Information processing device, generation control method, and generation control program

Abstract

This information processing device is provided with: a query decomposition unit that decomposes a query entered by a user to generate a plurality of decomposed queries; an allocation unit that allocates each of the plurality of decomposed queries to one of a plurality of trained models that have been trained through machine learning to be able to generate answers to queries; and a generation control unit that transmits a decomposed query to each trained model according to the allocation determined by the allocation unit so that the trained model generates an answer. This information processing device also makes it possible to optimize answers to queries.

Description

Information processing device, generation control method, and generation control program

[0001] The present disclosure relates to an information processing device, a generation control method, and a generation control program.

[0002] In recent years, with the development of large language models (LLMs), question-answering systems using natural language processing technology have become widespread. LLMs acquire a wide range of knowledge by learning large amounts of text data, and can generate appropriate answers to questions entered in natural language. For example, Patent Document 1 listed below discloses an information providing device that generates answers to questions about labor management using a language model.

[0003] Japanese Patent No. 7353695

[0004] In technologies for generating answers to queries input by users using trained models such as language models, including the information providing device described in Patent Literature 1, there is room for further improvement in the performance of generating answers. An exemplary object of the present disclosure is to provide a technology that enables improvement in the performance of generating answers using trained models.

[0005] An information processing device according to an exemplary aspect of the present disclosure includes a query decomposition unit that decomposes a query input by a user to generate a plurality of decomposed queries, an allocation unit that allocates each of the plurality of decomposed queries to one of a plurality of trained models that have been machine-learned to generate an answer to the query, and a generation control unit that sends the decomposed query to each trained model in accordance with the allocation determined by the allocation unit, causing the trained model to generate an answer.

[0006] A generation control method according to an exemplary aspect of the present disclosure includes at least one processor executing a query decomposition process in which a query input by a user is decomposed to generate a plurality of decomposed queries; an allocation process in which each of the plurality of decomposed queries is assigned to one of a plurality of trained models that have been machine-learned to generate an answer to the query; and a generation control process in which the decomposed query is sent to each trained model in accordance with the allocation determined by the allocation process, causing the trained model to generate an answer.

[0007] A generation control program according to an exemplary aspect of the present disclosure causes a computer to function as a query decomposition unit that decomposes a query input by a user to generate a plurality of decomposed queries, an allocation unit that allocates each of the plurality of decomposed queries to one of a plurality of trained models that have been machine-learned to generate an answer to the query, and a generation control unit that sends the decomposed query to each trained model in accordance with the allocation determined by the allocation unit, causing the trained model to generate an answer.

[0008] According to one exemplary aspect of the present disclosure, an exemplary effect is achieved in that a technology can be provided that enables improvement in the performance of generating answers using a trained model.

[0009] 1 is a block diagram showing a configuration of an information processing device according to the present disclosure; FIG. 2 is a flow diagram showing a flow of a generation control method according to the present disclosure; FIG. 3 is a diagram showing an example configuration of a response system according to the present disclosure; FIG. 4 is a block diagram showing a configuration of an information processing device according to the present disclosure; FIG. 5 is a flow diagram showing a flow of processing executed by the information processing device shown in FIG. 4; FIG. 6 is a block diagram showing a configuration of a computer that functions as an information processing device according to the present disclosure.

[0010] The following are examples of embodiments of the present invention. However, the present invention is not limited to the exemplary embodiments shown below, and various modifications are possible within the scope of the claims. For example, embodiments obtained by appropriately combining the technologies (part or all of the products or methods) employed in the exemplary embodiments shown below may also be included in the scope of the present invention. Furthermore, embodiments obtained by appropriately omitting some of the technologies employed in the exemplary embodiments shown below may also be included in the scope of the present invention. Furthermore, the effects mentioned in the exemplary embodiments shown below are examples of effects expected in the exemplary embodiments, and do not define the scope of the present invention. In other words, embodiments that do not exhibit the effects mentioned in the exemplary embodiments shown below may also be included in the scope of the present invention.

[0011] [First Exemplary Embodiment] A first exemplary embodiment, which is an example of an embodiment of the present invention, will be described. Note that the scope of application of each technique employed in this exemplary embodiment is not limited to this exemplary embodiment. In other words, each technique employed in this exemplary embodiment can also be employed in other exemplary embodiments included in the present disclosure, as long as no particular technical obstacles arise.

[0012] (Points of the Invention) The received job (i.e., the input query) is decomposed (i.e., divided into multiple parts) and assigned (i.e., each part of the query is assigned to a different LLM).

[0013] (Configuration of information processing device) The information processing device according to this exemplary embodiment includes a query decomposition means that decomposes an input query to generate a plurality of decomposed queries, an allocation means that determines an allocation of each of the plurality of decomposed queries to one of a plurality of language models, and a generation control means that causes each of the plurality of language models to generate an answer to the plurality of decomposed queries in accordance with the allocation.

[0014] According to the above configuration, it is possible to generate an answer to a query using multiple language models, which makes it possible to generate highly accurate answers that take advantage of the characteristics of each language model, and to generate answers to complex queries in a short time.

[0015] The query decomposition means decomposes a question using, for example, a question decomposition model that has been machine-learned to decompose a question into multiple parts. The query decomposition means may record dependencies between decomposed queries (e.g., an answer to one decomposed query becomes part of another decomposed query). The recorded dependencies are used when aggregating answers to each decomposed query.

