Information processing system

Abstract

It suppresses the occurrence of hallucination. [Solution] The information processing system (1) includes an acquisition means (111) for acquiring a question and a response from a large-scale language model to the question, and a determination means (112) for determining whether the response is incorrect based on the relationship between the question and the response and predetermined relationship information. The predetermined relationship information indicates the relationship between multiple question sentences and multiple responses from a large-scale language model corresponding to each of the multiple question sentences.

Description

【Technical Field】 【0001】 The present invention relates to the technical field of information processing systems. 【Background Art】 【0002】 As this type of system, for example, a system has been proposed in which a query data based on a document is generated for a language model, and a pair of the document and the query data is used for learning a search model for a chatbot (see Patent Document 1). 【Prior Art Documents】 【Patent Documents】 【0003】 【Patent Document 1】 Japanese Patent Application Laid-Open No. 2023-076413 【Summary of the Invention】 【Problems to be Solved by the Invention】 【0004】 As a chatbot, a chatbot using a mechanism (Retrieval-Augmented Genration: RAG) that gives a large language model (Large Language Models: LLM) an independent information source by combining a large language model and the search for a specific information source (hereinafter, appropriately referred to as a "knowledge base") has been proposed. By the way, when the data necessary for answering a user's question is not registered in the knowledge base, data unrelated to the above question may be extracted by the search of the knowledge base, and the large language model may generate a false answer using the extracted data. Note that a large language model is a language model constructed using a very large dataset and deep learning technology. 【0005】 The present invention has been made in view of the above problems, and an object thereof is to provide an information processing system capable of suppressing the occurrence of hallucination. 【Means for Solving the Problems】 【0006】 An information processing system according to one aspect of the present invention comprises: acquisition means for acquiring a question and a response from a large-scale language model to the question; and determination means for determining whether the response is incorrect based on the relationship between the question and the response and predetermined relationship information, wherein the predetermined relationship information indicates the relationship between a plurality of question and a plurality of responses from a large-scale language model corresponding to each of the plurality of question. [Brief explanation of the drawing] 【0007】 [Figure 1] This is a diagram showing the configuration of an information processing system according to an embodiment. [Figure 2] This figure shows an example of a screen related to a chatbot. [Figure 3] This block diagram shows an example of the configuration of a computing device according to the embodiment. [Modes for carrying out the invention] 【0008】 Embodiments of the information processing system will be described with reference to Figures 1 to 3. In Figure 1, the information processing system 1 comprises an information processing device 10, a server 20, a knowledge base 30, and a database 40. The information processing device 10, server 20, knowledge base 30, and database 40 are configured to communicate with each other via a network NW. Server 20 is a server for operating a large-scale language model (LLM). For this reason, server 20 may be referred to as an LLM server. Server 20 may be a cloud server. 【0009】 Knowledge base 30 may contain multiple text data entries. These text data entries may be fragments of text that have been divided from a single document. These fragments may be referred to as "chunks." Methods for dividing text from a single document include, for example, dividing it at fixed lengths, dividing it at sentence levels based on sentence delimiters, or dividing it based on structure such as Markdown. Knowledge base 30 may also contain multiple text data entries that have been vectorized. In other words, knowledge base 30 may be a vector database / vector store. In addition to text data, knowledge base 30 may also contain image data. 【0010】 (Chatbot) Server 20 and Knowledge Base 30 may provide a chatbot service using RAG. For example, user U may use the chatbot service via terminal device 50. In this case, user U may operate terminal device 50 to launch an application for using the chatbot service. User U may operate terminal device 50 to enter a question into the input field of the chat application. Here, "question" is not limited to interrogative sentences. For example, "question" may be a sentence that includes expressions such as requests, instructions, or commands, such as "Tell me about ****" or "Answer me about ****". Therefore, "question" is a concept that includes not only interrogative sentences but also sentences that include expressions such as requests, instructions, or commands. In other words, "question" may mean a sentence that seeks an answer from the other party. Note that terminal device 50 may be a personal computer, a tablet terminal, or a smartphone. 