Information processing device, information processing method, and recording medium

Abstract

An information processing device according to the present disclosure comprises a generating unit, a first processing unit, and a second processing unit. The generating unit generates first input information on the basis of a plurality of pieces of material information and relationships between the plurality of pieces of material information. The first processing unit accepts input of the first input information and executes first processing employing a first model that outputs first output information. The second processing unit accepts input of second input information including the first output information, and executes second processing employing a second model that outputs second output information.

Description

Information processing device, information processing method, and recording medium

[0001] This disclosure relates to an information processing device, an information processing method, a program, and a recording medium.

[0002] Technology related to this disclosure is disclosed in Patent Document 1. Patent Document 1 discloses a technology that adds information to a prompt entered by a user using RAG (Retrieval Augmented Generation) technology, and processes the prompt with added information using Generative AI (Artificial Intelligence).

[0003] Patent No. 7530134

[0004] By inputting appropriate information into the model, the desired answer can be obtained from that model. One example of the purpose of this disclosure is to provide a new technique for inputting appropriate information into a model.

[0005] According to one aspect of this disclosure, a program is provided that causes a computer to function as: a generation means for generating first input information based on a plurality of material information and the relationships between the plurality of material information; a first processing means for receiving the input of the first input information and executing a first process using a first model that outputs first output information; and a second processing means for receiving the input of second input information including the first output information and executing a second process using a second model that outputs second output information.

[0006] Furthermore, according to one aspect of this disclosure, an information processing device is provided, comprising: generation means for generating first input information based on a plurality of material information and the relationships between the plurality of material information; first processing means for receiving the input of the first input information and executing a first process using a first model that outputs first output information; and second processing means for receiving the input of second input information including the first output information and executing a second process using a second model that outputs second output information.

[0007] Furthermore, according to one aspect of this disclosure, an information processing method is provided in which one or more computers generate first input information based on a plurality of material information and the relationships between the plurality of material information, execute a first process using a first model that receives the input of the first input information and outputs first output information, and execute a second process using a second model that receives the input of second input information including the first output information and outputs second output information.

[0008] According to one example of this disclosure, a new technology is realized for inputting appropriate input information into a model.

[0009] Figure 1 is a diagram showing an example of a functional block diagram of an information processing device. Figure 2 is a flowchart showing an example of the processing flow of an information processing device. Figure 3 is a diagram illustrating an example of the overall structure of an information processing device. Figure 4 is a diagram showing an example of the hardware configuration of an information processing device. Figure 5 is a flowchart showing another example of the processing flow of an information processing device. Figure 6 is a diagram illustrating a comparative example.

[0010] The embodiments of this disclosure will be described below with reference to the drawings. In this disclosure, the drawings are associated with one or more embodiments. In all drawings, similar components are denoted by the same reference numerals, and their descriptions are omitted where appropriate.

[0011] <<First Embodiment>> Figure 1 is a functional block diagram showing an overview of the information processing device 10. Figure 2 is a flowchart showing an example of the processing flow executed by the information processing device 10.

[0012] As shown in Figure 1, the information processing device 10 includes a generation unit 11, a first processing unit 12, and a second processing unit 13. These functional units execute the processes shown in the flowchart of Figure 2.

[0013] In S10, the generation unit 11 generates first input information based on multiple material information and the relationships between the multiple material information. In S11, the first processing unit 12 executes a first process using the first model. The first model receives the input of the first input information and outputs the first output information. In S12, the second processing unit 13 executes a second process using the second model. The second model receives the input of second input information including the first output information and outputs the second output information.

[0014] In this manner, the information processing device 10 executes the first process and the second process. The second process is a process using a second model, which is, for example, a generative AI such as a large-scale language model. The first process is a process for generating second input information to be input to the second model, and is executed using the first model. When the first model receives the input of the first input information, it searches for information related to the first input information and outputs the search result (first output information). The information processing device 10 generates second input information that includes the first output information output from the first model.

[0015] The information processing device 10 generates, by a distinctive means, first input information to be input into the first model as described above, that is, "a query for searching for information to be included in the second input information to be input into the second model." Specifically, the information processing device 10 generates first input information based on multiple material information and the relationships between the multiple material information. With such an information processing device 10, it is possible to generate first input information not simply by processing multiple material information, but by processing multiple material information while considering the relationships between the multiple material information. For example, multiple material information can be weighted based on the relationships between the multiple material information, or some of the multiple material information can be excluded from processing based on the relationships between the multiple material information.

[0016] According to the information processing device 10 that generates the first input information using such a distinctive process, the information to be included in the second input information can be appropriately searched in the first process. Then, by inputting the second input information, which includes the appropriate information, into the second model and processing it, the desired output can be obtained from the second model.

[0017] Thus, according to the information processing device 10 of the first embodiment, a new technology is realized for inputting appropriate input information (second input information) into a model (second model).

[0018] <<Second Embodiment>> <Overview> The information processing device 10 of the second embodiment is a concrete implementation of the configuration of the information processing device 10 of the first embodiment.

[0019] Using Figure 3, an example of the overall structure and overview of the processing performed by the information processing device 10 will be explained.

[0020] The information processing device 10 is used, for example, to determine the content of the computer's statements in a chat service where a computer and a human converse. The information processing device 10 is preferably used in so-called multi-turn conversations. The content of the computer's statements is generated using a generative AI such as a large-scale language model. By inputting a prompt into the generative AI, the content of the computer's statements (output) is generated.

[0021] By generating an appropriate prompt and inputting it into the generating AI, an appropriate response can be obtained from the generating AI. The information processing device 10 can generate prompts that include past conversation history, extracted information, and instruction sentences, for example, as shown in the figure.