[0016] The allocation means, for example, identifies in advance the field to which each language model is suited from the learning data of the language model. Then, the allocation means analyzes each decomposed query to identify its field and allocates the query to a language model suited to the identified field. For example, a decomposed query related to the medical field is allocated to a language model suited to the medical field.

[0017] Furthermore, the allocation means allocates, for example, to a language model with an answer accuracy / answer speed according to the decomposed query, as in the above example. For example, a language model with a high answer accuracy is allocated to a decomposed query with a high degree of importance. Note that user attributes (age, interests, occupation, etc.), user feedback on past answers, the cost of using each language model, etc. may also be taken into consideration. Furthermore, a reliability level may be assigned to each language model, and the adoption order may be determined according to the reliability level, or a language model with a low reliability may be replaced.

[0018] The generation control means transmits a decomposition query to each language model (a server having that language model) in accordance with the allocation, and obtains a response.

[0019] (Flowchart) 1. Obtain an input query. 2. Decompose the input query to generate multiple decomposed queries. 3. Assign each decomposed query to a language model. 4. Have the language model generate an answer to the decomposed query.

[0020] Second Exemplary Embodiment A second exemplary embodiment, which is an example of an embodiment of the present invention, will be described in detail with reference to the drawings. This exemplary embodiment is a basic form for the exemplary embodiments described below. The scope of application of each technique employed in this exemplary embodiment is not limited to this exemplary embodiment. That is, each technique employed in this exemplary embodiment can also be employed in other exemplary embodiments included in the present disclosure, to the extent that no particular technical obstacles arise. Furthermore, each technique shown in the drawings referenced to explain this exemplary embodiment can also be employed in other exemplary embodiments included in the present disclosure, to the extent that no particular technical obstacles arise. These matters also apply to the third exemplary embodiment.

[0021] (Configuration of information processing device 1) The configuration of the information processing device 1 according to this exemplary embodiment will be described with reference to Fig. 1. Fig. 1 is a block diagram showing the configuration of the information processing device 1. As shown in Fig. 1, the information processing device 1 includes a query decomposition unit 101, an allocation unit 102, and a generation control unit 103.

[0022] The query decomposition unit 101 decomposes a query input by a user to generate a plurality of decomposed queries. Here, a "query" refers to a query to the information processing device 1. The content of the query is not particularly limited, and may be, for example, a question such as "What's the weather like in Tokyo tomorrow?" or a command such as "Please create a summary of the input text." Therefore, the "query" in the following embodiments may be interpreted as an inquiry, question, command, request, or the like. The query may be written in, for example, a natural language or an artificial language such as a programming language.

[0023] The query may be input as text data or as voice data. In the latter case, the input voice data can be processed in the same way as the former by converting it into text data through voice recognition processing.

[0024] The allocation unit 102 allocates each of the multiple decomposed queries generated by the query decomposition unit 101 to one of multiple trained models that have been machine-learned to generate an answer to the query. When a natural language query is used, a language model such as an LLM may be used as the trained model. For example, the language model may be a Generative Pre-Trained Transformer (GPT) that predicts a character string that is likely to follow an input character string and outputs a sentence containing the input character string. Other language models that may be used include a Text-to-Text Transfer Transformer (T5), a Bidirectional Encoder Representations from Transformers (BERT), a Robustly optimized BERT approach (RoBERTa), and an Efficiently Learning an Encoder that Classifies Token Replacements Accurately (ELECTRA).

[0025] The time required to complete answer generation can be expected to be shortened by having multiple trained models generate answers in parallel according to the allocation by the allocation unit 102. Furthermore, by having multiple trained models share the task of answer generation, the load of answer generation on each trained model can be reduced.

[0026] The answer generated by the trained model may be text data, or may be data in other formats such as image data, audio data, or numerical values. It is preferable to apply the multiple trained models with different characteristics, such as different training data used for training or different formats of output data. This makes it possible to generate highly accurate answers that take advantage of the characteristics of each trained model. It is also possible to apply the multiple trained models with common characteristics. In this case, too, it is possible to expect the effects of shortening the time required to complete answer generation and distributing the load.

[0027] The generation control unit 103 transmits the decomposition query to each trained model in accordance with the allocation determined by the allocation unit 102, causing an answer to be generated. Some or all of the plurality of trained models may be provided in the information processing device 1, or all of the plurality of trained models may be provided in a device external to the information processing device 1. When using a trained model provided in an external device, the generation control unit 103 transmits the decomposition query to the device to cause an answer to be generated.

[0028] As described above, the information processing device 1 according to this exemplary embodiment is configured to include a query decomposition unit 101 that decomposes a query input by a user to generate a plurality of decomposed queries, an allocation unit 102 that allocates each of the plurality of decomposed queries to one of a plurality of trained models that have been machine-learned to generate an answer to the query, and a generation control unit 103 that sends the decomposed queries to each trained model in accordance with the allocation determined by the allocation unit 102, causing the trained model to generate an answer.

