Information processing device, information processing system, and information processing method

Abstract

The goal is to significantly improve work efficiency by retrieving specific information from a large volume of documents. [Solution] The knowledge DB 20 consists of a large number of files 21 to 24, and the information processing device 10 extracts relevant information based on the question sentences according to the question order of the file group linkage rule 18 for the files 21 to 24. In that question order, it generates an answer sentence for the generating AI 40 with prompts based on the question sentence or the previous answer sentence and related information, and outputs the final answer sentence.

Description

Technical Field

[0001] The present invention relates to, for example, an information processing apparatus, an information processing system, and an information processing method.

Background Art

[0002] Today, large language models are being utilized to obtain target information from a large number of documents. As an example of this, a technique has been proposed for extracting an extraction target from a large language model using a main prompt that indicates the extraction target and a detailed prompt that sets the indication in detail (Patent Document 1). In addition, in a method using RAG (Retrieval Augmented Generation), a large language model can be selected from, for example, GPT-3 (registered trademark), GPT-4 (registered trademark), PaLM (registered trademark), etc. for each prompt node, and a technique for providing (transmitting) a prompt and obtaining an output result in units of nodes has been proposed (Patent Document 2).

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Patent Document 2

Summary of the Invention

Problems to be Solved by the Invention

[0004] As described above, in the conventional document search technology, there is a problem that it is impossible to reach the target information when obtaining the target information from a large number of documents.

[0005] The present invention has been made in view of such a background, and an object thereof is to provide an information processing apparatus, an information processing system, and an information processing method capable of obtaining target information from a large number of documents. [Means for solving the problem]

[0006] To solve the above-mentioned problems and achieve the above objectives, one embodiment of the present invention is an information processing device connected to a knowledge database composed of a generating AI and a plurality of different types of knowledge data sets, comprising: an information processing unit that extracts knowledge data from a plurality of different types of knowledge data sets in a predetermined order based on an input question sentence, repeatedly executes a process in which the answer sentence obtained from the generating AI becomes the next question sentence based on the question sentence and the extracted knowledge data as prompts, and finally outputs the obtained answer sentence.

[0007] Another embodiment of the present invention is an information processing system connected to a generating AI, comprising: a knowledge database composed of a plurality of different types of knowledge data sets; and an information processing device that extracts knowledge data from the plurality of different types of knowledge data sets in a predetermined order based on an input question sentence, repeatedly executes a process in which the answer sentence obtained from the generating AI becomes the next question sentence based on the question sentence and the extracted knowledge data as prompts, and finally outputs the obtained answer sentence.

[0008] Furthermore, another embodiment of the present invention is an information processing method for a device connected to a knowledge database comprising a generating AI and a plurality of different types of knowledge data sets, characterized in that it includes the steps of extracting knowledge data from a plurality of different types of knowledge data sets in a predetermined order based on an input question sentence, repeatedly executing a process in which the answer sentence obtained from the generating AI becomes the next question sentence based on the question sentence and the extracted knowledge data as prompts, and finally outputting the obtained answer sentence. [Effects of the Invention]

[0009] According to the present invention, it is possible to obtain the desired information from a large amount of documents, thereby significantly improving the efficiency of business operations. [Brief explanation of the drawing]

[0010] [Figure 1] This is a block diagram showing an example of the functionality of a cascade RAG system according to one embodiment of the present invention. [Figure 2] This is a configuration diagram showing an example of the hardware configuration of an information processing device according to one embodiment of the present invention. [Figure 3] This is an explanatory diagram illustrating the overview of file transfer using one embodiment of the present invention. [Figure 4] This is an explanatory diagram illustrating the procedure for a file transfer scenario according to one embodiment of the present invention. [Figure 5] This is a flowchart illustrating the main operation of one embodiment of the present invention. [Modes for carrying out the invention]

[0011] Embodiments of the present invention will be described below with reference to the drawings. Note that the following description and drawings are merely illustrative examples for explaining the present invention, and have been omitted or simplified as appropriate for clarity of explanation. Furthermore, the present invention can be implemented in various other forms. Also, unless otherwise specified, each component may be singular or plural.