【0011】 Terminal device 50 may search the knowledge base 30 based on the input question. Terminal device 50 may send first information, which includes the input question and text data as search results from the knowledge base 30, to server 20. Server 20 may input the question and text data included in the first information as a prompt to a large-scale language model. Server 20 may obtain the answer to the question output from the large-scale language model. Server 20 may send second information indicating the answer to terminal device 50. Upon receiving the second information, terminal device 50 may display the answer indicated by the second information on a screen related to the chat application. 【0012】 As a result, the terminal device 50 may display the screen 51 shown in Figure 2. For example, the screen 51 may include an area 511 for displaying the question entered by user U and an area 512 for displaying the answer from the large-scale language model. For example, below area 512, buttons 513 and 513 may be placed for user U to input their evaluation of the large-scale language model's answer. In the example shown in Figure 2, user U's evaluation is a two-level rating of "GOOD" and "BAD". However, user U's evaluation is not limited to a two-level rating, but may be a three-level or higher rating. 【0013】 User U may operate the terminal device 50 to select button 513 or 514 (in other words, User U may evaluate the response of the large-scale language model). If User U's evaluation is "GOOD" (in other words, a high evaluation), the terminal device 50 may send third information to the server 20, which includes the response evaluated by User U (for example, the response displayed in area 512 of Figure 2) and the question related to the response evaluated by User U (for example, the question displayed in area 511 of Figure 2). Upon receiving the third information, the server 20 may determine the relationship between the question included in the third information (i.e., User U's question) and the answer (i.e., the response that User U highly evaluated). For example, the server 20 may vectorize the question and the answer. That is, the server 20 may project the question and the answer onto a vector space. The server 20 may determine the relationship between the question and the answer by calculating the inner product of the vectorized question and the vectorized answer. Server 20 may register information indicating the relationship between the question and the answer (e.g., the dot product) in the database 40. The dot product of the vectorized question and the vectorized answer may be called the relevance score. 【0014】 (Information processing device 10) In Figure 1, the information processing device 10 comprises an arithmetic unit 11, a storage device 12, a communication device 13, an input device 14, and an output device 15. The arithmetic unit 11, storage device 12, communication device 13, input device 14, and output device 15 are connected via a data bus 16. The information processing device 10 may be a personal computer, a tablet terminal, or a smartphone. 【0015】 The arithmetic unit 11 may have a processor. The arithmetic unit 11 may have a single processor or multiple processors. In other words, the arithmetic unit 11 may have one or more processors. Furthermore, the processor may be a multi-core processor. If the arithmetic unit 11 has a single processor that is a multi-core processor, then logically, the arithmetic unit 11 can be said to have multiple processors. 【0016】 The processor may be at least one of the following: CPU (Central Processing Unit), GPU (Graphics Processing Unit), FPGA (Field Programmable Gate Array), and TPU (Tensor Processing Unit). 【0017】 The storage device 12 may be at least one of the following: RAM (Random Access Memory), ROM (Read Only Memory), hard disk drive, magneto-optical disk drive, SSD (Solid State Drive), and optical disk array. In other words, the storage device 12 may be implemented by a single device or by multiple devices. 【0018】 The communication device 13 may be capable of communicating with devices outside the information processing device 10. The communication device 13 may use either wired or wireless communication. 【0019】 The input device 14 is a device capable of receiving information input to the information processing device 10 from an external source. The input device 14 may include an operating device (e.g., a keyboard, mouse, touch panel, etc.) that can be operated by the user of the information processing device 10. The input device 14 may include a recording medium reader capable of reading information recorded on a recording medium that can be attached to and detached from the information processing device 10, such as a USB (Universal Serial Bus) memory. When information is input to the information processing device 10 via the communication device 13 (in other words, when the information processing device 10 acquires information via the communication device 13), the communication device 13 may function as an input device. 【0020】 The output device 15 is a device capable of outputting information to the outside of the information processing device 10. The output device 15 may have a display device capable of outputting visual information such as characters and images as the above information. Incidentally, the output device 15 may have a speaker capable of outputting auditory information such as voice as the above information. The output device 15 may have a vibration motor capable of outputting tactile information such as vibration as the above information. The output device 15 may have a printer. The output device 15 may be capable of outputting information to a recording medium detachable from the information processing device 10, such as a USB memory. Incidentally, when the information processing device 10 outputs information via the communication device 13, the communication device 13 may function as an output device. 