[0029] According to the above configuration, it is possible to generate an answer to a query using multiple trained models. This makes it possible, for example, to generate highly accurate answers that take advantage of the characteristics of each trained model, or to generate answers to complex queries in a short time. In other words, according to the above configuration, it is possible to improve the performance of generating answers using trained models. Note that improved answer generation performance may include not only improved accuracy in terms of answer content, but also improved answer generation speed (which can also be rephrased as a reduction in the time until an answer is presented), reduced load when generating an answer, and the like.

[0030] (Generation Control Program) The functions of the information processing device 1 described above can also be realized by a program. The generation control program according to this exemplary embodiment causes a computer to function as a query decomposition unit that decomposes a query input by a user to generate multiple decomposed queries, an allocation unit that allocates each of the multiple decomposed queries to one of multiple trained models that have been machine-learned to generate answers to the query, and a generation control unit that sends the decomposed queries to each trained model in accordance with the allocation determined by the allocation unit, causing the trained models to generate answers. This generation control program provides the effect of enabling improved performance in generating answers using trained models.

[0031] (Flow of the Generation Control Method) The flow of the generation control method according to this exemplary embodiment will be described with reference to Fig. 2. Fig. 2 is a flow diagram showing the flow of the generation control method. Note that the execution entity of each step in this generation control method may be a processor provided in the information processing device 1, or a processor provided in another device, or the execution entity of each step may be a processor provided in each different device.

[0032] In S1, at least one processor accepts a query input by a user. The query may be input by any method. For example, the query may be input via an input device such as a keyboard or a mouse, or via a voice input device such as a microphone.

[0033] In S2 (query decomposition process), at least one processor decomposes the query input in S1 to generate a plurality of decomposed queries.

[0034] In S3 (allocation process), at least one processor allocates each of the multiple decomposed queries generated in S2 to one of multiple trained models that have been machine-learned to generate an answer to the query.

[0035] In S4 (generation control process), at least one processor sends the decomposition query generated in S2 to each trained model according to the allocation determined by the process in S3, causing the model to generate an answer.

[0036] As described above, in the generation control method according to this exemplary embodiment, at least one processor executes a query decomposition process that decomposes a query input by a user to generate multiple decomposed queries, an allocation process that allocates each of the multiple decomposed queries to one of multiple trained models that have been machine-learned to generate answers to the query, and a generation control process that sends the decomposed queries to each trained model in accordance with the allocation determined by the allocation process, causing the trained models to generate answers. This generation control method has the effect of enabling improved performance in generating answers using trained models.

[0037] [Third Exemplary Embodiment] (Configuration of Response System 5A) A response system 5A according to this exemplary embodiment will be described with reference to Fig. 3. Fig. 3 is a diagram showing an example configuration of the response system 5A. As shown in the figure, the response system 5A includes an information processing device 1A, a generative model 21A, a generative model 22A, and a terminal device 3A. These components included in the response system 5A are connected to each other so as to be able to communicate with each other via a network.

[0038] The answer system 5A is a system that has the function of accepting query input from a user via the terminal device 3A, causing the generative models 21A and 22A to generate an answer to the query, and presenting the generated answer to the user.

[0039] The information processing device 1A is a device that has a function of causing the generative models 21A and 22A to generate an answer to a query input by a user. For example, the information processing device 1A may be a platform server provided on a cloud. As will be described in detail later, the information processing device 1A decomposes a query input by a user to generate multiple decomposed queries and allocates each of the generated multiple decomposed queries to one of the generative models 21A and 22A. The information processing device 1A then transmits the decomposed queries to the generative models 21A and 22A according to the determined allocation, causes them to generate answers, and integrates and presents the generated answers to the user. This makes it possible to improve answer generation performance compared to using only one of the generative models 21A and 22A.

[0040] The generative models 21A and 22A are trained models that have been machine-learned to generate answers to queries. For example, when accepting input of a query written in natural language, general-purpose language models that have been machine-learned to learn the arrangement of components (such as words) of sentences written in natural language or the arrangement of sentences in a text may be used as the generative models 21A and 22A. Note that the trained models used by the information processing device 1A are not limited to generative models; the information processing device 1A can also use inference models such as prediction models and classification models. It is also possible to use a model that combines a language model and an inference model to generate an answer in natural language based on the inference results of the inference model. Therefore, in the following description, the term "generative model" can be replaced with any "trained model."

[0041] It is preferable that the generative model 21A and the generative model 22A have different characteristics. For example, the generative models 21A and 22A may have different numbers of parameters to learn. Generally, the greater the number of parameters to learn, the higher the accuracy of the answer, but the longer it takes to generate the answer. In the answering system 5A, answers to decomposition queries for which priority should be placed on answer speed can be generated by a generative model with a smaller number of parameters to learn, and answers to decomposition queries for which priority should be placed on answer accuracy can be generated by a generative model with a larger number of parameters to learn. This makes it possible to achieve both answer speed and answer accuracy.

[0042] Furthermore, for example, by using different training data for at least one of machine learning and fine-tuning, it is possible to generate generative models 21A and 22A with different characteristics. For example, a general-purpose language model may be used as is as the generative model 21A, and a general-purpose language model may be fine-tuned as the generative model 22A using combinations of questions and their answers in a specific technical field as training data. In this case, answers to decomposed queries including general questions may be generated by the generative model 21A, and answers to decomposed queries including questions related to a specific technical field may be generated by the generative model 22A.