[0012] In the following explanation, identical or similar structures may be denoted by the same symbol, and redundant explanations may be omitted. Also, in the following explanation, various types of information may be described using expressions such as "information" and "table," but these types of information may be represented by data structures other than these. Furthermore, while expressions such as "identification information," "identifier," "name," "ID," and "number" may be used to represent identification information, these can be substituted for each other. In the following explanation, "database" will be abbreviated as "DB" and "table" as "TBL."

[0013] First, the configuration will be described using FIGS. 1 and 2. FIG. 1 is a block diagram showing a functional example of the cascade RAG system according to the present embodiment, and FIG. 2 is a configuration diagram showing a hardware configuration example of the information processing apparatus according to the present embodiment.

[0014] The cascade RAG system 1, which is an information processing system, is composed of, for example, an information processing apparatus 10 and a knowledge DB 20 as shown in FIG. 1, and is connected via a network (not shown) to a user terminal 30, a generation AI 40 that generates a response sentence based on a prompt by an LLM (Large Language Model), etc.

[0015] Specifically, in the cascade RAG system 1, the information processing apparatus 10 includes an information processing unit 10A composed of a question reception unit 11, a knowledge DB question generation unit 12, an extraction unit 13, a prompt generation unit 14, a prompt transmission unit 15, a response acquisition unit 16, a response check unit 17, a file group cooperation rule 18, a response output unit 19, etc.

[0016] The question reception unit 11 receives a text question sentence from the user terminal 30. The knowledge DB question generation unit 12 generates a vector data question sentence for searching the knowledge DB 20 using the question sentence, and the extraction unit 13 extracts data from the knowledge DB 20 using the question sentence.

[0017] The prompt generation unit 14 generates a prompt for the generation AI 40 based on the question sentence generated by the knowledge DB question generation unit 12 and the result extracted by the extraction unit 13, that is, the relevant information. The prompt transmission unit 15 transmits the generated prompt to the generation AI 40.

[0018] Then, the response acquisition unit 16 acquires the response sentence generated based on the prompt transmitted by the prompt transmission unit 15, and the response check unit 17 checks the positive or negative of the response sentence by a score check determined in advance.

[0019] In this check, for example, when it is lower than a predetermined threshold value, the answer check unit 17 instructs the knowledge DB question generation unit 12 to do it again as there is a problem with the check. Further, if there is no problem with the check, the answer check unit 17 passes the answer sentence to the knowledge DB question generation unit 12 for the next question and notifies the subsequent file group link rule 18 of normal, and if there is a problem with the check, it transmits an abnormality intended to be redone within the same file group. This transmission may be, for example, data of normal "0" and abnormal "1".

[0020] The file group link rule 18 is information stored as a route for linking the question order of four file groups 21 to 24, which are knowledge data groups stored in the knowledge DB 20, that is, a link rule. Of course, the link rule is set according to the number of file groups stored in the knowledge DB 20. If there are four file groups, there are three connecting routes, and if there are seven file groups, there are six connecting routes.

[0021] This file group link rule 18 is configured by preparing different link rules in advance according to the type of question sent from the user terminal 30. As will be described later, in the link between file groups, the items provided in each file become link elements, and the items and file groups to be linked differ according to the type of question.

[0022] When the file group link rule 18 receives normal "0" from the answer check unit 17, it designates the transition to the next file group, which is the next link destination, for the knowledge DB question generation unit 12. When it receives abnormal "1" from the answer check unit 17, it designates the file group that is the same link destination as the previous time for the knowledge DB question generation unit 12.

[0023] Note that in the case of normal "0", the answer check unit 17 outputs the answer sentence received from the answer acquisition unit 16 to the answer output unit 19, and the answer output unit 19 outputs the answer sentence to the user terminal 30 as the final answer sentence.

[0024] The knowledge database 20 consists of a large amount of knowledge data composed of multiple different file groups, and for example, it consists of storage that stores four file groups 21 to 24. Each file group 21, 22, 23, and 24 has an item, which is the main element that links the file groups together according to the file group linkage rule 18 described above. In other words, file groups 21 to 24 are configured to achieve linkage through items according to the file group linkage rule 18.

