Information processing method, information processing device, and computer program

Abstract

An answer generating unit 24 of an information processing device 10 generates an answer sentence in token units using generative AI on the basis of an input prompt. In the process of generating the answer sentence, a determining unit 28 of the information processing device 10 determines whether or not a predetermined condition is satisfied for each token. The answer generating unit 24 of the information processing device 10 interrupts the generation of the answer sentence when it is determined that the predetermined condition is satisfied. When it is determined that the predetermined condition is satisfied, a display control unit 46 of the information processing device 10 outputs information relating to the status of the generation of the answer sentence.

Description

Information processing method, information processing device, and computer program

[0001] The present disclosure relates to data processing technology, and more particularly to an information processing method, an information processing device, and a computer program.

[0002] Writing sentences using generative AI (also known as large language models (LLMs)) is becoming more common. Patent Document 1 listed below describes a technology that suppresses erroneous output by determining, based on a confidence level, whether to output an output result (an answer sentence based on a probability distribution P(y|x)) when an input sentence x is input to a language model M1.

[0003] Japanese Patent Application Laid-Open No. 2022-185799

[0004] Currently, there is a problem with generative AI generating information that is inconsistent with the facts (also known as "hallucination"). In particular, hallucination can occur when the generative AI has insufficient training data. One objective of the present disclosure is to provide a technology that suppresses the occurrence of hallucination when creating text using generative AI, while supporting learning about insufficient training data.

[0005] In order to solve the above problem, an information processing method of one aspect of the present disclosure is an information processing method in a computer, which includes generating an answer sentence on a token-by-token basis using a generation AI based on an input prompt, determining whether or not a predetermined condition is met for each token in the process of generating the answer sentence, and if it is determined that the predetermined condition is met, suspending the generation of the answer sentence, and if it is determined that the predetermined condition is met, further outputting information regarding the status of the generation of the answer sentence.

[0006] Another aspect of the present disclosure is an information processing device. The device includes an answer generation unit that generates an answer sentence on a token-by-token basis using a generation AI based on an input prompt, a determination unit that determines whether a predetermined condition is satisfied for each token in the process of generating the answer sentence, and an output unit. If the answer generation unit determines that the predetermined condition is satisfied, it suspends generation of the answer sentence, and if the output unit determines that the predetermined condition is satisfied, it outputs information regarding the status of generation of the answer sentence.

[0007] Any combination of the above components, or any conversion of the expression of the present disclosure between a system, a computer program, a recording medium on which a computer program is recorded, etc., is also valid as an aspect of the present disclosure.

[0008] According to the technology disclosed herein, it is possible to suppress the occurrence of hallucination when creating text using generative AI while supporting learning of missing training data.

[0009] FIG. 1 is a diagram showing an example of generating an answer sentence when there is sufficient learning data; FIG. 2 is a diagram showing an example of generating an answer sentence when there is insufficient learning data; FIG. 3 is a block diagram showing functional blocks of the information processing device of the first embodiment; FIG. 4 is a flowchart showing the operation of the information processing device of the first embodiment; FIG. 5 is a flowchart showing details of the answer generation process of S12 in FIG. 4; and FIG. 6 is a flowchart showing the operation of the information processing device of the second embodiment.

[0010] The subject of the device or method disclosed herein includes a computer. The computer executes a program to realize the functions of the subject of the device or method disclosed herein. The computer includes, as its main hardware configuration, a processor that operates according to the program. The type of processor is not important as long as it can realize the functions by executing the program. The processor is composed of one or more electronic circuits, including a semiconductor integrated circuit (IC) or a large-scale integration (LSI). The electronic circuits may be integrated into a single chip or may be provided on multiple chips. The multiple chips may be integrated into a single device or may be provided on multiple devices. The program may be recorded on a non-transitory recording medium such as a computer-readable read-only memory (ROM), an optical disk, or a hard disk drive, or on a temporary storage medium such as a computer-readable random access memory (RAM). The program may be stored in advance on a recording medium, or may be supplied to a recording medium or storage medium via a wide area communication network including the Internet.

[0011] As the use of generative AI to create sentences is becoming more common, hallucination, in which generative AI generates information that is inaccurate, is becoming a problem. Hallucination can occur particularly when the generative AI does not have enough training data.

[0012] When generating an answer using generative AI, the next word that is likely to follow a certain word (also called a "token") is predicted, and the word that is likely to follow next is selected. Figure 1 shows an example of an answer sentence generated when there is sufficient training data. In Figure 1, an answer sentence is generated on a token-by-token basis in response to the input information (hereinafter also referred to as a "prompt"), "What is the capital of Japan?" As shown in Figure 1, when there is sufficient training data, there are words that are likely to be selected among the next words, and therefore a word that expresses a fact ("Tokyo" in Figure 1) is likely to be selected.