[0043] The terminal device 3A is a device that serves as an interface with the user in the answer system 5A. Specifically, the terminal device 3A has a function of accepting a query input by the user and presenting the user with an answer to the query (generated under the control of the information processing device 1A). While Fig. 3 shows an example in which the terminal device 3A is a smartphone, a stationary personal computer or the like can also be used as the terminal device 3A.

[0044] When a general-purpose computer such as a smartphone or a personal computer is used as the terminal device 3A, a predetermined client application may be installed on the terminal device 3A to enable the use of the service provided by the response system 5A. In this case, when using the service provided by the response system 5A, the user simply operates the terminal device 3A to launch the client application. When starting to use the service, the user may be prompted to enter pre-registered authentication information to log in to the service.

[0045] In the example of Fig. 3, a query such as "Create a training report for this month and a menu for next month" is input to the terminal device 3A. In this manner, the reply system 5A can also be used for healthcare purposes. The query input to the terminal device 3A is transmitted to the information processing device 1A. Note that the query is preferably transmitted in an encrypted state using a secure communication protocol such as HTTPS (HyperText Transfer Protocol Secure).

[0046] The information processing device 1A decomposes the received query to generate a plurality of decomposed queries. For example, the information processing device 1A decomposes the input query into two queries: "Create a training report for this month" and "Create a training menu for next month."

[0047] Next, the information processing device 1A assigns each query generated by the decomposition, i.e., the decomposed queries, to the generative model 21A or 22A. Here, for example, the generative model 21A is provided on a server that stores various data indicating the user's training results and is capable of generating answers using that data, while the generative model 22A is a model trained to be able to create a training menu. In this case, the information processing device 1A assigns the decomposed query "Create a report on this month's training" to the generative model 21A, and assigns the decomposed query "Create a training menu for next month" to the generative model 22A.

[0048] Next, the information processing device 1A transmits each decomposition query to the generative model 21A or 22A. The transmission of the decomposition query is performed, for example, via an API (Application Programming Interface). Then, the answer to the decomposition query generated by the generative model 21A or 22A is also transmitted to the information processing device 1A, for example, via the API. The information processing device 1A integrates the answers generated by the generative models 21A and 22A, formats them as necessary, and transmits them to the terminal device 3A. It is preferable that this answer be transmitted in an encrypted state, similar to the query.

[0049] The terminal device 3A that receives the answer decodes and outputs the received answer as necessary. In the example of Fig. 3, a "This Month's Report" and a "Recommended Menu for Next Month" are displayed on the terminal device 3A as answers to the above query. In this way, the answering system 5A can present answers that include a "This Month's Report" with appropriate content using various data indicating the user's training results and a "Recommended Menu for Next Month" with appropriate content based on the user's prior learning results.

[0050] As described above, a user can use an advanced question-answering service that uses multiple generative models, namely, generative models 21A and 22A, simply by inputting a query in natural language into the user's own terminal device 3A. This service provides efficient answers by switching between generative models 21A and 22A transparently to the user.

[0051] The query may be input as text or as voice. The answer to the query may be presented as text as in the example of FIG. 3 or as voice. The answer system 5A may include three or more generative models. The generative models included in the answer system 5A may be stored inside the information processing device 1A or may be stored in another device. For example, the terminal device 3A may include the generative models.

[0052] Here, conventional answering systems that answer queries using a single model may not be able to provide sufficient answers to complex questions, such as questions that span multiple fields or questions that require advanced specialized knowledge. Furthermore, even if the content of the answer is appropriate, it may take a long time to generate it. In this regard, answering system 5A is capable of providing appropriate and prompt answers even to complex questions.

[0053] (Configuration of Information Processing Device 1A) The configuration of the information processing device 1A according to this exemplary embodiment will be described with reference to FIG. 4. FIG. 4 is a block diagram showing the configuration of the information processing device 1A. Note that the information processing device 1A may be a device whose main function is to control the generation of answers to queries, or may be a general-purpose device that also has other functions. Furthermore, the information processing device 1A may be a stationary device as shown in FIG. 3, or may be a portable device. For example, it is also possible to provide the functions of the information processing device 1A to a terminal device 3A shown in FIG. 3.

[0054] As shown in FIG. 4 , the information processing device 1A includes a control unit 10A that controls the various components of the information processing device 1A and a storage unit 11A that stores various data used by the information processing device 1A. The information processing device 1A also includes a communication unit 12A that enables the information processing device 1A to communicate with other devices, an input unit 13A that accepts input to the information processing device 1A, and an output unit 14A that enables the information processing device 1A to output data. The control unit 10A includes a query decomposition unit 101A, an allocation unit 102A, a generation control unit 103A, a presentation control unit 104A, a feature information generation unit 105A, a compatibility evaluation unit 106A, a load information acquisition unit 107A, and an update unit 108A. Details of the feature information generation unit 105A and the update unit 108A will be described later.

[0055] The query decomposing unit 101A, like the query decomposing unit 101 described in the second exemplary embodiment, decomposes a query input by a user to generate multiple decomposed queries. The method for generating the decomposed queries is not particularly limited. For example, the query decomposing unit 101A may generate the decomposed queries using a question decomposition model that has been machine-learned to decompose a question into multiple queries. A known model such as DecompRC may also be applied as the question decomposition model. Furthermore, for example, if the query input by the user is in text format, the query decomposing unit 101A may decompose the query at the positions of periods and commas. Furthermore, for example, the query decomposing unit 101A may perform morphological analysis on the query input by the user and decompose the query based on the results of the morphological analysis.