[0025] Regarding the aforementioned file sets, the number of hierarchical levels prepared is not limited to four, depending on the industry or business type; it may be four or fewer, or four or more. Furthermore, the number of items that serve as linking elements, as mentioned above, may be one or more. In addition, each file set 21 to 24 consists of multiple files.

[0026] The user terminal 30 may be a PC or a mobile device such as a smartphone, and the generating AI 40 may be a text generation AI such as ChatGPT, but it may also be a combination of a text generation AI and an image processing AI. The text generation AI is, for example, a language model that has been trained to estimate and output the next token from an input sequence of tokens, and is a language model such as an attention model or a transfer model (e.g., ChatGPT, BERT).

[0027] Examples of transfer models include GPT (Generative Pre-trained Transformer) and BERT (Bidirectional Encoder Representations from Transformers). Furthermore, the generative AI40 may be fine-tuned, and may also be a multimodal generative AI that mixes text and images.

[0028] The information processing device 10 is configured to communicate with the knowledge DB 20, user terminal 30, and generation AI 40 via the network 50, as shown in Figure 2, for example.

[0029] The information processing device 10 consists of, for example, a CPU 101 that controls the entire device, a memory 103 that stores programs 102 such as processes to be executed by the CPU 101 (see Figure 6) and various data being executed, an operating device 104 equipped with a keyboard and a display device, an external storage device 105 that registers and stores various data in formats such as DB and TBL, a communication IF 106 connected to the network 50 and responsible for various external communications, and a bus 107 connected to each unit within the device and responsible for internal data and signal communication within the device.

[0030] Next, we will explain the operation using Figures 3, 4, and 5. Figure 3 is an explanatory diagram illustrating the overview of file linking according to this embodiment, Figure 4 is an explanatory diagram illustrating the procedure of the file linking scenario according to this embodiment, and Figure 5 is a flowchart illustrating the main operations according to this embodiment.

[0031] First, let's explain the concept of file linking using Figure 3. In Knowledge DB20, file groups 21, 22, 23, and 24 contain files A211, B221, C231, and D241, respectively. Furthermore, files A211, B221, C231, and D241 contain items W, X, Y, and Z, respectively.

[0032] The file group linking rule 18 follows a flow similar to the file linking transition 200 processed by the information processing unit 10A, where file A211 and file B221 are connected by item W and item X, file B221 and file C231 are connected by item X and item Y, and file C231 and file D241 are connected by item Y and item Z.

[0033] In this way, the information processing unit 10A receives a question from the user terminal 30 and first extracts knowledge data for item W from file A211 in the knowledge DB 20 based on that question. Subsequently, the extracted knowledge data, i.e., the related information and the question, are sent to the generating AI 40 as a prompt, and the above process is repeated until all the connections are satisfied.

[0034] In other words, the first link is between item W in file A211 and item X in file B221, and knowledge data for item X is extracted. In this way, the next link is between item X in file B221 and item Y in file C231, and knowledge data for item Y is extracted. The final link is between item Y in file C231 and item Z in file D241, and knowledge data for item Z is extracted. The knowledge data for item Z is then output to the user terminal 30 as an answer.

[0035] Next, we will explain the more specific details of the collaboration using Figure 4. The overall flow is as explained in Figure 3. For example, in the case of financial transactions, when identifying which part of the online ledger is being updated by the input fields of the accounting and settlement system, we will explain the flow when the question is "the part updating liquidity account opening transactions." Since the processing flow has already been explained, only the main functions will be explained below.

[0036] In this case, the user terminal 30 inputs the above natural language question, and the knowledge database question generation unit 12 extracts related information for item W based on the question, generating vector data to confirm, for example, "the correspondence between a predetermined number and a predetermined code." The extraction unit 13 performs knowledge data extraction on item W in the file group 211 based on the vector data to obtain related information 1.