[0022] The information processing device 10 can generate "extracted information" to be included in the prompt through a distinctive process. Specifically, the information processing device 10 generates a query to search for information to be included in the prompt based on the content of multiple past statements and the relationships between the multiple statements. In the example shown in Figure 3, the information processing device 10 assigns a weight (W shown in the figure) to each of the multiple statements based on the relationships between the multiple statements. １ Or W ３The information processing device 10 then integrates multiple statements, taking these weights into consideration, to generate a query for searching for information to include in the prompt. The information processing device 10 searches for and extracts information related to the query generated in this way, and includes the extracted information in the prompt. The configuration of such an information processing device 10 will be described in detail below.

[0023] <Hardware Configuration> First, an example of the hardware configuration of the information processing device 10 will be described. Each functional unit of the information processing device 10 is realized by any combination of hardware and software. Those skilled in the art will understand that there are various variations in the implementation method and the device. The software includes programs that are pre-installed at the time of shipment of the device, as well as programs downloaded from recording media such as CDs (Compact Discs) or from servers on the Internet.

[0024] Figure 4 is a block diagram illustrating the hardware configuration of the information processing device 10. As shown in Figure 4, the information processing device 10 includes a processor 1A, memory 2A, input / output interface 3A, peripheral circuitry 4A, and bus 5A. The peripheral circuitry 4A includes various modules. The information processing device 10 does not necessarily have peripheral circuitry 4A. The information processing device 10 may also be composed of multiple physically and / or logically separated devices. In this case, each of the multiple devices may have the above hardware configuration.

[0025] Bus 5A is a data transmission path for the processor 1A, memory 2A, peripheral circuit 4A, and input / output interface 3A to send and receive data to and from each other. The processor 1A is a processing unit such as a CPU (Central Processing Unit) or GPU (Graphics Processing Unit). Memory 2A is a memory such as RAM (Random Access Memory) or ROM (Read Only Memory). The input / output interface 3A includes interfaces for acquiring information from input devices, external devices, external servers, external sensors, cameras, etc., and interfaces for outputting information to output devices, external devices, external servers, etc. The input / output interface 3A also includes interfaces for connecting to a communication network such as the Internet. Input devices include, for example, a keyboard, mouse, microphone, physical buttons, touch panel, etc. Output devices include, for example, a display, projection device, speaker, printer, mailer, etc. The processor 1A can issue commands to each module and perform calculations based on their calculation results.

[0026] The information processing device 10 may be a server. Alternatively, the information processing device 10 may be a user terminal. Examples of user terminals include, but are not limited to, personal computers, smartphones, tablet devices, mobile phones, game consoles, wearable devices, home appliances, and in-vehicle devices. The functions of the information processing device 10 described below are realized by installing a server, a user terminal, and a predetermined program.

[0027] <Functional Configuration> Next, the functional configuration of the information processing device 10 will be described in detail. Figure 1 is an example of a functional block diagram of the information processing device 10. As shown in the figure, the information processing device 10 has a generation unit 11, a first processing unit 12, and a second processing unit 13.

[0028] The generation unit 11 generates first input information based on multiple material information and the relationships between the multiple material information.

[0029] "Multiple material information" refers to information used to generate the first input information. For example, in a chat service where a computer and a person converse, past statements can serve as material information. In a multi-turn conversation, for instance, the statements from one turn can be treated as a single piece of material information.

[0030] In the second embodiment, the number of material information items used to generate the first input information is predetermined. For example, a predetermined number of recent statements in a series of conversations on the chat service, i.e., a predetermined number of statements starting from the most recent, are used to generate the first input information. One statement can represent one turn's worth of statements. The predetermined number is exemplified by, but is not limited to, 3, 4, 5, etc. If, immediately after starting a conversation on the chat service, the number of statements made up to that point is less than the predetermined number, all of the statements made up to that point are used to generate the first input information.

[0031] Thus, in the second embodiment, except for the immediate aftermath of the conversation, the number of material pieces (number of statements) used to generate the first input information is fixed and does not change. The following embodiment will describe an example in which the number of material pieces used to generate the first input information changes.

[0032] "The relationships between multiple pieces of information" refers to the relationships between the timing of statements made by multiple pieces of information. For example, the relationships between multiple pieces of information can indicate the order in which statements are made by multiple pieces of information.

[0033] The generation unit 11 can acquire multiple material information and information indicating the relationships between the multiple material information from the means for providing the chat service. The information processing device 10 may also include means for providing the chat service. Alternatively, a device physically and / or logically separate from the information processing device 10 may also include means for providing the chat service. In this case, the information processing device 10 and the device with the means for providing the chat service are communicated with each other. The means for providing the chat service is implemented using widely known technology.

[0034] "Acquisition" includes at least one of the following: the self-device going to obtain data or information stored in other devices or storage media (active acquisition), and the self-device inputting data or information output from other devices (passive acquisition). Examples of active acquisition include making requests to other devices and receiving their responses, and accessing other devices or storage media to read data. Examples of passive acquisition include receiving information distributed (or transmitted, push-notified, etc.). Furthermore, acquisition may involve selecting and obtaining data or information from the received data or information, or selecting and receiving distributed data or information.

[0035] Next, the process of generating the first input information will be described.

[0036] First, based on the relationships among the plurality of material information, the generation unit 11 sets weights for each of the plurality of material information. The generation unit 11 assigns a greater weight to the material information indicating the speech content with a speech timing closer to the current time, that is, the newer speech content.

[0037] In the second embodiment, the number of material information used to generate the first input information and the weights assigned to each of the predetermined number of material information are predetermined. The generation unit 11 sets weights for each of the plurality of material information according to the predetermined content.