[0056] Here, an example will be described in which the query decomposition unit 101A decomposes queries with dependencies using a model such as DecompRC. When the query input by the user is "Which companies in industries where sales are declining this year are expected to recover next year?", the query decomposition unit 101A generates, for example, the following three queries Q1, Q2, and Q3. Q1: In which industry are sales declining this year? Q2: Which companies in [Answer A1] are expected to recover next year? Q3: What is the specific company name in [Answer A2]? In this case, query Q2 is in the format where answer A1 of query Q1 is substituted for "[Answer A1]," making it a subquery that depends on query Q1. Similarly, query Q3 is in the format where answer A2 of query Q2 is substituted for "[Answer A2]," making it a subquery that depends on query Q2.

[0057] In this way, the query decomposition unit 101A decomposes a query input by a user into multiple subqueries, and the generation control unit 103A acquires appropriate related documents and related tables for each subquery, thereby enabling efficient derivation of a final answer. In particular, by decomposing a query in consideration of the dependency relationships between subqueries, answers can be obtained in stages, making it possible to generate appropriate answers even for complex queries.

[0058] Furthermore, for example, the query decomposing unit 101A may decompose a query input by a user into a language model. In this case, the query decomposing unit 101A may generate a command statement that includes the query input by the user and commands the query to be divided into multiple sentences according to its content, and input the generated command statement to the language model. This causes the decomposed query to be output from the language model. Note that if at least one of the generation models 21A and 22A is a language model, the query decomposing unit 101A may cause the generation model 21A or 22A, which is a language model, to generate the decomposed query.

[0059] The allocating unit 102A allocates each of the multiple decomposed queries to one of multiple generative models that have been machine-learned so as to be able to generate an answer to the query, similar to the allocating unit 102 described in exemplary embodiment 2. For example, as described based on FIG. 3 , the allocating unit 102A may allocate the decomposed queries to the generative models 21A and 22A.

[0060] Similar to the generation control unit 103 described in exemplary embodiment 3, the generation control unit 103A sends a decomposition query to multiple generative models in accordance with the allocation determined by the allocation unit 102A, causing them to generate answers.

[0061] As described above, the information processing device 1A includes the query decomposition unit 101A that decomposes a query input by a user to generate multiple decomposed queries, the allocation unit 102A that allocates each of the multiple decomposed queries to one of the generative models 21A and 22A, and the generation control unit 103A that transmits the decomposed queries to the generative models 21A and 22A in accordance with the allocation determined by the allocation unit 102A, thereby causing the generative models 21A and 22A to generate answers. Thus, the information processing device 1A has the effect of being able to improve answer generation performance, similar to the information processing device 1 described in exemplary embodiment 2.

[0062] The presentation control unit 104A integrates answers generated by multiple generative models and presents the integrated answers to the user. In addition to the effects of the information processing device 1, the information processing device 1A equipped with the presentation control unit 104A has the effect of being able to present the generated multiple answers in a form that is easy for the user to understand. Details of the method for integrating answers will be described later.

[0063] The compatibility evaluation unit 106A evaluates the compatibility of each of the multiple decomposed queries with the multiple generative models. The allocation unit 102A then allocates the decomposed queries based on the compatibility evaluation results by the compatibility evaluation unit 106A. In addition to the effects of the information processing device 1, the information processing device 1A equipped with the compatibility evaluation unit 106A has the effect of making it possible to perform appropriate allocation in consideration of the compatibility of each combination of decomposed queries and generative models. The compatibility evaluation method will be described in detail later.

[0064] The load information acquisition unit 107A acquires, for each of a plurality of generative models, load information indicating the load status in answer generation using the generative model. The allocation unit 102A then allocates decomposition queries using the load information acquired by the load information acquisition unit 107A. Thus, in addition to the effects of the information processing device 1, the information processing device 1A has the effect of being able to perform allocation taking into account the load status in answer generation using each generative model. For example, the allocation unit 102A may assign a decomposition query to a generative model with a smaller load indicated in the load information, which is expected to result in a faster answer.

[0065] The load information acquisition unit 107A may generate load information or may acquire load information generated by another device via the communication unit 12A or the input unit 13A. It is preferable that the load information be updated in real time. The load information may be discrete information, such as high or low load, or may indicate the magnitude of the load using a continuous value (e.g., a numerical value between 0 and 1). The load in answer generation using a generative model can also be expressed as the load on the device that causes the generative model to generate an answer (the device equipped with the generative model). Any method, including known methods, can be used to calculate the load. For example, the load information may include the number of queries that have already been input to the target generative model but for which no answer has been generated, or the predicted time from when the generative model is instructed to generate an answer until answer generation is completed.

[0066] (Generating an Answer Taking Dependencies into Account) When a query input by a user is decomposed to generate multiple decomposed queries, a dependency relationship may arise between the multiple decomposed queries. For example, suppose the query "Please explain the sales strategy for product X and the benefits of adopting that sales strategy" is decomposed into decomposed query A, "Please explain the sales strategy for product X," and decomposed query B, "Please explain the benefits of adopting the sales strategy." In this case, to generate an appropriate answer to decomposed query B, it is necessary to refer to the answer to decomposed query A. In other words, it can be said that decomposed query B depends on decomposed query A.