[0013] Figure 2 shows an example of generating answer sentences when training data is insufficient. In Figure 2, answer sentences are generated on a token-by-token basis in response to the prompt, "Please give me an example sentence to introduce myself." As shown in Figure 2, when training data is insufficient, there are no words with an exceptionally high selection probability, and the distribution of the selection probability for each word candidate is nearly uniform, which may result in the selection of a word that is inaccurate with the actual situation. For example, in Figure 2, words other than the correct answer, "music," are likely to be selected, meaning that hallucination is likely to occur.

[0014] There is a need for a system that will suspend answer generation when the learning data for the generation AI is insufficient and the probability of the next word being correct is low (i.e., when the probability distribution of the selected words is close to uniform).If the prompt simply includes an instruction to "suspend answer generation if you are not confident," there is a possibility that the generation AI will not follow the instruction, so a suspension mechanism that does not rely on prompts is required.

[0015] Therefore, in the embodiment, in the process of generating answer sentences for input prompts on a token-by-token basis using the generation AI, it is determined whether a predetermined interruption condition is met for each token, and if the interruption condition is met, the generation of the answer sentence is interrupted and information about the status of the generation of the answer sentence is output. This makes it possible to suppress the occurrence of hallucination in sentence creation using the generation AI while supporting learning of missing training data.

[0016] <First Embodiment> Figure 3 is a block diagram showing functional blocks of an information processing device 10 according to a first embodiment. Each block shown in the block diagram in this specification can be realized in terms of hardware by a computer processor (e.g., a CPU), memory, and other elements, electronic circuits, and mechanical devices, and in terms of software by a computer program, etc. However, the functional blocks shown here are realized by the cooperation of these elements. Therefore, it will be understood by those skilled in the art that these functional blocks can be realized in various ways by combining hardware and software.

[0017] The information processing device 10 includes an input acceptance unit 12, a display unit 14, a generation AI unit 16, and a processing unit 18. The input acceptance unit 12 accepts information input by a user who is attempting to create a sentence using the generation AI. The input acceptance unit 12 inputs the accepted information to the processing unit 18. The display unit 14 displays various information according to the control of the processing unit 18. The input acceptance unit 12 and the display unit 14 may be integrally formed as a touch screen.

[0018] The generative AI unit 16, which can also be referred to as an LLM unit, executes a sentence generation process using generative AI (LLM). The generative AI unit 16 includes a training data storage unit 20, a model storage unit 22, an answer generation unit 24, a probability calculation unit 26, a determination unit 28, and an information extraction unit 30. The training data storage unit 20 stores training data used for training the generative AI. The model storage unit 22 stores data of the training model (which can also be referred to as a mathematical model or function approximator) of the generative AI.

[0019] The answer generation unit 24 generates answer sentences on a token-by-token basis, based on the prompt input from the processing unit 18, using the learning model of the generation AI stored in the model storage unit 22. The probability calculation unit 26 uses publicly known technology to calculate a probability distribution for multiple word candidates that can be selected for the token being processed, in other words, calculates the selection probability of each of the multiple word candidates.

[0020] The determination unit 28 determines whether a predetermined interruption condition is satisfied for each token during the process in which the answer generation unit 24 generates an answer sentence. If the determination unit 28 determines that the interruption condition is satisfied, the answer generation unit 24 interrupts the generation of the answer sentence. The interruption condition in the first embodiment is a condition related to the probability distribution of multiple word candidates calculated by the probability calculation unit 26. Specifically, the interruption condition is satisfied when at least one of the following conditions 1 and 2 is satisfied.

[0021] Condition 1: This condition is met when the selection probability for each of multiple word candidates that can be selected for each token is below a predetermined first threshold for all of the word candidates. The first threshold may be, for example, 5%. Condition 2: This condition is met when the difference in selection probability for each of multiple word candidates that can be selected for each token is below a predetermined second threshold. The second threshold may be, for example, 1%. The thresholds described in the present disclosure may be appropriate values ​​determined based on the knowledge of the system developer or the results of experiments using the information processing device 10.

[0022] When the determination unit 28 determines that the interruption condition is met, the information extraction unit 30 identifies and extracts information about the status of answer sentence generation by the generation AI (answer generation unit 24). The information extraction unit 30 identifies information about the status of answer sentence generation based on information about the token determined to meet the interruption condition.