[0037] Then, the prompt generation unit 14 generates a prompt P1 based on the related information 1 and the question text concerning item W, and the prompt transmission unit 15 transmits the prompt P1 to the generation AI 40. The answer text 1 is obtained from the generation AI 40, and its contents are, for example, such that AAA is a predetermined code and BBB is a predetermined number.

[0038] Next, the question generation unit 12 for the knowledge database extracts related information for item X based on the answer sentence 1 according to the file group linkage rule 18. For example, vector data is generated to confirm "the CCC input item corresponding to the BBB input item, by referring to the transaction-specific input / output explanation document for the target BBB number." The extraction unit 13 extracts knowledge data for item X in the file group 221 based on this vector data to obtain related information 2, and the prompt generation unit 14 generates prompt P2.

[0039] Furthermore, according to file group linking rule 18, related information 3 is extracted from item Y of file group 231 and related information 4 is extracted from item Z of file group 241. Prompts P3 and P4 are generated for generated AI 40, respectively, and answer sentences 2 and 3 are obtained.

[0040] In this way, the collaborative process is executed repeatedly, and the final answer statement 4 is obtained. This answer statement 4 is obtained as text, for example, "The part of the liquidity account opening transaction being updated is... The information referenced is item W in file A, item X in file B,..." Of course, in line with the spirit of the present invention, depending on the type of question, representations such as tables and diagrams are also possible.

[0041] Finally, an example of operation will be explained using Figure 5. First, the information processing device 10 provides data to the accessing user terminal 30 for displaying a question input screen (step S2001). On the user terminal 30, the question input screen is displayed on a display (not shown) based on the input data from the information processing device 10 (step S1001).

[0042] On the user terminal 30, the user inputs a question into the question input screen and sends it to the information processing device 10 (step S1002). When the information processing device 10 receives the question, it performs the process of obtaining the final answer using the four file groups 21 to 24, as already explained in Figures 3 and 4.

[0043] For the step-by-step processing, relevant information (knowledge data) is extracted (search process) from the knowledge DB20 (step S2003), and in response, a search is performed on the specified file group within the knowledge DB20 according to the file group linkage rule 18 (step S3001).

[0044] The search results are provided to the information processing device 10 as related information (step S3002). The information processing device 10 generates a prompt based on the question text and related information (step S2004) and sends it to the generating AI 40 (step S2005).

[0045] The generating AI 40 generates a response based on the prompt (step S4001) and sends it to the information processing device 10 (step S4002). When the information processing device 10 receives the response from the generating AI 40 (step S2006), it determines whether the response is the final response (step S2007).

[0046] If File Group Linking Rule 18 determines that a final answer has not yet been reached (NO route in step S2007), the answer is set as the next question, and the next set of files is specified as the extraction target according to File Group Linking Rule 18 (step S2008). The process then returns to step S2003, and the same process is repeated.

[0047] Furthermore, if the file group linking rule 18 determines that a final response has been reached (the YES route in step S2007), the final response text is sent to the user terminal 30 (step S2009). Then, the received response text is displayed on a display (not shown) in the user terminal 30 (step S1003). The response text check is as described above.

[0048] As explained above, this embodiment allows for the acquisition of target information from a large volume of documents, significantly improving work efficiency and enabling its use in carrying out tasks.

[0049] In particular, in financial institutions such as banks, identifying which part of the online ledger is being updated by the input fields of the bookkeeping and settlement system requires identification from a large amount of knowledge data, so this is expected to be effective.

[0050] Furthermore, the information to be acquired is not limited to a single specific type; it will be acquired if the conditions are met.

[0051] Of course, the application field of this invention is not limited to the financial sector; since it identifies target information from a large amount of data (files), it can be applied to all industries.

[0052] Furthermore, while the above-described embodiment used items as elements in the linking rules, the present invention is not limited to this, and other elements such as data types may be applied.

[0053] Furthermore, each of the above-mentioned configurations, functional units, processing units, processing means, etc., may be implemented in hardware, in whole or in part, for example, by designing them as integrated circuits. Alternatively, each of the above-mentioned configurations, functions, etc., may be implemented in software by having the processor interpret and execute programs that realize each function. Information such as programs, tables, and files that realize each function can be stored in memory, hard disks, SSDs (Solid State Drives), or other recording devices, or in recording media such as IC cards, SD cards, or DVDs.