[0067] In such a case, the generation control unit 103A generates an answer to the decomposed query A, and then generates an answer to the decomposed query B by referring to that answer. This makes it possible to generate an appropriate answer that takes dependency into consideration. For example, if the answer to the decomposed query A is "get an influencer to introduce the product," the generation control unit 103A may generate an answer by inputting this answer and the decomposed query B into the generation model. This makes it possible to generate an answer such as "sales can be expected to increase at low cost," for example.

[0068] (Allocation Based on Comprehensive Evaluation) As described above, the allocation unit 102A allocates decomposed queries based on the suitability evaluation results obtained by the suitability evaluation unit 106A. Also, as described above, the allocation unit 102A may allocate decomposed queries using load information acquired by the load information acquisition unit 107A. The allocation unit 102A may then allocate decomposed queries using both the suitability evaluation results and the load information. This makes it possible to allocate decomposed queries to generative models that are suitable for the decomposed queries while distributing the load in generation using each generative model.

[0069] For example, both the compatibility evaluation result and the load information may be expressed as numerical values. In this case, the allocation unit 102A may calculate, for each combination of a decomposed query to be allocated and a plurality of generative models, the difference between the numerical value indicating the compatibility evaluation result and the numerical value indicating the load information (hereinafter referred to as the overall evaluation value). The allocation unit 102A may then allocate the generative model with the highest overall evaluation value to the decomposed query. For example, if the compatibility evaluation result between a decomposed query X and a generative model 21A is 0.8 (the larger this numerical value, the higher the compatibility) and the load information of the generative model 21A is 0.2 (the larger this numerical value, the higher the load), the overall evaluation value of the combination of the decomposed query X and the generative model 21A is 0.6. In this case, if the compatibility evaluation result between the decomposed query X and the generative model 22A is 0.6 and the load information of the generative model 22A is 0.1, the overall evaluation value of the combination of the decomposed query X and the generative model 22A is 0.5. Therefore, in this example, the allocation unit 102A allocates the decomposed query X to the generative model 21A that has a larger overall evaluation value.

[0070] (Method for Evaluating Compatibility Between Decomposed Query and Generative Model) Any method can be applied as the method for evaluating compatibility by the compatibility evaluation unit 106A as long as it can obtain valid evaluation results. For example, the compatibility evaluation unit 106A may evaluate the compatibility between the decomposed query and the generative model using query feature information indicating the features of the decomposed query and model feature information indicating the features of the generative model. This provides the effect of enabling allocation taking into consideration the features of both the decomposed query and the generative model, in addition to the effect achieved by the information processing device 1.

[0071] The query feature information may be generated by the feature information generation unit 105A. The feature information generation unit 105A generates query feature information indicating features of the decomposed query. The method for generating the query feature information is not particularly limited. For example, the feature information generation unit 105A may generate the query feature information of the decomposed query using a feature extraction model that has been machine-learned to generate query feature information (e.g., a feature vector) indicating features of the input query.

[0072] On the other hand, the model feature information may be generated in advance and stored in the storage unit 11A or the like. The method for generating the model feature information is not particularly limited. For example, the model feature information may be information generated using training data used in machine learning of the generative model. This provides, in addition to the effects achieved by the information processing device 1, an effect of enabling allocation that takes into account the characteristics of both the decomposition query and the generative model by utilizing the training data.

[0073] For example, if the generative model is a large-scale language model, a large-scale text corpus is used for training. By analyzing sentences and words contained in such a text corpus, it is possible to generate model feature information that indicates the characteristics of the generative model (large-scale language model) generated by training using such a text corpus.

[0074] For example, characteristic words may be extracted from the text corpus used for training using a technique such as TF-IDF (Term Frequency - Inverse Document Frequency), and some or all of the extracted words may be used as model feature information. Alternatively, some or all of the extracted words may be synthesized to generate model feature information. For example, some or all of the extracted words may be input to a language model, and words or sentences derived from those words may be output, and the output words or sentences may be used as model feature information.

[0075] Furthermore, for example, multiple generative models may be artificially classified, in which case the artificial classification serves as model feature information. For example, classifications such as "good at answering mathematical questions," "good at everyday conversation," and "fast processing" may serve as model feature information. Sentences contained in the above-described text corpus may be input to a language model to determine which classification they fall into or their degree of conformance to each classification. Furthermore, when a fine-tuned language model is used as the generative model, model feature information may be generated from the training data used for fine-tuning.

[0076] When a feature vector generated using a feature extraction model is used as the query feature information, the model feature information may also be a feature vector generated using the feature extraction model. This makes it possible to easily determine the similarity between the query feature information and the model feature information within the same feature space. This also makes it possible to assign a generative model having similar features to the decomposed query to the decomposed query.