[0038] Here, using a specific example, the process of setting weights for each of the plurality of material information according to the predetermined content will be described. In this specific example, the number of material information used to generate the first input information is determined to be 3. And the weight of the material information indicating the latest speech content is W ３ , the weight of the material information indicating the speech content one turn before that is W ２ , the weight of the material information indicating the speech content one turn before that is W １ is determined. W ３ , W <000000​​​​​​​​Each value is determined so as to satisfy the relationship of "」. The generation unit 11 sets weights for each of the plurality of material information, for example, according to the content determined in advance in this way.

[0039] After setting weights for each of the plurality of material information, the generation unit 11 generates integrated information obtained by integrating the plurality of material information using the weights as the first input information. Specifically, as shown in FIG. 3, the generation unit 11 vectorizes each of the plurality of material information. Vectorization of text (text) can be realized by using a widely known technique. Then, the generation unit 11 calculates the weighted sum (integrated information) of the vectors of the plurality of material information using the weights of each of the plurality of material information.

[0040] Returning to FIG. 1, the first processing unit 12 executes a first process using the first model.

[0041] The "first model" is a model that, when receiving an input of the first input information, searches for information related to the first input information and outputs the search result as the first output information. As shown in FIG. 3, the first model can search a predetermined database for information related to the first input information.

[0042] The types of information related to the first input information include, but are not limited to, documents, images (including still images and moving images), audio, and other data. For example, a plurality of vectorized documents are stored in the database. In addition, the database may store images, audio, and other data to which vectorized tag information is added. The tag information can include descriptions and other related information of images, audio, and other data.

[0043] The first model searches for information related to the first input information from the database according to a predetermined search rule. For example, the first model may search for information whose similarity to the first input information is greater than or equal to a threshold value. In addition, the first model may search for information with the highest similarity to the first input information. In addition, the first model may search for a predetermined number of information from the ones with higher similarity to the first input information.

[0044] The "first process" is the process of inputting the first input information generated by the generation unit 11 into the first model and obtaining the search results (first output information) output from the first model.

[0045] Returning to Figure 1, the second processing unit 13 executes a second process using the second model.

[0046] The "second model" is a generative AI such as a large-scale language model. The second model may have the ability to process images, audio, and other data in addition to natural language. The second model accepts input of first output information and second input information including an instruction sentence, and outputs the response to the instruction sentence as second output information. The second model is implemented using widely known technologies. Generative AI is a concept that includes a large-scale language model (LLM). Generative AI may also include an image analysis model. Generative AI interprets the content of a prompt input in natural language, generates a response to the input prompt based on pre-trained data, and outputs the response expressed in natural language. Generative AI can learn a vast amount of various data, for example, that is publicly available on the internet. Generative AI can also accept input such as an image along with the prompt, interpret the content of the input image, and generate a response that reflects the result of the interpretation. Generative AI can also generate and output an image. In this embodiment, the information processing device 10 can input a prompt entered by an operator to the generative AI. Furthermore, the information processing device 10 of this embodiment can read prompts registered in advance by the operator and input them into the generating AI. Also, the information processing device 10 of this embodiment can read prompts from a plurality of prompts registered in advance by the operator according to predetermined rules and input them into the generating AI. Furthermore, the information processing device 10 of this embodiment can generate prompts by filling predetermined information into the blank spaces of a prompt (template) registered in advance by the operator and input them into the generating AI. The information processing device 10 of this embodiment may obtain the information to be filled into the blank spaces from another device (such as a server), generate it itself, or receive input from the operator. Also, the information processing device 10 of this embodiment can generate prompts according to predetermined rules and input the generated prompts into the generating AI. The information processing device 10 of this embodiment can input reference information to be referenced along with the prompts into the generating AI. The information processing device 10 of this embodiment can input reference information entered by the operator into the generating AI.Furthermore, the information processing device 10 of this embodiment can acquire reference information from other devices (such as servers) and input it into the generating AI. Also, the information processing device 10 of this embodiment can generate reference information according to predetermined rules and input the generated reference information into the generating AI. The generating AI is, for example, ChatGPT, Gemini, Firefly, Cloude, etc., but is not limited to these. The information processing device 10 of this embodiment may also include a generating AI. In addition, an external device configured to communicate with the information processing device 10 of this embodiment may also include a generating AI.

[0047] The "second process" is a process in which, after generating second input information based on the first output information generated by the first processing unit 12, the second input information is input to the second model, and the response (second output information) output from the second model is obtained.

[0048] Here, we will explain the process of generating the second input information. The second input information is the information that is input to the second model as described above, and is called a prompt.

[0049] The second processing unit 13 can generate a prompt that includes, for example, past conversation history, information extracted in the first processing (first output information), and instruction text, as shown in Figure 3.

[0050] "Past conversation history" refers to the content of past messages in a chat service where a computer and a human communicate.

[0051] For example, the second processing unit 13 may include a predetermined number of statements (material information) used by the generation unit 11 to generate the first input information in the prompt as past conversation history. As described above, the generation unit 11 can use a predetermined number of recent statements from a series of conversations on the chat service to generate the first input information. In this case, the second processing unit 13 can include a predetermined number of recent statements from a series of conversations on the chat service in the prompt as past conversation history.

[0052] As another example, the second processing unit 13 may include in the prompt, in addition to a predetermined number of statements (material information) used by the generation unit 11 to generate the first input information, other statements as further past conversation history. Other statements are statements from a series of conversations on the chat service that were not used to generate the first input information. For example, other statements are statements made before the most recent predetermined number of statements. The second processing unit 13 may include all statements made before the most recent predetermined number of statements as past conversation history in the prompt. Alternatively, the second processing unit 13 may include a portion of the statements made before the most recent predetermined number of statements as past conversation history in the prompt.