【0021】 The storage device 12 can store desired data. The storage device 12 may store the computer program CP executed by the arithmetic device 11. The storage device 12 may temporarily store data temporarily used by the arithmetic device 11 when the arithmetic device 11 is executing the computer program CP. 【0022】 Incidentally, the computer program CP may be recorded on a computer-readable and non-temporary recording medium. In this case, the computer program CP may be stored in the storage device 12 by reading the recording medium using a recording medium reading device (not shown) provided in the information processing device 10. Incidentally, as the above recording medium, at least one of an optical disk, a magnetic medium, a magneto-optical disk, a semiconductor memory, and any other medium capable of storing a program may be used. Incidentally, the computer program CP may be acquired from a device (not shown) outside the information processing device 10 via the communication device 13. In other words, the computer program CP may be downloaded from an external device to the storage device 12 of the information processing device 10. 【0023】 The arithmetic unit 11 (e.g., a processor) may execute the processing to be performed by the information processing apparatus 10 together with the storage device 12 in which the computer program CP is stored (in other words, together with the storage device 12 and the computer program CP stored in the storage device 12). For example, by executing the computer program CP, a logical functional block for executing the processing to be performed by the information processing apparatus 10 may be realized in the arithmetic unit 11 (e.g., within the processor). 【0024】 By the way, in the above-described chatbot service, data (e.g., text data) related to the question sentence of the user (e.g., user U) may not be registered in the knowledge base 30. Even in this case, when the knowledge base 30 is searched, some data is often extracted as a search result. When the user's question sentence and the data as the search result are input to the large language model, the large language model may generate a false answer (that is, hallucination may occur). In addition, it may be difficult for the user to determine whether the answer generated by the large language model is a hallucination. 【0025】 Therefore, the information processing system 1 according to the present embodiment determines whether the answer generated by the large language model is incorrect. As shown in FIG. 3, the arithmetic unit 11 of the information processing apparatus 10 includes an acquisition unit 111, a determination unit 112, and a control unit 113 for performing the above determination. The acquisition unit 111, the determination unit 112, and the control unit 113 may be realized as the above-described logical functional blocks. However, at least one of the acquisition unit 111, the determination unit 112, and the control unit 113 may be realized as a physical processing circuit. Alternatively, at least one of the acquisition unit 111, the determination unit 112, and the control unit 113 may be realized in a form in which logical functional blocks and physical processing circuits are mixed. 【0026】 The acquisition unit 111 acquires a question sentence input by a user (e.g., user U) and a response from a large-scale language model to that question sentence from the server 20. In this embodiment, the response from the large-scale language model may be transmitted from the server 20 to the information processing device 10 before it is transmitted to a terminal device (e.g., terminal device 50). 【0027】 The determination unit 112 may determine the relationship between a question and an answer. For example, the determination unit 112 may vectorize both the question and the answer. That is, the determination unit 112 may project the question and the answer onto a vector space. The determination unit 112 may determine the relationship between the question and the answer by calculating the inner product of the vectorized question and the vectorized answer. 【0028】 The determination unit 112 determines whether an answer is incorrect based on the relationship between a question and an answer, and the relationships between multiple question statements and multiple answers corresponding to those multiple question statements, which are registered (or stored) in the database 40. As described above, the answers for which relationships are sought registered in the database 40 are answers that have been highly rated by the user (for example, user U). For this reason, the answers for which relationships are sought registered in the database 40 are very likely to be correct answers (in other words, answers in which hallucination does not occur). Therefore, it can be said that the relationships registered in the database 40 indicate the relationship between a question and the correct answer to that question. For this reason, if the relationship between a question and an answer is similar to the relationship registered in the database 40, the determination unit 112 may determine that the answer is not incorrect. On the other hand, if the relationship between a question and an answer is not similar to the relationship registered in the database 40, the determination unit 112 may determine that the answer is incorrect. Furthermore, the determination unit 112 may determine whether an answer is incorrect or not by using a learning model that, upon input of a question and an answer, outputs whether or not the answer is incorrect. 