[0023] The processing unit 18 controls learning of the generation AI, instructs the generation AI to generate answer sentences, acquires answer sentences generated by the generation AI, performs user interface processing, etc. The processing unit 18 includes a generation AI control unit 40, a generation AI output acquisition unit 42, a question generation unit 44, a display control unit 46, and a learning control unit 48.

[0024] The generation AI control unit 40 controls the generation AI unit 16. For example, the generation AI control unit 40 inputs to the generation AI unit 16 a prompt received by the input receiving unit 12 and an instruction to generate an answer sentence based on the prompt.

[0025] The generation AI output acquisition unit 42 acquires information (hereinafter also referred to as "generation AI output information") output from the generation AI unit 16. The generation AI output information includes an answer sentence by the generation AI, and also includes information regarding the status of the generation of the answer sentence by the generation AI.

[0026] The display control unit 46 controls the content to be displayed on the display unit 14. The display control unit 46 causes the display unit 14 to display the generation AI output information acquired by the generation AI output acquisition unit 42. For example, the display control unit 46 causes the display unit 14 to display an answer sentence by the generation AI, and also causes the display unit 14 to display information regarding the status of the generation of the answer sentence by the generation AI.

[0027] The question generation unit 44 generates a question requesting information necessary for generating an answer sentence, based on information regarding the status of the generation of an answer sentence by the generation AI, acquired by the generation AI output acquisition unit 42. The display control unit 46 displays information about the question generated by the question generation unit 44 on the display unit 14.

[0028] The input receiving unit 12 receives input of information necessary for generating an answer sentence. When the input of information necessary for generating an answer sentence is received, the learning control unit 48 generates learning data for the generation AI based on the input.

[0029] 3 may be implemented in a computer program and installed in the storage of the information processing device 10. The processor of the information processing device 10 may load the computer program into main memory and execute it to perform the functions of the functional blocks implemented in the computer program.

[0030] The multiple functional blocks of the information processing device 10 shown in FIG. 3 may be distributed across multiple devices. For example, the functions of the generation AI unit 16 and the functions of the processing unit 18 may be implemented in different devices connected via a communication network including a LAN, a WAN, the Internet, etc. Also, the functions of the generation AI unit 16 may be implemented in a server on the cloud, while the functions of the processing unit 18 may be implemented in a user terminal (such as a PC or smartphone). Also, some of the functional blocks of the generation AI unit 16 shown in FIG. 3 may be arranged in the processing unit 18, and conversely, some of the functional blocks of the processing unit 18 shown in FIG. 3 may be arranged in the generation AI unit 16. For example, the learning data storage unit 20 may be arranged in the processing unit 18.

[0031] The operation of the information processing device 10 of the first embodiment will be described. FIG. 4 is a flowchart showing the operation of the information processing device 10 of the first embodiment. The information processing device 10 provides a writing assistant service to a user. The input receiving unit 12 passes information input by a user to an input means such as a keyboard to the processing unit 18. The generation AI control unit 40 of the processing unit 18 waits until a prompt is input (N in S10). When a prompt is input (Y in S10), the generation AI control unit 40 inputs the input prompt to the generation AI unit 16 (S11). The generation AI unit 16 executes an answer generation process based on the input prompt (S12).

[0032] 5 is a flowchart showing details of the answer generation process of S12 in FIG. 4. The answer generation unit 24 tokenizes the input prompt and inputs it into the learning model of the generation AI. The answer generation unit 24 generates an answer sentence to the prompt on a token-by-token basis. In the process of generating the answer sentence, the probability calculation unit 26 calculates the selection probability of each of multiple candidate words in the next token (i.e., the target token for the word selection process) using publicly known technology (S20). For example, the probability calculation unit 26 may calculate the probability distribution of candidate words for the next token by inputting a vector of the dimension of the number of candidate words into a softmax function.

[0033] The determination unit 28 determines whether the probability distribution of the candidate words for the next token calculated in S20 satisfies the interruption conditions (the above-described conditions 1 and 2 in the first embodiment). If the probability distribution of the candidate words for the next token does not satisfy the interruption conditions (if both conditions 1 and 2 are satisfied in the first embodiment) (N in S21), the answer generation unit 24 determines a word for the next token from among the multiple candidate words. Specifically, the answer generation unit 24 selects the candidate word with the highest selection probability from among the multiple candidate words as the word for the next token (S22).

[0034] When the answer generation unit 24 selects the word of the last token, it generates an end token. If an end token has not been generated (N in S23), the process returns to S20 and the next token is processed. When an end token has been generated (Y in S23), the answer generation unit 24 outputs an answer sentence including the end token to the processing unit 18 (S24).

