be further classified into many more components depending on the processing content. Alternatively, each component can be classified to perform even more processing.

[0058] Furthermore, each processing unit (processing unit 110, question answering unit 111, notation variation correction unit 112, terminology conversion unit 113, RAG data acquisition unit 114, response request unit 115) may be constructed using dedicated hardware (ASIC, GPU, etc.) to implement its respective function. Also, the processing of each processing unit may be executed on a single piece of hardware or on multiple pieces of hardware.

[0059] Figure 4 shows an example of the hardware configuration of an operator terminal. The operator terminal 300 includes an input device 301, a processor 302, storage 303, memory 304, a display 305, a communication device 306, and a bus 307 connecting the devices.

[0060] The input device 301 is a variety of input devices such as a keyboard, mouse, or touch panel. The input reception unit 320 of the operator terminal 300 is realized by the input device 301 and the processor 302. The processor 302 is a computing device such as a CPU or GPU, and it executes processing according to a program recorded in the memory 304 or storage 303. In the operator terminal 300, processing is performed by the processor 302, which operates according to a program read from the memory 304 or storage 303. The processing unit 310 and the question processing unit 311 realize their respective functions by having the processor 302 execute a program.

[0061] Memory 304 is a storage device such as RAM (Random Access Memory) or flash memory, and functions as a storage area where programs and data are temporarily read. Storage 303 is a writable and readable storage device.

[0062] The display 305 is a display device such as a liquid crystal display or an organic EL display. The display unit 330 is realized by the display 305 and the processor 302.

[0063] The communication device 306 is an interface for connecting the operator terminal 300 to an external device for communication. For example, the communication device 306 uses an antenna that can utilize predetermined radio waves (e.g., 5GHz band, 2.4GHz band, etc.) to establish a connection with the inquiry response device 100 using the Wi-Fi standard for wireless communication. The communication unit 340 of the operator terminal 300 is realized by the communication device 306 and the processor 302.

[0064] Furthermore, the processing of each component of the operator terminal 300 may be performed on one piece of hardware or on multiple pieces of hardware. Also, the processing of each component of the operator terminal 300 may be implemented by one program or by multiple programs.

[0065] The input device 301, processor 302, storage 303, memory 304, display 305, and communication device 306 are connected to each other by connecting wires such as a bus 307. The above is an example of the hardware configuration of the operator terminal 300.

[0066] Each configuration of the operator terminal 300 can be further classified into many more components depending on the processing content. Alternatively, each component can be classified to perform even more processing.

[0067] Furthermore, each processing unit (processing unit 310, question processing unit 311, input reception unit 320, display unit 330, communication unit 340) may be constructed using dedicated hardware (ASIC, GPU, etc.) to realize its respective function. Also, the processing of each processing unit may be executed on a single piece of hardware or on multiple pieces of hardware.

[0068] The Generative AI Service 200 is a service provided outside of the Inquiry Response System 1, which uses Generative AI. However, it is not limited to this, and it may also use Generative AI maintained by the Inquiry Response System 1. In that case, the Generative AI includes Generative AI that uses Large Language Models (LLMs). The Generative AI is a pre-trained model, such as GPT-4, which has been trained in advance using language data to create a Large Language Model using a neural network (NN) through machine learning or deep learning. The Generative AI is not limited to neural networks (NNs), and other known methods may be used. Furthermore, it is desirable to use a Generative AI that has been tuned to achieve high accuracy in program code generation through techniques such as Few-shot Learning and fine-tuning.

[0069] Alternatively, the response request unit 115 may pass the RAG information, obtained as a result of searching the RAG database 400, to the generation AI service 200.

[0070] Next, the operation of the inquiry response system 1 in this embodiment will be described.

[0071] First, in this embodiment, a user of the inquiry response system 1, such as a company's customer service representative, acts as an operator and uses the operator terminal 300 to input questions from customers or collaborators as text. This becomes the inquiry text (query), and the operator terminal 300 requests a response from the inquiry response device 100. The inquiry is then completed when the generating AI service 200 or the inquiry response device 100 responds to the operator terminal 300 with the answer.

[0072] It should be noted that the types of inquiries made in such inquiry-based work are often biased towards certain types of inquiries depending on the organization. For example, these may include inquiries about product specifications or confirmation of usage rules. In this embodiment, we will explain assuming that the main types of inquiries are questions and answers, inquiries about design information, and inquiries about regulations, but this is not limited to these, and may differ depending on the organization to which it applies.