[0053] The "instruction text" is prepared in advance. The second processing unit 13 can include the pre-prepared instruction text in the prompt. For example, the instruction text could be, "Please decide on a response to the last statement in the 'past conversation history' by referring to the 'extracted information'," or it could be something else.

[0054] The second processing unit 13 can input the prompt (second input information) generated in this way into the second model and obtain the response (second output information) output from the second model. The second processing unit 13 can then input the second output information into the means for providing the chat service.

[0055] Next, an example of the processing flow of the information processing device 10 will be explained using the flowchart in Figure 5. The purpose here is simply to explain the processing flow. Details of each process have been described above, so explanations will be omitted here as appropriate.

[0056] In S20, the information processing device 10 acquires a predetermined number of material information and information indicating the relationships between the multiple material information. For example, the information processing device 10 can acquire a predetermined number of material information indicating the most recent predetermined number of statements, and information indicating the order in which those statements were made, from the means providing the chat service.

[0057] In S21, the information processing device 10 sets weights for a predetermined number of material information obtained in S20 based on their relationships with each other. For example, the information processing device 10 can increase the weight of material information that indicates a statement that is closer to the present time, i.e., a more recent statement.

[0058] In S22, the information processing device 10 generates first input information by integrating a predetermined number of material information using the weights set in S21. For example, the information processing device 10 vectorizes each of the predetermined number of material information. Then, the information processing device 10 can calculate a weighted sum of the vectors of the multiple material information (first input information) using the weights of each of the multiple material information.

[0059] In S23, the information processing device 10 searches for information related to the first input information using the first model and outputs the search results as the first output information.

[0060] In S24, the information processing device 10 generates first output information and second input information including an instruction statement. For example, the information processing device 10 can generate past conversation history on the chat service, the first output information output in S23, and second input information including a pre-prepared instruction statement.

[0061] In S25, the information processing device 10 inputs the second input information generated in S24 as a prompt to the second model and obtains the response of the instruction statement output from the second model.

[0062] In S26, the information processing device 10 outputs the response to the instruction obtained in S25. For example, the output response is input to the means that provides the chat service. The means that provides the chat service can output the input response as the next computer-side statement in the ongoing conversation on the chat service.

[0063] As described above, the information processing device 10 can utilize generative AI and RAG technologies. Generative AI and RAG are implemented, for example, by a neural network. A neural network includes multiple artificial neurons, each with synapses connecting them. Each synapse has a weight. When such a neural network receives an input, it performs calculations using the weights associated with each synapse and produces an output corresponding to the input. A model representing the connection relationships between neurons and synapses is stored in memory, for example, in software form. Alternatively, the model may be implemented as a dedicated circuit. Similarly, the weights of each synapse are also stored in memory in software form. Alternatively, a circuit representing the weights may be implemented in a dedicated circuit. When constructing generative AI using multiple models, it is not necessarily required that all models be stored in the same memory. There are many different types of models using such neural networks. Generative AI and RAG may be realized by employing and substituting a wide variety of models, such as Transformers, Convolutional Neural Networks (CNNs), and Recurrent Neural Networks (RNNs).

[0064] <Effects and Effects> The information processing device 10 of the second embodiment can achieve the same effects and effects as the information processing device 10 of the first embodiment.

[0065] Furthermore, the information processing device 10 can be used to determine the content of the computer's statements in a chat service where a computer and a human converse. The content of the computer's statements is generated using generative AI such as a large-scale language model, but in this case, the following problems arise.

[0066] In multi-turn conversations, it's necessary to consider not only the most recent statement but also past statements to understand what the person is asking. In the multi-turn conversation shown in Figure 3, the most recent statement is "Please tell me your email address," but this statement alone doesn't tell us whose email address is being asked for. However, by considering the statements made before that, we can understand that the person is asking for "the email address of the contact person who handles inquiries about expense reimbursements for business trips," that is, "the email address of the XX Department XX Group."

[0067] As shown in the comparative example in Figure 6, if the first input information (a query to search for information to include in the prompt) is generated based only on the most recent utterance, there is a risk that the appropriate information cannot be found using that first input information. In this case, the appropriate information cannot be included in the prompt input to the second model. Consequently, the desired answer (email address of the XX department XX group) cannot be obtained from the second model.

[0068] In contrast, as shown in Figure 3, the information processing device 10 generates first input information based on multiple statements, specifically a predetermined number of recent statements. Based on this first input information, it becomes possible to search for appropriate information. In this case, the prompt to be input to the second model can include appropriate information. Then, the desired answer (email address of the XX department XX group) can be obtained from the second model.

[0069] Furthermore, the information processing device 10 can assign weights to a predetermined number of recent statements, taking into account their order of appearance, and generate first input information considering these weights. Specifically, the information processing device 10 can assign greater weights to more recent statements. Generally, statements that are closer to the present time have a stronger relationship with the most recent statements. Therefore, by assigning weights as described above and generating first input information considering these weights, the information processing device 10 can generate first input information that allows for more accurate retrieval of the information necessary to create a response to the most recent statements.

[0070] <<Third Embodiment>> The information processing device 10 of the third embodiment differs from the information processing device 10 of the first and second embodiments in the process of generating the first input information. Specifically, the information processing device 10 of the third embodiment sets weights for multiple statements based on the similarity of the multiple statements, rather than the relationship between the timing of statements of multiple statements, and generates the first input information using these weights. This will be explained in detail below.