[0077] For example, the model feature information may be a feature vector obtained by inputting words and sentences extracted from a text corpus or the like as described above and metadata of the generative model (indicating the application field and characteristics (e.g., response speed and accuracy) of the generative model) into a feature extraction model. The compatibility evaluation unit 106A may then calculate the similarity (e.g., cosine similarity) between the query feature information of the decomposed query and the model feature information of each generative model as a value indicating the compatibility between the decomposed query and the generative model. In this case, the allocation unit 102A may allocate the decomposed query to the generative model with the highest similarity. In this case, queries with similar meanings are allocated to the same generative model.

[0078] Additionally, queries previously input to a generative model and answers previously generated by the generative model can also be said to indicate the characteristics of the generative model. Therefore, model feature information may be generated using at least one of queries previously input to the generative model and answers generated in response to those queries.

[0079] For example, the model feature information may be words or sentences extracted from queries previously input to the generative model and answers generated in response to those queries, or feature vectors obtained by inputting those words or sentences into a feature extraction model. Alternatively, for example, answers that best represent the characteristics of the generative model may be extracted from answers previously output by the generative model, the similarity between that answer and other answers may be calculated, and the number or percentage of answers whose similarity is equal to or greater than a threshold may be used as the conformance to the features. For example, suppose that 100 answers generated by the generative model 22A include one that indicates the recommended exercise menu shown in FIG. 3, and that 59 of the other 99 answers have a similarity to this answer that is equal to or greater than a threshold. In this case, the model feature information of the generative model 22A may indicate a conformance to the "exercise menu generation" of 60%. In this method, the model feature information is generated based on answers actually generated by the generative model, so the model feature information reflects the results of directly evaluating the generative model.

[0080] (Updating the feature extraction model) The update unit 108A updates the feature extraction model. As described above, the feature extraction model is a machine-learned model used when generating query feature information from a decomposed query. More specifically, the update unit 108A updates the feature extraction model based on the evaluation results for the answer generated using the generative model. In addition to the effects achieved by the information processing device 1, the information processing device 1A including the update unit 108A has the effect of being able to generate more suitable answers as the information processing device 1A repeatedly presents answers.

[0081] The answer may be evaluated by the user. In this case, if the user gives a positive evaluation to the answer presented to the user by the presentation control unit 104A, the update unit 108A updates the feature extraction model so that the degree of similarity between the query feature information of the decomposed query generated by the feature extraction model and the model feature information of the generation model to which the decomposed query is input increases. Conversely, if the user gives a negative evaluation, the update unit 108A updates the feature extraction model so that the degree of similarity between the query feature information of the decomposed query generated by the feature extraction model and the model feature information of the generation model to which the decomposed query is input decreases.

[0082] The evaluation of the answer may be performed automatically, without relying on the user's evaluation (which may also be referred to as feedback). In this case, the index for evaluating the answer and the method for calculating the index may be determined in advance. For example, the generated answer may be input into a language model to generate information indicating whether the answer is good or bad.

[0083] (Other Allocation Examples) The method of allocating decomposed queries is not limited to the above example. For example, the allocation unit 102A may determine the allocation of decomposed queries using a model obtained by reinforcement learning. In this case, the "state" in reinforcement learning may be represented by various information that is preferably considered when determining an appropriate allocation. Examples of the information include the decomposed query to be allocated, the query before decomposition, the user to whom the answer is to be presented, and the history of past interactions with the user.

[0084] In addition, the "action" in reinforcement learning can be the selection of a generative model, and the "reward" can be the evaluation result of the generated answer. When determining allocation using a model obtained by reinforcement learning, updating the model based on the evaluation result of the generated answer can achieve the same effect as updating a feature extraction model.

[0085] The allocating unit 102A may also determine the allocation of decomposed queries according to a strategy obtained by online learning. In this case, the strategy can be updated based on the evaluation results of the generated answers, as in the case of applying reinforcement learning.

[0086] (Method of Integrating Answers) As described above, the presentation control unit 104A integrates answers generated by multiple generative models. The method of integrating answers is not particularly limited. For example, the presentation control unit 104A may integrate answers generated by multiple generative models by listing the answers in parallel. For example, in the example of FIG. 3 , the answers displayed on the terminal device 3A are answers generated by the generative model 21A and answers generated by the generative model 22A listed in parallel.

[0087] Furthermore, for example, the presentation control unit 104A may integrate answers using a trained model that has been machine-learned to enable the integration of multiple answers to generate a single cohesive answer. This trained model may be, for example, a general-purpose language model, a general-purpose language model that has been fine-tuned for answer integration, or a language model dedicated to answer integration that has been generated for answer integration.

[0088] Furthermore, the multiple generation models used to generate an answer may include multiple language models. In this case, the allocation unit 102A may allocate one decomposed query to multiple language models. In this case, the presentation control unit 104A may generate an answer by combining candidate outputs predicted by each of the multiple language models for one decomposed query. This provides the same effect as the information processing device 1, and also provides the effect of enabling the generation of more diverse and flexible answers.

[0089] Generally, a language model repeats the process of predicting a probability distribution for each of the candidate outputs that may follow an input string, indicating the probability that each candidate output follows the input string, and selecting a candidate output based on the predicted probability distribution, thereby generating an answer consisting of an array of multiple candidate outputs.