[0073] Figure 5 shows an example of the terminology correction response processing flow. The terminology correction response processing flow starts when the operator terminal 300 receives the inquiry text from the operator.

[0074] First, the question processing unit 311 of the operator terminal 300 receives the inquiry text and transmits the inquiry text to the inquiry response device 100 (step S001).

[0075] Then, the question answering unit 111 of the inquiry response device 100 receives the inquiry text. The notation variation correction unit 112 corrects the grammatical-level notation variations in the inquiry text (step S002).

[0076] Then, the notation variation correction unit 112 corrects inconsistencies in the general knowledge level of the inquiry text (step S003).

[0077] Then, the notation variation correction unit 112 corrects the notation variation at the unique domain level of the inquiry text (step S004).

[0078] In steps S002 to S004, the notation variation correction unit 112 performs text similarity determination using, for example, a general edit distance to correct the notation variation. As a method for correcting Japanese-specific notation variations, the notation variation correction unit 112 absorbs differences in hiragana / katakana / kanji, etc., or full-width / half-width by converting them to Roman characters. For example, for the term "スイカ" (suika), there can be Chinese character notation ("西瓜"), Japanese full-width katakana, and Japanese half-width katakana notations. On the other hand, the similarity distances among the three can be made almost zero by converting them to the notation "suika" after converting to Roman characters. The notation variation correction unit 112 uses this to correct the notation variation. In this embodiment, the inquiry text thus corrected for notation variation may be referred to as the corrected inquiry text.

[0079] Then, the term conversion unit 113 detects the notation variation in the community of the inquiry text using the term dictionary 121 and performs term conversion (step S005).

[0080] Specifically, the term conversion unit 113 analyzes the inquiry text to detect the dictionary phrases 121b included in the term dictionary 121. The term conversion unit 113 identifies the unified phrase 121c associated with the detected dictionary phrase 121b. Then, the term conversion unit 113 corrects by replacing the terms included in the inquiry text with the identified unified phrase.

[0081] In the replacement with the unified phrase, when the unified phrase should have conjugations applied (for example, in the case of verbs, adjectives, adjectival nouns, etc.), the term conversion unit 113 converts to the appropriate conjugated form and replaces. In this process, the term conversion unit 113 uses a language model such as an LLM to apply the appropriate conjugated form. In this embodiment, the inquiry text thus term-converted may be referred to as the inquiry text in unified expression.

[0082] Furthermore, if explanatory text is associated with each unified term in the term dictionary 121, the term conversion unit 113 converts the proper nouns included in the inquiry text into unified terms and adds explanatory text for the unified terms.

[0083] Then, the RAG data acquisition unit 114 extracts RAG data from the RAG database 400 using a unified query text (step S006). Specifically, the RAG data acquisition unit 114 extracts data associated with the unified query text (for example, question and answer history, etc.) from the RAG database 400.

[0084] Then, the response request unit 115 transmits the extracted RAG data and instructs the generation AI service 200 to provide a response to the unified expression inquiry text (step S007). Specifically, the response request unit 115 transmits the unified expression inquiry text and the RAG data extracted in step S006 to the generation AI service 200 and instructs the generation AI service 200 to provide a response to the inquiry (unified expression inquiry text).

[0085] Then, the AI ​​processing unit 210 generates an answer to the inquiry (step S008). Specifically, the AI ​​processing unit 210 learns from the RAG data (for example, through fusion learning) and generates an answer to the inquiry in natural language. The criteria for generating this answer are similar to the own criteria of the generation AI service 200, but generally, the range described in the RAG data is given priority and the information within that range is used. For example, if a question and answer history is provided as RAG data, the AI ​​processing unit 210 identifies questions similar to the inquiry text from the question and answer history and generates an answer that is similar in content. Alternatively, the answer request unit 115 may instruct the generation AI service 200 in advance on the answer criteria.

[0086] The AI ​​processing unit 210 then transmits the response data to the inquiry to the operator terminal 300 (step S009). The display unit 330 of the operator terminal 300 then displays the response data on the display (step S010).

[0087] The above is an example of the terminology correction response processing flow. According to this example of the terminology correction response processing flow, variations in notation can be absorbed and unified expressions can be used, that is, the same term can be used for terms of the same concept, thereby improving the accuracy and response speed of answers related to inquiries. Alternatively, the response request unit 115 may receive the text that will become the response generated in step S008, and then revert the unified terms used back to their notation in the inquiry text before sending it to the operator terminal 300.