[0035] If the probability distribution of candidate words for the next token satisfies the interruption condition (in the first embodiment, if at least one of Condition 1 and Condition 2 is satisfied) (Y of S21), the answer generation unit 24 interrupts the process of determining words for the next token, i.e., interrupts the generation of the answer sentence. The answer generation unit 24 passes the answer sentence up to the position where the word has been determined (hereinafter also referred to as the "sentence in progress") to the information extraction unit 30 (S25).

[0036] The information extraction unit 30 generates information (hereinafter also referred to as "missing information") about the token that satisfied the interruption condition (the token that caused the interruption of the answer sentence generation) as information about the status of the generation of the answer sentence. In the case of FIG. 2, the missing information includes information about "XX", and may include, for example, the position of "XX" in the sentence, the element name, the part of speech name, etc. The information extraction unit 30 may also include the generated sentence (i.e., the sentence in progress) before the token that satisfied the interruption condition, such as "My hobby is XX", in the missing information. The information extraction unit 30 outputs the generated missing information to the processing unit 18 (S26).

[0037] Returning to FIG. 4 , if the output information from the generation AI unit 16 includes an end token (Y in S13), the display control unit 46 displays the answer sentence generated by the generation AI on the display unit 14 (S14). If the output information from the generation AI unit 16 does not include an end token (N in S13), i.e., if the output information from the generation AI unit 16 includes missing information, the question generation unit 44 of the processing unit 18 acquires the missing information (S15). Based on the acquired missing information, the question generation unit 44 generates a question for the user to obtain information necessary for answer sentence generation (e.g., selecting a word in a token that satisfies the interruption condition). The display control unit 46 displays the question for the user generated by the question generation unit 44 on the display unit 14 (S16).

[0038] A method for generating a question for a user will be described below. Here, the prompt input by the generation AI control unit 40 to the generation AI unit 16 (hereinafter also referred to as the "input prompt") is "Please give me an example sentence for introducing myself," and the missing information output from the generation AI unit 16 is "My hobby is ____." Of the missing information, "____" is the information of the token that satisfies the interruption condition, and "My hobby is" is the sentence in progress.

[0039] Generation Method 1: The question generation unit 44 may generate a question for the user using a predetermined template. The template may be, for example, "In response to the input '%input prompt%', the assistant generated an answer up to '%sentence in progress%', but is not sure about the words that follow. Please tell me the correct answer." % represents a variable. In this case, the question generation unit 44 may generate the question sentence "In response to the input 'Please tell me an example sentence for introducing myself.', the assistant generated an answer up to 'My hobbies are,' but is not sure about the words that follow. Please tell me the correct answer."

[0040] Generation Method 2: The question generation unit 44 may generate a prompt (hereinafter also referred to as a "question generation instruction prompt") instructing the generation AI to generate a question for the user by inputting the prompt to the generation AI. The question generation instruction prompt may be, for example, "In response to the input '%input prompt%', the assistant generated an answer up to '%sentence in progress%', but is not confident about the words that follow. Please output a question sentence to ask the user for the correct answer to the words that follow." Note that the generation AI that generates a question for the user may be the same as the generation AI that generates an answer sentence, or it may be a different generation AI from the generation AI that generates the answer sentence. For example, the generation AI that creates the answer sentence may be located on a cloud server, while the generation AI that generates a question for the user may be implemented on a user terminal (such as a PC or smartphone).

[0041] Modifications of generation method 1 and generation method 2 will be described. The information extraction unit 30 of the generation AI unit 16 may output missing information including multiple candidate words (such as "walk," "music," and "movie" in the example of FIG. 2) in the token that satisfies the interruption condition. The question generation unit 44 may include multiple candidate words indicated by the missing information in the question generated by generation method 1 or 2. For example, the question generation unit 44 may generate the following question in response to the input of the question "Please tell me an example sentence for introducing myself." The assistant generated an answer up to "My hobbies are," but is not sure about the words that follow. Does this include the correct answer? [walk, music, movie]."

[0042] Returning to the explanation of FIG. 4 , the user inputs an answer to the question ("music" in the example of FIG. 2 ) into a predetermined input means. The input accepting unit 12 accepts the input of the user's answer. The learning control unit 48 waits until the user's answer to the question is input (N in S17). When the user's answer to the question is input (Y in S17), the learning control unit 48 controls the learning process based on the input user's answer (S18). The learning control unit 48 generates learning data for the generation AI based on the user's answer.