[0071] Figure 1 is an example of a functional block diagram of the information processing device 10. As shown in the figure, the information processing device 10 has a generation unit 11, a first processing unit 12, and a second processing unit 13.

[0072] The generation unit 11 generates first input information based on a plurality of material information and the relationships between the plurality of material information, similar to the first and second embodiments.

[0073] The multiple material information, as in the second embodiment, is, for example, past statements in a chat service where a computer and a person converse. In the third embodiment, as in the second embodiment, the number of material information items used to generate the first input information is predetermined. For example, a predetermined number of recent statements in a series of conversations on the chat service, that is, a predetermined number of statements starting from the most recent, are used to generate the first input information.

[0074] The "relationship between multiple material information" indicates the similarity of the multiple statements indicated by the multiple material information. For example, the relationship between multiple material information indicates the similarity between the latest statement and each of the other statements. Such information indicating the relationship between multiple material information can be generated by the generation unit 11. That is, the generation unit 11 can generate information indicating the relationship between multiple material information by calculating the similarity between the latest statement and each of the other statements. The calculation of the similarity between two sentences (texts) is achieved using widely known techniques. For example, each of the two sentences (texts) may be vectorized, and the similarity between the two vectors may be calculated.

[0075] Next, we will explain the process for generating the first input information.

[0076] First, the generation unit 11 sets a weight for each of the multiple material information based on the relationships between them. The generation unit 11 gives the largest weight to the material information that represents the most recent statement. Then, the generation unit 11 gives an even larger weight to the material information that represents statements with a greater similarity to the most recent statement.

[0077] Furthermore, the generation unit 11 may set the weight of the material information indicating statements whose similarity to the most recent statement is below a threshold value as the lower limit. The lower limit is, for example, the minimum value in the range of values ​​that can be set as weights, or a value equivalent to it. A sufficiently small value is set as the lower limit. For example, the lower limit may be "0". The similarity threshold is set in advance.

[0078] There are various specific means for setting such weights. For example, the generation unit 11 may determine the weight of each of the multiple material information using a decision model that determines the weight from the similarity. The decision model may be a calculation formula that calculates the weight from the similarity, or it may consist of a table that shows the relationship between similarity and weight.

[0079] After assigning weights to each of the multiple material information items, the generation unit 11 generates integrated information, which is created by integrating the multiple material information items using these weights, as the first input information. Specifically, as shown in Figure 3, the generation unit 11 vectorizes each of the multiple material information items. Vectorization of text can be achieved using widely known techniques. Then, the generation unit 11 calculates a weighted sum (integrated information) of the vectors of the multiple material information items using the weights of each of the multiple material information items.

[0080] Other configurations of the information processing device 10 can be the same as those in the first and second embodiments.

[0081] The information processing device 10 of the third embodiment can achieve the same effects as the information processing device 10 of the first and second embodiments. Furthermore, the information processing device 10 can set a weight for each statement considering its similarity to the most recent statement, and generate first input information (a query for searching for information to include in the prompt) considering this weight. Specifically, the information processing device 10 can increase the weight of statements that are more similar to the most recent statement. Generally, the closer a statement is to the most recent statement, the stronger its relationship to the most recent statement. Therefore, the information processing device 10 that sets weights as described above and generates first input information considering these weights can generate first input information that can search for the information necessary to create a response to the most recent statement with higher accuracy. As a result, responses that do not consider statements with a low relationship to the most recent statement (second output information) are more likely to be output from the second model.

[0082] Furthermore, the information processing device 10 can set the weight of material information indicating statements whose similarity to the latest statement is below a threshold to a lower limit (a sufficiently small value). With such an information processing device 10, the influence of statements with little relation to the latest statement can be reduced or eliminated, and the first input information can be generated. As a result, it becomes possible to generate first input information that allows for more accurate retrieval of the information necessary to create a response to the latest statement. In addition, responses that do not consider statements with little relation to the latest statement (second output information) are more likely to be output from the second model.

[0083] <<Fourth Embodiment>> The information processing device 10 of the fourth embodiment differs from the information processing device 10 of the first to third embodiments in the process of generating the first input information. Specifically, the information processing device 10 of the fourth embodiment sets weights for multiple statements based on both the relationship between the timing of statements and the similarity of multiple statements, and generates the first input information using these weights. This will be explained in detail below.

[0084] Figure 1 is an example of a functional block diagram of the information processing device 10. As shown in the figure, the information processing device 10 has a generation unit 11, a first processing unit 12, and a second processing unit 13.

[0085] The generation unit 11 generates first input information based on a plurality of material information and the relationships between the plurality of material information, similar to the first to third embodiments.

[0086] The multiple material information, as in the second and third embodiments, is, for example, past statements in a chat service where a computer and a person converse. In the fourth embodiment, as in the second and third embodiments, the number of material information items used to generate the first input information is predetermined. For example, a predetermined number of recent statements in a series of conversations on the chat service, that is, a predetermined number of statements starting from the most recent, are used to generate the first input information.

[0087] "The relationships between multiple pieces of material information" refers to the relationships between the timing of statements indicated by multiple pieces of material information, and the degree of similarity between the statements indicated by multiple pieces of material information. The content of these relationships between multiple pieces of material information, and the means for acquiring (generating) them, are as described in the second and third embodiments.

[0088] Next, we will explain the process for generating the first input information.