[0088] In the terminology correction response process described above, the notation variation correction unit 112 corrects notation variations in the inquiry text in steps S002 to S004. However, this is not the only method; the notation variation correction unit 112 may also instruct the generation AI service 200 to correct notation variations in the inquiry text. Such modified examples are explained below with reference to Figure 6.

[0089] Figure 6 shows an example of the terminology correction response process for another example. The terminology correction response process for this other example is basically the same as the terminology correction response process described above, but differs in that it uses a generation AI to correct variations in spelling. The differences will be explained below.

[0090] In the terminology correction response process for another example, steps S101 and S102 are executed instead of steps S002 to S004 in the example flow of the terminology correction response process.

[0091] The notation variation correction unit 112 instructs the AI ​​service 200 to correct grammatical-level notation variations, general knowledge-level notation variations, and unique domain notation variations in the inquiry text (step S101). Specifically, the notation variation correction unit 112 sends the inquiry text to the generation AI service 200, generates a prompt instructing the generation AI service 200 to correct grammatical-level notation variations, general knowledge-level notation variations, and unique domain notation variations and return the corrected text.

[0092] Then, the AI ​​processing unit 210 generates the corrected inquiry text and sends it back to the inquiry response device 100 (step S102).

[0093] The above is an example of a terminology correction response processing flow for another example. According to this example of a terminology correction response processing flow for another example, the generating AI can absorb variations in notation and use a unified expression, meaning that the same term can be used for terms of the same concept, thereby improving the accuracy and response speed of answers related to inquiries. In addition, in this case, the inquiry response device 100 does not need to implement LLM, etc., and the hardware resource requirements can be reduced.

[0094] The above describes an inquiry response system 1 to which an embodiment of the present invention is applied. According to this embodiment, it is possible to respond to inquiries with improved accuracy and response speed of answers related to inquiries.

[0095] The present invention is not limited to the embodiments described above. The embodiments described above can be modified in various ways within the scope of the technical idea of ​​the present invention. For example, in the embodiments described above, the term dictionary 121 is given, but it may be updated as needed. Such embodiments will be explained with reference to Figure 7. The inquiry response system according to the second embodiment is basically the same as the inquiry response system 1 described above, but differs in that it updates dictionary terms using a generation AI. The differences will be explained below.

[0096] In the second embodiment, the term conversion unit 113 receives the query text along with the term dictionary 121 and instructs the generation AI service 200 to correct the terms contained in the text. Specifically, the term conversion unit 113 instructs the generation AI service 200 to convert proper nouns contained in the query text into unified phrases. Furthermore, the term conversion unit 113 instructs the generation AI service 200 to ask follow-up questions and update the term dictionary 121 if there are any unknown proper nouns (e.g., those with low confidence scores) that are not included in the term dictionary 121. In addition, if the term dictionary 121 has explanatory text associated with each unified phrase, the term conversion unit 113 may instruct the generation AI service 200 to convert proper nouns contained in the query text into unified phrases and to add explanatory text for the unified phrases.

[0097] Figure 7 shows an example of the terminology correction response process according to the second embodiment. The terminology correction response process according to the second embodiment is basically the same as the terminology correction response process according to the other example described above, but differs in that it uses a generation AI to correct inconsistencies in notation.

[0098] In step S101, instead of the notation variation correction unit 112 instructing the AI ​​processing unit 210 to correct the notation variation, the notation variation correction unit 112 sends the terminology dictionary 121 to the generation AI service 200 and instructs the AI ​​processing unit 210 to also correct the notation variation in the community (step S201). The difference is that in step S102, the AI ​​processing unit 210 generates a unified expression inquiry text and sends it back to the inquiry response device 100 along with the updated terminology dictionary 121 (step S102'), while the processing in steps S202 to S204 is performed.

[0099] Specifically, first, the AI ​​processing unit 210 lists the unknown terms in the inquiry text and sends this list of unknown terms to the operator terminal 300 (step S202). The question processing unit 311 of the operator terminal 300 receives input of unified terms for the unknown terms via the input receiving unit 320 and sends this list of unified terms to the generation AI service 200 (step S203).