[0043] An example of learning process control of S18 is shown. The learning control unit 48 may generate learning data for fine-tuning the generation AI based on a pair of an input prompt and a user's answer, and store the generated learning data in the learning data storage unit 20. Alternatively, the learning control unit 48 may generate learning data for fine-tuning the generation AI based on a pair of a question to the user and the user's answer, and store the generated learning data in the learning data storage unit 20. The learning control unit 48 may regenerate a learning model of the generation AI using the learning data for fine-tuning stored in the learning data storage unit 20.

[0044] Another example of the learning process control of S18 is shown below. The learning control unit 48 may generate learning data based on pairs of input prompts and user responses, and store the generated learning data in a Retrieval Augmented Generation (RAG) database applicable to the generation AI. Alternatively, the learning control unit 48 may generate learning data based on pairs of questions to the user and the user responses, and store the generated learning data in a RAG database applicable to the generation AI. The RAG database is a database that searches for and extracts information necessary for the generation AI to generate answer sentences.

[0045] According to the information processing device 10 of the first embodiment, if the accuracy of the answer obtained using the generation AI is low, the generation of an answer sentence can be interrupted, thereby preventing the generation of an answer sentence that is inaccurate from the facts. Furthermore, according to the information processing device 10, when the generation of an answer sentence using the generation AI is interrupted, information regarding the token that caused the interruption of the answer sentence generation is presented to the user as information regarding the status of the answer sentence generation. This makes it easier to obtain the information necessary for generating an answer sentence from the user, making it easier to obtain an appropriate answer sentence using the generation AI.

[0046] The present disclosure has been described above based on the first embodiment. The first embodiment is an example, and it will be understood by those skilled in the art that various modifications are possible in the combination of each component or each treatment process of the first embodiment, and that such modifications are also within the scope of the present disclosure.

[0047] A modified example of the first embodiment will be described. After the learning process based on the user's answer in S18 of FIG. 4, the generation AI control unit 40 of the processing unit 18 may instruct the generation AI unit 16 to resume generating answer sentences for tokens determined to satisfy the interruption condition. In accordance with this instruction, the answer generation unit 24 of the generation AI unit 16 may resume the process of generating answer sentences for tokens determined to satisfy the interruption condition (the process of S12 of FIG. 4). This modified example makes it easier to obtain appropriate answer sentences using the generation AI that has undergone learning process based on the user's answer to the question.

[0048] Second Embodiment A second embodiment of the present disclosure will be described, focusing on differences from the first embodiment. Components of the second embodiment that are the same as or equivalent to those of the first embodiment will be assigned the same reference numerals as those of the first embodiment, and descriptions that overlap with those of the first embodiment will be omitted as appropriate.

[0049] The information processing apparatus 10 of the second embodiment includes the same functional blocks as the information processing apparatus 10 of the first embodiment shown in FIG.

[0050] 6 is a flowchart showing the operation of the information processing device 10 of the second embodiment. The processes from S30 to S36 in FIG. 6 are the same as the processes from S10 to S16 in FIG. 4, and therefore will not be described again. The input accepting unit 12 accepts input of a user's answer ("music" in the example of FIG. 2). The generation AI control unit 40 of the processing unit 18 waits until the user's answer to the question is input (N in S37).

[0051] When the user inputs an answer to the question (Y in S37), the generation AI control unit 40 instructs the generation AI unit 16 to resume generating answer sentences for the tokens determined to satisfy the interruption condition based on the input user answer (S38). This instruction includes the correct words indicated by the user answer, i.e., the correct words for the tokens determined to satisfy the interruption condition.

[0052] Returning to S32, the answer generation unit 24 of the generation AI unit 16 resumes generating answer sentences from the token determined to satisfy the interruption condition onwards, in accordance with instructions from the generation AI control unit 40. The answer generation unit 24 sets the correct word indicated by the user's answer as the word in the token determined to satisfy the interruption condition. Thereafter, the answer generation unit 24 selects words for tokens after the token determined to satisfy the interruption condition based on the probability distribution of multiple word candidates.

[0053] In the second embodiment, the generation of the answer sentence is completed to the end, and the final answer sentence is displayed in S34, after which the learning process is performed. That is, if insufficient information occurs during the generation of the answer sentence (Y in S39), similar to the process of S18 in FIG. 4, the learning control unit 48 of the processing unit 18 controls the learning process based on the user's answer to the question, and also generates learning data for the generation AI based on the user's answer (S40). The details of the learning process control are the same as in the first embodiment. If insufficient information does not occur during the generation of the answer sentence (N in S39), the process of S40 is skipped.

[0054] The information processing device 10 of the second embodiment also achieves the same effects as the information processing device 10 of the first embodiment. That is, according to the information processing device 10 of the second embodiment, it becomes easier to obtain an appropriate answer sentence from the generation AI based on a user's input to a question, and the accuracy of the answer by the generation AI can be improved by learning based on the user's input to the question.