[0089] First, the generation unit 11 sets a weight for each of the multiple material information based on their relationships with each other. The generation unit 11 can increase the weight of material information that indicates a statement that is closer to the current time. The generation unit 11 can also increase the weight of material information that indicates a statement that is more similar to the latest statement. The generation unit 11 may set the weight of material information that indicates a statement that is below a threshold in similarity to the latest statement as a lower limit. The lower limit is, for example, the minimum value in the range of values ​​that can be set as a weight, or a value equivalent to it. A sufficiently small value is set as the lower limit. For example, the lower limit may be "0". The similarity threshold is set in advance.

[0090] There are various specific means for setting such weights. For example, the generation unit 11 may determine a first weight for each of the multiple material information based on the relationship between the timing of multiple statements indicated by the multiple material information, using the method described in the second embodiment. Alternatively, the generation unit 11 may determine a second weight for each of the multiple material information based on the similarity of multiple statements indicated by the multiple material information, using the method described in the third embodiment. The first weight and the second weight may then be integrated to determine the weight for each of the multiple material information. Examples of integrating the first and second weights include calculating the average value or the weighted average value, but are not limited to these.

[0091] After assigning weights to each of the multiple material information items, the generation unit 11 generates integrated information, which is created by integrating the multiple material information items using these weights, as the first input information. Specifically, as shown in Figure 3, the generation unit 11 vectorizes each of the multiple material information items. Vectorization of text can be achieved using widely known techniques. Then, the generation unit 11 calculates a weighted sum (integrated information) of the vectors of the multiple material information items using the weights of each of the multiple material information items.

[0092] Other configurations of the information processing device 10 can be the same as those of the first to third embodiments.

[0093] The information processing device 10 of the fourth embodiment can achieve the same effects as the information processing device 10 of the first to third embodiments. Furthermore, the information processing device 10 can set a weight for each statement content, taking into account the order of statements of a predetermined number of recent statements and the similarity to the latest statement content, and can generate first input information (a query for searching for information to be included in the prompt) taking these weights into account. With such an information processing device 10, it becomes possible to generate first input information that can search with higher accuracy for the information necessary to create a response to the latest statement content.

[0094] <<Fifth Embodiment>> The information processing device 10 of the fifth embodiment differs from the information processing device 10 of the first to fourth embodiments in the process of generating the first input information. Specifically, in the information processing device 10 of the fifth embodiment, the number of material information used to generate the first input information is not fixed but changes. This will be explained in detail below.

[0095] Figure 1 is an example of a functional block diagram of the information processing device 10. As shown in the figure, the information processing device 10 has a generation unit 11, a first processing unit 12, and a second processing unit 13.

[0096] The generation unit 11 determines the content of the statement to be used to generate the first output information based on the similarity between the most recent statement in a series of conversations on the chat service and each of the other statements. In other words, the generation unit 11 determines the content of the statement to be used as material information from among multiple statements.

[0097] For example, the generation unit 11 does not include statements whose similarity to the most recent statement is below a threshold, as well as statements made before that, in the material information used to generate the first output information. The generation unit 11 then uses the remaining statements as material information to generate the first output information. All of the remaining statements have a similarity to the most recent statement that is greater than the threshold.

[0098] Statements whose similarity to the most recent statement is below a threshold have a low relationship to the most recent statement. It is considered that the topic has shifted in such statements. Statements made before such a shift in topic have a low relationship to the most recent statement. Therefore, the generation unit 11 does not include statements whose similarity to the most recent statement is below a threshold, as well as statements made before that, in the material information used to generate the first output information.

[0099] The generation unit 11 determines a plurality of material information to be used to generate the first output information, and then generates the first input information based on the plurality of material information and the relationships between them. The generation unit 11 can perform this processing using the method described in the first to fourth embodiments.

[0100] Other configurations of the information processing device 10 can be the same as those of the first to fourth embodiments.

[0101] The information processing device 10 of the fifth embodiment can achieve the same effects as the information processing device 10 of the first to fourth embodiments. Furthermore, the information processing device 10 can determine which statements to use for generating the first output information based on the similarity between the most recent statement in a series of conversations on the chat service and each of the other statements. Specifically, the information processing device 10 does not include statements whose similarity to the most recent statement is below a threshold, and statements made before that, as material information for generating the first output information. The information processing device 10 then uses the other statements as material information for generating the first output information.

[0102] Statements whose similarity to the most recent statement is below a threshold have a low relationship to the most recent statement. It is considered that the topic has shifted in such statements. Statements made before such a shift in topic have a low relationship to the most recent statement. Therefore, the information processing device 10 does not include statements whose similarity to the most recent statement is below a threshold, as well as statements made before that, in the material information used to generate the first output information.

[0103] This information processing device 10 enables the generation of first input information that allows for more accurate retrieval of information necessary to create a response to the latest statement. Furthermore, it makes it easier for the second model to output a response (second output information) that does not consider statements that are less relevant to the latest statement.

[0104] <<Sixth Embodiment>> The information processing device 10 of the sixth embodiment differs from the information processing device 10 of the first to fifth embodiments in the process of generating the first input information. Specifically, the number of material information used to generate the first input information in the information processing device 10 of the sixth embodiment is not fixed but changes. The information processing device 10 of the sixth embodiment uses all past statements as material information to generate the first input information. Therefore, the more statements there are (the more turns there are in the conversation), the more material information used to generate the first input information increases. This will be explained in detail below.

[0105] Figure 1 is an example of a functional block diagram of the information processing device 10. As shown in the figure, the information processing device 10 has a generation unit 11, a first processing unit 12, and a second processing unit 13.

[0106] The generation unit 11 acquires all the content of statements in a series of conversations on the chat service as raw material information. Then, the generation unit 11 assigns weights to all the raw material information.