【0029】 If the determination unit 112 determines that one of the answers is not incorrect, the control unit 113 permits the server 20 to transmit (in other words, output) the answer to the terminal device (for example, terminal device 50). On the other hand, if the determination unit 112 determines that one of the answers is incorrect, the control unit 113 prohibits the server 20 from transmitting the answer to the terminal device. In this case, the server 20 may send information to the terminal device causing it to output a message such as "No suitable answer was found." 【0030】 (Technical effects) In the information processing system 1 according to this embodiment, before the answer generated by the large-scale language model is transmitted to the terminal device (for example, terminal device 50), the information processing device 10 determines whether the answer is incorrect. If the answer is determined to be incorrect, transmission of the answer to the terminal device is prohibited. In other words, in the information processing system 1, although the large-scale language model may generate an incorrect answer, it is possible to suppress the presentation of an incorrect answer to the user. Therefore, according to the information processing system 1, the occurrence of hallucination can be suppressed. 【0031】 Although the above-described embodiment is based on a chatbot service using RAG, the present invention is also applicable to chatbot services that do not use a knowledge base (for example, knowledge base 30). 【0032】 Various aspects of the invention derived from the embodiments described above are described below. 【0033】 An information processing system according to one aspect of the invention comprises an acquisition means for acquiring a question and a response from a large-scale language model to the question, and a determination means for determining whether the response is incorrect based on the relationship between the question and the response and predetermined relationship information, wherein the predetermined relationship information indicates the relationship between a plurality of question and a plurality of responses from a large-scale language model corresponding to each of the plurality of question. In the above embodiment, the "acquisition unit 111" corresponds to an example of the "acquisition means," and the "determination unit 112" corresponds to an example of the "determination means." 【0034】 In the information processing system according to the above embodiment, each of the multiple responses may be a response that has been highly rated by the user. 【0035】 In the information processing system according to the above embodiment, the predetermined relationship information may represent a relationship in a vector space, and the determination means may determine whether the first answer is incorrect based on the relationship in the vector space between the first question and the first answer, and the relationship information. 【0036】 In the information processing system according to the above embodiment, if it is determined that the first answer is incorrect, the system may be equipped with control means to prohibit the output of the first answer. In the above embodiment, the "control unit 113" is an example of the "control means". 【0037】 The present invention is not limited to the embodiments described above, and can be modified as appropriate without contradicting the gist or idea of ​​the invention as can be read from the claims and specification as a whole. Information processing systems involving such modifications are also included within the technical scope of the present invention. [Explanation of symbols] 【0038】 1... Information processing system, 10... Information processing device, 20... Server, 30... Knowledge base, 40... Database, 111... Acquisition unit, 112... Determination unit, 113... Control unit

Claims

[Claim 1] An acquisition means for acquiring a question sentence and a response from a large-scale language model to the said question sentence, A determination means for determining whether the answer to the first question is incorrect, based on the relationship between the first question and the first answer, and predetermined relationship information, Equipped with, The predetermined relationship information indicates the relationship between multiple question sentences and multiple responses from a large-scale language model corresponding to each of the multiple question sentences. Information processing system. [Claim 2] Each of the above responses is a response that was highly rated by the user. The information processing system according to claim 1. [Claim 3] The aforementioned predetermined relationship information represents relationships in a vector space. The determination means determines whether the first answer is incorrect based on the relationship between the first question and the first answer in the vector space and the relationship information. The information processing system according to claim 1. [Claim 4] The system includes a control means that prohibits the output of the first answer if it is determined that the first answer is incorrect. The information processing system according to claim 1.

Images