[0055] The present disclosure has been described above based on the second embodiment. The second embodiment is merely an example, and it will be understood by those skilled in the art that various modifications are possible in the combination of each component or each treatment process of the second embodiment, and that such modifications are also within the scope of the present disclosure.

[0056] A modified example applicable to the first and second embodiments described above will be described. When the determination unit 28 determines that a predetermined interruption condition is met (i.e., when the answer generation unit 24 is not confident that the answer content is correct), the answer generation unit 24 of the information processing device 10 may not interrupt the generation of an answer; in other words, may continue generating an answer. In this case, the answer generation unit 24 may generate an answer that is different from the answer generated when the determination unit 28 determines that the predetermined interruption condition is not met (i.e., when the answer generation unit 24 is confident that the answer content is correct). The different answer may have a different writing style (including expressions, endings, etc.; the same applies below) from the answer generated when the answer is confident.

[0057] For example, when it is determined that the interruption condition is satisfied, the answer generation unit 24 may include, in the answer content, a predetermined mark, icon, or sentence indicating that the answer is not confident, or may include, in the answer content, a predetermined mark, icon, or sentence indicating that information for generating an answer is insufficient, unlike when it is determined that the interruption condition is not satisfied. Additionally or alternatively, the answer generation unit 24 may include, in the answer content, a number indicating the selection probability calculated for the word (token) determined to satisfy the interruption condition.

[0058] For example, even if it is determined that the interruption condition is satisfied, the answer generation unit 24 may continue generating an answer to the end (until completion), and then input another prompt to the generation AI unit 16 to rewrite the generated answer sentence into a style that indicates (or suggests) that the answer is not confident. The rewriting target may be the entire sentence including the word (token) about which the user is not confident in its correctness, or may be only the sentence including the word (token) up to a delimiter such as a punctuation mark, or may be only the word (token) about which the user is not confident.

[0059] For example, if the token "walk" in the answer sentence "My hobby is walking," satisfies the interruption condition, the answer generation unit 24 may rewrite it to one of the following answer sentences: "Maybe my hobby is walking." "Maybe my hobby is walking?" "My hobby is walking?"

[0060] For example, the model storage unit 22 of the information processing device 10 may store a model of the generation AI to be used when it is determined that the interruption condition is not satisfied (herein referred to as a "normal model"), and a model of the generation AI to be used when it is determined that the interruption condition is satisfied (herein referred to as a "special model"). The special model may be a model that is fine-tuned to generate answer sentences in a style that indicates (or suggests) a lack of confidence in the answer content. When it is determined that the interruption condition is satisfied, the answer generation unit 24 may switch the model of the generation AI that generates answers from the normal model to the special model, and resume (in other words, continue) generating answer sentences.

[0061] After switching the model of the generation AI that generates an answer from the normal model to the special model, the answer generation unit 24 may continue to use the special model to generate the answer sentence to the end. Alternatively, if the answer generation unit 24 determines that the interruption condition is not satisfied for another token to be processed after switching the model of the generation AI that generates an answer from the normal model to the special model, the answer generation unit 24 may switch the model of the generation AI that generates an answer from the special model back to the normal model, in other words, return to the normal model and continue generating the answer sentence.

[0062] Any combination of the above-described embodiments and modifications is also useful as an embodiment of the present disclosure. A new embodiment resulting from a combination will have the effects of the combined embodiments or modifications. It will also be understood by those skilled in the art that the functions to be performed by each component recited in the claims can be realized by each component shown in the embodiments and modifications, either alone or in combination.