[0107] The generation unit 11 can increase the weight of material information that indicates a statement that is closer to the current time. Furthermore, the generation unit 11 can increase the weight of material information that indicates a statement with a greater similarity to the most recent statement.

[0108] Furthermore, the generation unit 11 may set the weight of material information indicating statements whose similarity to the most recent statement is below a threshold as the lower limit. Also, the generation unit 11 may set the weight of material information indicating statements whose similarity to the most recent statement is below a threshold, and statements made earlier than that, as the lower limit. Also, the generation unit 11 may set the weight of material information indicating statements made at a timing more than a threshold away from the timing of the most recent statement as the lower limit.

[0109] The lower limit is, for example, the minimum value in the range of values ​​that can be set as weights, or a value equivalent to it. A sufficiently small value is set as the lower limit. For example, the lower limit may be "0". The similarity threshold and the utterance timing threshold are set in advance. The utterance timing threshold may also be indicated by, for example, the number of turns. In this case, the generation unit 11 can set the weight of the material information indicating utterances made at utterance timings that are more than a predetermined number of turns (threshold) away from the utterance timing of the most recent utterance as the lower limit.

[0110] There are various specific means for setting such weights. For example, the generation unit 11 can set the weights in accordance with the methods described in the second to fifth embodiments.

[0111] After assigning weights to each of the multiple material information items, the generation unit 11 generates integrated information, which is created by integrating the multiple material information items using these weights, as the first input information. Specifically, as shown in Figure 3, the generation unit 11 vectorizes each of the multiple material information items. Vectorization of text can be achieved using widely known techniques. Then, the generation unit 11 calculates a weighted sum (integrated information) of the vectors of the multiple material information items using the weights of each of the multiple material information items.

[0112] Other configurations of the information processing device 10 can be the same as those of the first to fifth embodiments.

[0113] The information processing device 10 of the sixth embodiment can achieve the same effects as the information processing device 10 of the first to fifth embodiments. Furthermore, the information processing device 10 can generate first input information (a query for searching for information to be included in the prompt) using all the content of statements in a series of conversations on the chat service. With this information processing device 10, it becomes possible to search for the information necessary to create a response to the latest statement by considering all past statements without omission. As a result, a response that considers all past statements (second output information) is more likely to be output from the second model.

[0114] However, in a series of conversations on a chat service, the topic may change midway through. The content of statements made before the topic change may not be very relevant to the latest statement and may not need to be considered when creating a response to the latest statement. Nevertheless, if all statements are used in the same way to generate the first input information, the accuracy of the search results and the output from the second model will be poor. Therefore, the information processing device 10 can suppress this problem by appropriately setting weights for multiple statements. Specifically, the information processing device 10 can set the weight of material information indicating statements that have little relation to the latest statement to a lower limit. This reduces or eliminates the influence of statements that have little relation to the latest statement. As a result, the above problem that may arise from using all statements can be suppressed.

[0115] <<Modifications>> Here, modifications applicable to the first to sixth embodiments are described. These modifications can also achieve the same effects and advantages as the first to sixth embodiments.

[0116] The generation unit 11 does not need to determine the weights of the multiple material information. The generation unit 11 may also generate the first input information by integrating the multiple material information without weighting. For example, the generation unit 11 can vectorize each of the multiple material information and generate their sum as the first input information. Even if the multiple material information is not weighted and is treated equally to generate the first input information (a query for searching for information to include in the prompt), the information to include in the prompt can be detected with high accuracy by utilizing the multiple material information.

[0117] Although this disclosure has been described above with reference to embodiments, this disclosure is not limited to the embodiments described above. Various modifications to the structure and details of this disclosure are possible, which can be understood by those skilled in the art within the scope of this disclosure. Furthermore, each embodiment can be combined with other embodiments as appropriate.

[0118] Furthermore, the flowcharts used in the above description show multiple steps (processes) in sequence. However, the execution order of the steps performed in each embodiment is not limited to the order in which they are described. In each embodiment, the order of the illustrated steps can be changed to the extent that it does not impede the content.

[0119] Some or all of the above embodiments may also be described as follows, but are not limited to the following: 1. A program that causes a computer to function as: a generation means for generating first input information based on a plurality of material information and the relationships between the plurality of material information; a first processing means for receiving the input of the first input information and executing a first process using a first model that outputs first output information; and a second processing means for receiving the input of second input information including the first output information and executing a second process using a second model that outputs second output information. 2. The program according to 1, wherein the plurality of material information represents a plurality of statements in a conversation, and the relationships between the plurality of material information represent at least one of the relationship between the timing of the statements of the plurality of statements and the similarity of the plurality of statements. 3. The program according to 2, wherein the generation means sets a weight for each of the plurality of material information based on the relationships between the plurality of material information, and generates integrated information by integrating the plurality of material information using the weights as the first input information. 4. 1. The generation means is a program according to 3 that, in the process of setting the weights, increases the weight of the material information that indicates a statement that is closer to the current time. 5. The generation means is a program according to 3 or 4 that, in the process of setting the weights, increases the weight of the material information that indicates a statement that has a greater similarity to the latest statement. 6. The generation means is a program according to any one of 3 to 5 that, in the process of setting the weights, sets the weight of the material information that indicates a statement that was made at a timing that is more than a threshold away from the timing of the latest statement to a lower limit. 7. The generation means is a program according to any one of 3 to 6 that, in the process of setting the weights, sets the weight of the material information that indicates a statement that has a similarity to the latest statement that is below a threshold to a lower limit.8. A program according to any one of 1 to 7, wherein the first model, upon receiving the input of the first input information, searches for information related to the first input information and outputs the search results as the first output information, and the second model, upon receiving the input of the first output information and the second input information including an instruction, outputs the answer to the instruction as the second output information. 9. An information processing device having: generation means for generating first input information based on a plurality of material information and the relationships between the plurality of material information; first processing means for executing a first process using a first model that receives the input of the first input information and outputs the first output information; and second processing means for executing a second process using a second model that receives the input of second input information including the first output information and outputs the second output information. 10. An information processing method comprising: one or more computers generating first input information based on multiple material information and the relationships between the multiple material information; executing a first process using a first model that receives the input of the first input information and outputs first output information; and executing a second process using a second model that receives the input of second input information including the first output information and outputs second output information.