[0054] Furthermore, the arrangement of the various functional units, processing units, and databases described above is merely an example. The arrangement of the various functional units, processing units, and databases can be changed to the optimal arrangement from the standpoint of the performance, processing efficiency, and communication efficiency of the hardware and software of these devices.

[0055] Furthermore, the configuration of the database (schema, etc.) that stores the various types of data mentioned above can be flexibly modified from the perspective of efficient resource utilization, improved processing efficiency, improved access efficiency, and improved search efficiency. [Explanation of Symbols]

[0056] 1. Cascade RAG System 1A Information Processing Unit 10 Information Processing Devices 11. Question Reception Department 12 Question generator for knowledge DB 13 Extraction part 14. Prompt generation unit 15 Prompt transmission unit 16 Answer acquisition part 17. Answer Check Section 18 File group linking rules 19. Answer Output Section 20 Knowledge DB 30 User terminals 40 Generation AI 50 Networks 101 CPU 102 Programs 103 memory 104 Operating Devices 105 External storage device 106 Communication IF 107 Bus

Claims

1. An information processing device connected to a knowledge database consisting of generating AI and multiple different types of knowledge data sets, An information processing unit extracts knowledge data from a plurality of different types of knowledge data sets according to a predetermined order based on an input question, and repeatedly executes a process in which the answer obtained from the generating AI becomes the next question based on a prompt based on the question and the extracted knowledge data, and finally outputs the obtained answer. An information processing device characterized by comprising:

2. The information processing apparatus according to claim 1, characterized in that the generating AI has a large-scale language model.

3. The information processing device according to claim 1, wherein the information processing device has a memory for storing linkage rules for linking a plurality of different types of knowledge data sets in a predetermined order and repeating the processing.

4. The information processing device according to claim 3, wherein each of the plurality of different types of knowledge data groups is composed of a file group, and the linking rule is a rule that links the linking elements contained in the files of each file group.

5. The information processing device according to claim 4, characterized in that the linking element is an item.

6. The information processing apparatus according to claim 1, wherein the information processing apparatus has a prompt generation unit that generates the prompt based on the question and the extracted knowledge data.

7. The information processing apparatus according to claim 1, wherein the information processing unit has a question generation unit that converts the question sentence into vector data and asks a question to the knowledge data group.

8. The information processing device according to claim 7, wherein the information processing device has an answer checking unit that instructs the question generation unit to redo the answer according to the accuracy of the answer obtained from the generated AI.

9. An information processing system connected to a generating AI, A knowledge database composed of multiple different types of knowledge data sets, An information processing device that extracts knowledge data from a plurality of different types of knowledge data sets in a predetermined order based on an input question, repeatedly executes a process in which the answer obtained from the generating AI becomes the next question based on a prompt based on the question and the extracted knowledge data, and finally outputs the obtained answer. An information processing system characterized by comprising the following features.

10. The information processing system according to claim 9, characterized in that the generating AI has a large-scale language model.

11. The information processing system according to claim 9, wherein the information processing device has a memory for storing linkage rules for linking a plurality of different types of knowledge data sets in a predetermined order and repeating the processing.

12. An information processing system according to claim 9, characterized in that the information processing device and the knowledge database constitute a cascade RAG (Retrieval Augmented Generation) system.

13. An information processing method for a device connected to a knowledge database consisting of a generation AI and multiple different types of knowledge data sets, An information processing method characterized by including the steps of extracting knowledge data from a plurality of different types of knowledge data sets in a predetermined order based on an input question, repeatedly executing a process in which the answer obtained from the generating AI becomes the next question based on a prompt based on the question and the extracted knowledge data, and finally outputting the obtained answer.

14. The information processing method according to claim 13, wherein each of the plurality of different types of knowledge data groups is composed of a file group, and the processing is repeated by linking the linking elements contained in the files of each file group.

15. The information processing method according to claim 14, characterized in that the linking element is an item.

Images