[0063] <Supplementary Notes> The above description of the embodiments and variations discloses the following technologies. [Technology 1] An information processing method in a computer, comprising: generating an answer sentence on a token-by-token basis using a generation AI based on an input prompt; determining whether a predetermined condition is satisfied for each token during the process of generating the answer sentence, and suspending the generation of the answer sentence if it is determined that the predetermined condition is satisfied; and outputting information related to the status of the generation of the answer sentence if it is determined that the predetermined condition is satisfied. According to this information processing method, when the generation of an answer sentence using the generation AI is suspended, information related to the status of the generation of the answer sentence is output, making it easier to obtain information necessary for generating an answer sentence from the user and to obtain an appropriate answer sentence using the generation AI. [Technology 2] The information processing method described in Technology 1, in which the predetermined condition is a condition related to the probability distribution of multiple word candidates. According to this information processing method, the generation of an answer sentence can be suspended if the accuracy of an answer using the generation AI is low, thereby preventing the generation of an answer sentence that is inaccurate. [Technology 3] The information processing method described in Technology 2, wherein the predetermined condition is that the selection probability for each of a plurality of word candidates that can be selected for each token is below a predetermined threshold for each of the word candidates. According to this information processing method, it is possible to suspend the generation of an answer sentence when the accuracy of the answer using the generation AI is low, thereby preventing the generation of an answer sentence that is inaccurate. [Technology 4] The information processing method described in Technology 2, wherein the predetermined condition is that the difference in selection probability for each of a plurality of word candidates that can be selected for each token is below a predetermined threshold. According to this information processing method, it is possible to suspend the generation of an answer sentence when the accuracy of the answer using the generation AI is low, thereby preventing the generation of an answer sentence that is inaccurate. [Technology 5] The information processing method described in Technology 1, wherein information on the status of the generation of the answer sentence is identified based on information about a token that is determined to satisfy the predetermined condition. According to this information processing method, by outputting information generated based on information about the token that caused the generation of the answer sentence to be suspended, it is easier to obtain information necessary for generating the answer sentence from the user.[Technology 6] The information processing method according to Technology 1 further includes generating and outputting a question requesting information necessary for generating the answer sentence, based on information regarding the status of generation of the answer sentence. This information processing method makes it easier to obtain information necessary for generating the answer sentence from a user. [Technology 7] The information processing method according to Technology 6 further includes accepting input regarding information necessary for generating the answer sentence, and, when the input is accepted, generating learning data for the generation AI based on the input. This information processing method makes it possible to improve the accuracy of answers provided by the generation AI through learning based on the user's input in response to a question. [Technology 8] The information processing method according to Technology 7 further includes restarting generation of the answer sentence for tokens determined to satisfy the predetermined condition, based on the generated learning data. This information processing method makes it easier to obtain appropriate answer sentences from the generation AI. [Technology 9] The information processing method described in Technology 6 further includes: accepting input regarding information required for generating the answer sentence; when the input is accepted, resuming generation of the answer sentence for tokens determined to satisfy the predetermined condition based on the input; and when generation of the answer sentence is completed, generating training data for the generation AI based on the input. This information processing method makes it easier to obtain an appropriate answer sentence from the generation AI, and also improves the accuracy of answers provided by the generation AI through learning based on user input for questions. [Technology 10] The information processing method described in any of Technology 7 to 9, wherein the training data is training data for fine-tuning the generation AI, generated based on a pair of the prompt and the input, or a pair of the question and the input. This information processing method improves the accuracy of answers provided by the generation AI based on user input for questions.[Technology 11] The information processing method according to any one of Technology 7 to Technology 9, wherein the training data is training data used to build a Retrieval-Augmented Generation (RAG) database applicable to the generation AI, generated based on a pair of the prompt and the input, or a pair of the question and the input. This information processing method can improve the accuracy of answers using the generation AI based on user input for questions. [Technology 12] The information processing method according to Technology 1, wherein the information on the status of the generation of the answer sentence is information about tokens determined to satisfy the predetermined condition, and includes resuming the generation of the answer sentence for the token determined to satisfy the predetermined condition using a second model different from the first model used in the process of generating the answer sentence. This information processing method outputs information about the token determined to satisfy the predetermined condition, and can further output a final answer sentence even if the predetermined condition is satisfied. [Technology 13] The information processing method according to Technology 12, wherein the second model is a model that generates the answer sentence using an expression suggesting lack of confidence in the content of the answer sentence. According to this information processing method, it is possible to notify the output destination (user, etc.) that the user is not confident in the content of the answer sentence. [Technology 14] The information processing method according to Technology 12 or 13, wherein, after resuming generation of the answer sentence using the second model, if it is determined that the subsequent token does not satisfy the predetermined condition, the system switches back to the first model to generate the answer sentence. According to this information processing method, an answer sentence can be generated using an appropriate model depending on whether the token to be processed satisfies the predetermined condition. [Technology 15] An information processing device comprising: an answer generation unit that generates an answer sentence on a token-by-token basis using a generation AI based on an input prompt; a determination unit that determines whether the predetermined condition is satisfied for each token in the process of generating the answer sentence; and an output unit, wherein the answer generation unit suspends generation of the answer sentence if it determines that the predetermined condition is satisfied; and the output unit outputs information regarding the status of generation of the answer sentence if it determines that the predetermined condition is satisfied.According to this information processing device, when the generation of an answer sentence using the generation AI is interrupted, information regarding the status of the generation of the answer sentence is output, making it easier to obtain the information necessary for generating the answer sentence from the user and to obtain an appropriate answer sentence using the generation AI. [Technology 16] A computer program for causing a computer to execute the following: generating an answer sentence on a token-by-token basis using the generation AI based on an input prompt; determining whether or not a predetermined condition is satisfied for each token in the process of generating the answer sentence, and if it is determined that the predetermined condition is satisfied, interrupting the generation of the answer sentence; and if it is determined that the predetermined condition is satisfied, further outputting information regarding the status of the generation of the answer sentence. According to this computer program, when the generation of an answer sentence using the generation AI is interrupted, information regarding the status of the generation of the answer sentence is output, making it easier to obtain the information necessary for generating the answer sentence from the user and to obtain an appropriate answer sentence using the generation AI.