[0120] Some or all of the appendices 2 through 8, which are dependent on the program described in appendice 1 above, may also be dependent on the information processing device in appendice 9 and the information processing method in appendice 10 in the same dependent relationship as between appendice 1 and appendices 2 through 8. Furthermore, without departing from the embodiments described above, some or all of the configurations described as appendices can be realized in various hardware, software, various recording means for recording software, or systems.

[0121] This application claims priority based on Japanese Patent Application No. 2024-173092, filed on 2 October 2024, and incorporates all of its disclosures herein.

[0122] 10 Information processing device 11 Generation unit 12 First processing unit 13 Second processing unit 1A Processor 2A Memory 3A Input / Output I / F 4A Peripheral circuit 5A Bus

Claims

1. A recording medium that stores a program causing a computer to function as: a generation means for generating first input information based on a plurality of material information and the relationships between the plurality of material information; a first processing means for receiving the input of the first input information and executing a first process using a first model that outputs first output information; and a second processing means for receiving the input of second input information including the first output information and executing a second process using a second model that outputs second output information.

2. The recording medium according to claim 1, wherein the plurality of material information indicates the content of multiple statements made in a conversation, and the relationship between the plurality of material information indicates at least one of the relationship between the timing of the multiple statements and the degree of similarity between the multiple statements.

3. The recording medium according to claim 2, wherein the generation means sets a weight for each of the plurality of material information based on the relationships between the plurality of material information, and generates integrated information by integrating the plurality of material information using the weights as the first input information.

4. The recording medium according to claim 3, wherein the generation means, in the process of setting the weights, increases the weight of the material information that indicates a statement that is closer to the present time in terms of the timing of the statement.

5. The recording medium according to claim 3, wherein the generation means, in the process of setting the weights, increases the weight of the material information that indicates a statement with a greater similarity to the most recent statement.

6. The recording medium according to claim 3, wherein the generation means, in the process of setting the weight, sets the weight of the material information indicating a statement made at a statement timing that is more than a threshold away from the statement timing of the most recent statement to a lower limit value.

7. The recording medium according to claim 3, wherein the generation means, in the process of setting the weight, sets the weight of the material information indicating a statement whose similarity to the latest statement is below a threshold value to a lower limit value.

8. The recording medium according to any one of claims 1 to 7, wherein the first model, upon receiving the input of the first input information, searches for information related to the first input information and outputs the search results as the first output information, and the second model, upon receiving the input of the first output information and the second input information including an instruction, outputs the response to the instruction as the second output information.

9. An information processing apparatus comprising: generation means for generating first input information based on multiple material information and the relationships between the multiple material information; first processing means for receiving the input of the first input information and executing a first process using a first model that outputs first output information; and second processing means for receiving the input of second input information including the first output information and executing a second process using a second model that outputs second output information.

10. The information processing apparatus according to claim 9, wherein the plurality of material information items represent the content of multiple statements made in a conversation, and the relationship between the plurality of material information items represents at least one of the relationship between the timing of the multiple statements and the degree of similarity between the multiple statements.

11. The information processing apparatus according to claim 10, wherein the generation means sets a weight for each of the plurality of material information based on the relationships between the plurality of material information, and generates integrated information by integrating the plurality of material information using the weights as the first input information.

12. The information processing apparatus according to claim 11, wherein the generation means, in the process of setting the weights, increases the weight of the material information that indicates a statement that is closer to the present time in terms of the timing of the statement.

13. The information processing apparatus according to claim 11, wherein the generation means, in the process of setting the weights, increases the weight of the material information that indicates a statement with a greater similarity to the most recent statement.

14. The information processing apparatus according to claim 11, wherein the generation means, in the process of setting the weight, sets the weight of the material information indicating a statement made at a statement timing that is more than a threshold away from the statement timing of the most recent statement to a lower limit.

15. An information processing method in which one or more computers generate first input information based on a plurality of material information and the relationships between the plurality of material information; execute a first process using a first model that receives the input of the first input information and outputs first output information; and execute a second process using a second model that receives the input of second input information including the first output information and outputs second output information.

16. The information processing method according to claim 15, wherein the plurality of material information indicates the content of multiple statements made in a conversation, and the relationship between the plurality of material information indicates at least one of the relationship between the timing of the multiple statements and the degree of similarity between the multiple statements.

17. The information processing method according to claim 16, wherein one or more computers set weights for each of the plurality of material information based on the relationships between them, and generate integrated information by integrating the plurality of material information using the weights as the first input information.

18. The information processing method according to claim 17, wherein one or more computers, in the process of setting the weights, increase the weight of the material information that indicates a statement that is closer to the present time in terms of the timing of the statement.

19. The information processing method according to claim 17, wherein one or more computers, in the process of setting the weights, increase the weight of the material information that indicates a statement of speech that has a greater similarity to the most recent statement of speech.

20. The information processing method according to claim 17, wherein, in the process of setting the weights, one or more computers set the weight of the material information indicating a statement made at a timing that is more than a threshold away from the timing of the most recent statement to a lower limit.

Images