[0100] Then, the AI ​​processing unit 210 updates the terminology dictionary 121 using the unknown terms and the standardized terms, respectively (step S204). Specifically, the AI ​​processing unit 210 associates the unknown terms and the standardized terms with dictionary terms 121b and standardized terms 121c, respectively, and stores them in the terminology dictionary 121. Then, in step S102', the AI ​​processing unit 210 returns the updated terminology dictionary to the inquiry response device 100 along with the inquiry text of the standardized expression (with the explanatory text added if an explanatory text is further associated). When the inquiry response device 100 receives the updated terminology dictionary 121, it stores it in the storage unit 120.

[0101] The above is an example of the terminology correction response process flow according to the second embodiment. According to the terminology correction response process according to the second embodiment, when the inquiry text contains a term not included in the term dictionary 121, it is possible to prompt the operator to input a standardized term each time and update the term dictionary 121.

[0102] The above is an example of an inquiry response system according to the second embodiment. According to the inquiry response system according to the second embodiment, the terminology dictionary 121 is easily enriched by updating it each time.

[0103] Furthermore, although the example of the query response system according to the second embodiment does not mention the unified expression of terminology in the RAG database 400, it may be implemented. By unifying the terminology in the RAG database 400, the terminology used between the search target and the query content will be unified, which will reduce omissions and inappropriate matches in searches and improve the search response speed. An example of this will be described below. Needless to say, this is not limited to the query response system according to the second embodiment, but can also be similarly applied to the query response system according to the first embodiment.

[0104] Figure 8 shows an example of the flow of the unified representation RAG data generation process. The flow of the unified representation RAG data generation process starts when the inquiry response device 100 receives the target data to be registered in the RAG database 400 from the system administrator.

[0105] First, the RAG data registration unit (not shown) included in the processing unit 110 of the inquiry response device 100 receives the target data (for example, additional data of the question and answer history, etc.) (step S301).

[0106] Then, the notation variation correction unit 112 instructs the AI ​​processing unit 210 to correct notation variations at the grammatical level, general knowledge level, and proprietary domain level of the target data, as well as notation variations in the community, along with the terminology dictionary 121 (and further, if explanatory text is associated with each unified term in the terminology dictionary 121, to add explanatory text for the unified term) (step S302). Specifically, the notation variation correction unit 112 sends the target data and terminology dictionary 121 to the generation AI service 200, generates a prompt instructing the generation AI service 200 to correct notation variations at the grammatical level, general knowledge level, proprietary domain, and community level, and to add explanatory text for the unified term before returning it.

[0107] The AI ​​processing unit 210 then enumerates the unknown terms in the target data and sends them to the inquiry response device 100 as an unknown term list (step S303). The RAG data registration unit of the inquiry response device 100 accepts input of unified terms for the unknown terms via an input unit (not shown) and sends them to the generation AI service 200 as a unified term list (step S304).

[0108] Then, the AI ​​processing unit 210 updates the term dictionary 121 using the unknown terms and the unified terms, respectively (step S305). Specifically, the AI ​​processing unit 210 associates the unknown terms and the unified terms with the dictionary terms 121b and the unified terms 121c, respectively, and stores them in the term dictionary 121.

[0109] The AI ​​processing unit 210 then generates unified expression target data from the target data (step S306). Specifically, the AI ​​processing unit 210 corrects grammatical-level inconsistencies, general knowledge-level inconsistencies, and inconsistencies in proprietary domains from the target data, generates corrected target data, and then adds terminology conversion to generate unified expression target data. Furthermore, if explanatory text is associated with each unified term in the term dictionary 121, the AI ​​processing unit 210 converts proper nouns included in the target data into unified terms and adds explanatory text for the unified terms. The AI ​​processing unit 210 then returns the updated term dictionary, along with the unified expression target data, to the inquiry response device 100.

[0110] Then, the RAG data registration unit stores the data subject to the unified representation in the RAG database 400 (step S307).

[0111] Then, the RAG data registration unit stores the updated terminology dictionary in the terminology dictionary 121 (step S308).

[0112] The above is an example of the flow of the unified expression RAG data generation process. According to this example of the unified expression RAG data generation process flow, the generation AI can absorb variations in notation and use unified expressions for the RAG data to be registered in the RAG database 400. In other words, it becomes possible to use the same term for terms of the same concept, thereby improving the accuracy and response speed of answers related to inquiries. Furthermore, even in this case, the inquiry response device 100 does not need to implement LLM or the like, and the hardware resource requirements can be reduced.

[0113] Furthermore, the technical elements of the embodiments described above may be applied individually, or they may be divided into multiple parts, such as program components and hardware components, and applied accordingly.