[0064] The present disclosure relates to data processing technology, and more particularly to an information processing method, an information processing device, and a computer program.

[0065] 10 Information processing device, 12 Input receiving unit, 14 Display unit, 24 Answer generation unit, 26 Probability calculation unit, 28 Determination unit, 30 Information extraction unit, 40 Generation AI control unit, 44 Question generation unit, 46 Display control unit, 48 Learning control unit.

Claims

1. An information processing method in a computer, comprising: generating an answer sentence on a token-by-token basis using a generation AI based on an input prompt; determining whether or not a predetermined condition is met for each token in the process of generating the answer sentence, and suspending the generation of the answer sentence if it is determined that the predetermined condition is met; and outputting information relating to the status of the generation of the answer sentence if it is determined that the predetermined condition is met.

2. The information processing method according to claim 1, wherein the predetermined condition is a condition relating to a probability distribution of a plurality of word candidates.

3. The information processing method according to claim 2, wherein the predetermined condition is that the selection probability for each of a plurality of word candidates that can be selected for each token is below a predetermined threshold for each of the word candidates.

4. The information processing method according to claim 2, wherein the predetermined condition is that the difference in selection probability between each of a plurality of word candidates that can be selected for each token is below a predetermined threshold.

5. The information processing method according to claim 1, wherein the information relating to the status of the generation of the answer sentence is identified based on information about a token determined to satisfy the predetermined condition.

6. The information processing method according to claim 1, further comprising generating and outputting a question for requesting information necessary for generating the answer sentence based on information relating to the status of generation of the answer sentence.

7. The information processing method according to claim 6, further comprising: accepting input of information required to generate the answer sentence; and, when the input is accepted, generating learning data for the generation AI based on the input.

8. The information processing method according to claim 7, further comprising restarting generation of the answer sentence for a token determined to satisfy the predetermined condition based on the generated learning data.

9. The information processing method of claim 6, further comprising: accepting input of information required for generating the answer sentence; when the input is accepted, resuming generation of the answer sentence for tokens determined to satisfy the predetermined condition based on the input; and when generation of the answer sentence is completed, generating learning data for the generation AI based on the input.

10. An information processing method according to any one of claims 7 to 9, wherein the learning data is learning data for fine-tuning the generative AI, generated based on a pair of the prompt and the input, or a pair of the question and the input.

11. An information processing method according to any one of claims 7 to 9, wherein the training data is training data used to construct a RAG (Retrieval-Augmented Generation) database applicable to the generation AI, generated based on a pair of the prompt and the input, or a pair of the question and the input.

12. The information processing method of claim 1, wherein the information regarding the status of the generation of the answer sentence is information about a token determined to satisfy the specified condition, and the method includes resuming the generation of the answer sentence for the token determined to satisfy the specified condition using a second model different from the first model used in the process of generating the answer sentence.

13. The information processing method according to claim 12, wherein the second model is a model that generates the answer sentence using an expression that suggests a lack of confidence in the content of the answer sentence.

14. An information processing method as described in claim 12 or 13, wherein after resuming generation of the answer sentence using the second model, if it is determined that the subsequent token does not satisfy the specified condition, the answer sentence is generated by switching back to the first model.

15. An information processing device comprising: an answer generation unit that generates answer sentences on a token-by-token basis using generation AI based on an input prompt; a judgment unit that judges whether or not a predetermined condition is met for each token in the process of generating the answer sentence; and an output unit, wherein the answer generation unit suspends generation of the answer sentence if it is determined that the predetermined condition is met, and the output unit outputs information regarding the status of generation of the answer sentence if it is determined that the predetermined condition is met.

16. A computer program for causing a computer to perform the following steps: generate answer sentences on a token-by-token basis using generation AI based on an input prompt; determine whether or not a predetermined condition is met for each token in the process of generating the answer sentences, and if it is determined that the predetermined condition is met, suspend the generation of the answer sentences; and if it is determined that the predetermined condition is met, further output information regarding the status of the generation of the answer sentences.

Images