[0114] The present invention has been described above, focusing on its embodiments. [Explanation of symbols]

[0115] 1...Inquiry response system, 50...Network, 100...Inquiry response device, 110...Processing unit, 111...Question answering unit, 112...Notation variation correction unit, 113...Terminology conversion unit, 114...RAG data acquisition unit, 115...Answer request unit, 120...Storage unit, 121...Terminology dictionary, 130...Communication unit, 200...Generating AI service, 210...AI processing unit, 220...Communication unit, 300...Operator terminal, 310...Processing unit, 311...Question processing unit, 320...Input reception unit, 330...Display unit, 340...Communication unit.

Claims

1. A question and answer section that accepts text inquiries, A term conversion unit that corrects terms contained in the text using a term dictionary stored in a memory unit, A RAG data acquisition unit extracts data related to the text of the aforementioned inquiry from a predetermined RAG database as RAG data, A response request unit transmits the extracted RAG data and instructs a predetermined generation AI service to provide an answer to the inquiry, An inquiry response device characterized by comprising the following:

2. An inquiry response device according to claim 1, A notation variation correction unit corrects any grammatical inconsistencies, general knowledge level inconsistencies, or domain-specific inconsistencies among the terms contained in the aforementioned text. An inquiry response device characterized by comprising the following:

3. An inquiry response device according to claim 1, The term conversion unit receives the text along with the term dictionary and instructs the generation AI service to correct the terms contained in the text. A question response device characterized by the following features.

4. An inquiry response device according to claim 1 or 2, The aforementioned glossary defines standardized terms for each proper noun used within a given organization. The term conversion unit converts the proper nouns contained in the text into the unified phrases. A question response device characterized by the following features.

5. The inquiry response device according to claim 3, The aforementioned glossary defines standardized terms for each proper noun used within a given organization. The term conversion unit instructs the generation AI service to convert the proper nouns contained in the text into the unified phrases. A question response device characterized by the following features.

6. The inquiry response device according to claim 5, The term conversion unit instructs the generation AI service to convert the proper nouns contained in the text into the unified terms, and instructs the generation AI service to update the term dictionary if there are any proper nouns that are not included in the term dictionary. A question response device characterized by the following features.

7. An inquiry response device according to claim 4, The aforementioned glossary defines a standardized term and its explanatory text for each proper noun used within a given organization. The term conversion unit converts the proper nouns contained in the text into the standardized phrases and adds the explanatory text. A question response device characterized by the following features.

8. The inquiry response device according to claim 5, The aforementioned glossary defines a standardized term and its explanatory text for each proper noun used within a given organization. The term conversion unit instructs the generation AI service to convert the proper nouns contained in the text into the standardized phrases and to add the explanatory text. A question response device characterized by the following features.

9. An inquiry response system including an operator terminal and an inquiry response device, The operator terminal, upon receiving the input of the inquiry text, performs a question processing step of passing the text to the inquiry response device. The aforementioned inquiry response device, A question-answering step that receives the text of the aforementioned inquiry, A term conversion step in which terms contained in the text are corrected using a term dictionary stored in the memory unit, A RAG data acquisition step involves extracting data related to the text of the aforementioned inquiry as RAG data from a predetermined RAG database. A response request step involves transmitting the extracted RAG data and instructing a predetermined generation AI service to provide an answer to the inquiry, An inquiry response system characterized by implementing the following.

10. An inquiry response program for an information processing device that implements an inquiry response function, The aforementioned information processing device is It comprises a memory unit for storing a glossary and a processing unit, In the aforementioned processing unit, A question-answering step that accepts text inquiries, A term conversion step of correcting terms contained in the text using a term dictionary stored in the memory unit, A RAG data acquisition step involves extracting data related to the text of the aforementioned inquiry as RAG data from a predetermined RAG database. A response request step involves transmitting the extracted RAG data and instructing a predetermined generation AI service to provide an answer to the inquiry, An inquiry response program characterized by enabling the following actions.

11. A query response method for implementing a query response function in an information processing device, The aforementioned information processing device is It comprises a memory unit for storing a glossary and a processing unit, In the aforementioned processing unit, A question-answering step that accepts text inquiries, A term conversion step of correcting terms contained in the text using a term dictionary stored in the memory unit, A RAG data acquisition step involves extracting data related to the text of the aforementioned inquiry as RAG data from a predetermined RAG database. A response request step involves transmitting the extracted RAG data and instructing a predetermined generation AI service to provide an answer to the inquiry, A method for responding to inquiries, characterized by having the following action taken.

Images