[0090] Therefore, when multiple language models are caused to generate answers for a single decomposed query, each language model predicts a probability distribution of candidate outputs. Therefore, the presentation control unit 104A can sample candidate outputs to be used as components of an integrated answer from among the candidate outputs of each language model based on these probability distributions. The generation control unit 103A then causes each language model to generate a probability distribution of the candidate output following the sampled candidate output. Similarly, the presentation control unit 104A performs sampling and the generation control unit 103A performs generation of probability distributions repeatedly to generate an integrated answer.

[0091] For example, suppose a decomposed query such as "What are popular pets in City X?" is input to language model A and language model B. Language model A infers that the probability that the candidate output of "cat" will follow the decomposed query is 0.4, while language model B infers that the probability that the candidate output of "dog" will follow the decomposed query is 0.3. In this case, the presentation control unit 104A may determine that the candidate output of "cat," which has a higher probability, will be the output following the decomposed query. By repeating this process, an answer that combines the outputs of language models A and B can be generated.

[0092] Any sampling method, including known techniques, can be applied. For example, a greedy method that always selects the most probable candidate output, a method that randomly selects a candidate output based on probability, or top-k sampling that selects a candidate output from the top k most probable texts can be applied.

[0093] Furthermore, for example, the presentation control unit 104A may generate an integrated answer by beam search. In beam search, the presentation control unit 104A searches multiple candidate outputs generated by each language model and outputs the most likely output sequence as an answer. Specifically, the presentation control unit 104A stores the candidate outputs generated by each language model in multiple candidate sequences called beams. For example, if the beam width is 3, three candidate outputs are stored as candidate sequences at each step. In the next step, the presentation control unit 104A adds the candidate outputs generated by the language model to each candidate output stored as a candidate sequence and calculates a score (likelihood). The top three candidate outputs with the highest scores are then selected as a new beam, and the process proceeds to the next step. The presentation control unit 104A repeats the above process and finally selects the sequence of candidate outputs with the highest score as the output. In beam search, the outputs of the language models are combined to generate an optimal answer that is in line with the context.

[0094] The presentation control unit 104A may also integrate numerical values ​​included in each answer generated by multiple generative models for one decomposition query. For example, suppose that in response to a decomposition query such as "What is the probability of rain in Tokyo tomorrow?", generative model A generates the answer "60%," and generative model B generates the answer "80%." In this case, the presentation control unit 104A may calculate the arithmetic mean of the numerical values ​​included in each answer to generate the answer "70%." Furthermore, if reliability is set for generative models A and B, the presentation control unit 104A may calculate a weighted average value by applying a weight according to the reliability, and generate an answer including the calculated weighted average value.

[0095] (Processing Flow) The processing flow executed by the information processing device 1A will be described with reference to Fig. 5. Fig. 5 is a flow diagram showing the processing flow executed by the information processing device 1A. Note that the flow in Fig. 5 includes each step of the generation control method according to this exemplary embodiment.

[0096] In S11, the query decomposing unit 101A accepts a query input by a user. In S12 (query decomposing process), the query decomposing unit 101A decomposes the query accepted in S11 to generate a plurality of decomposed queries.

[0097] In S13, the feature information generation unit 105A generates query feature information indicating features of the decomposed query generated in S12. For example, the feature information generation unit 105A may generate the query feature information by inputting the decomposed query generated in S12 into a feature extraction model. Note that this process is performed for each of the multiple decomposed queries generated in S12.

[0098] In S14, the compatibility evaluation unit 106A evaluates the compatibility with the plurality of generative models for each of the plurality of decomposed queries generated in S12. For example, the compatibility evaluation unit 106A may calculate the similarity between the query feature information generated in S13 and the model feature information of each generative model stored in advance in the storage unit 11A or the like, and use the calculated similarity as the evaluation result of the compatibility.

[0099] In S15, the load information acquisition unit 107A acquires load information for each of the multiple generative models, indicating the load status for answer generation using the generative model. Note that the processing of S15 only needs to be performed before S16, and does not necessarily have to be performed immediately before S16. However, it is preferable to acquire load information indicating the load status as close as possible to the timing when the generative model is caused to generate an answer. From this perspective, it is preferable to acquire the load information immediately before S16.

[0100] In S16 (allocation processing), the allocation unit 102A allocates each of the multiple decomposition queries generated in S12 to one of the multiple generative models based on the evaluation result of S14 and the load information acquired in S15. At this time, the allocation unit 102A performs allocation so that the evaluation result of S14 is high and so that an excessive load is not placed on some generative models (or so that the load is distributed).

[0101] In S17 (generation control process), the generation control unit 103A transmits a decomposition query to each generative model in accordance with the allocation determined in the process of S16, and causes the model to generate an answer.

[0102] In S18, the presentation control unit 104A integrates the answers generated in the process of S17 and presents them to the user, thereby ending the process in Fig. 5. After the process of S18, the process may return to S11 and accept input of a new query.

[0103] [Example of Using a Trained Model Other Than a Generative Model] As described above, the trained model used by the information processing device 1A is not limited to a generative model, and the information processing device 1A can also use an inference model such as a prediction model or a classification model. For example, assume that a query such as "Predict store X's sales this month and plan next month's ordering policy for product A based on the sales" is input. In this case, the query decomposition unit 101A decomposes this query into a decomposed query Q1 of "Predict store X's sales this month" and a decomposed query Q2 of "Plan next month's ordering policy for product A based on [the answer to Q1]."

[0104] Here, the decomposition query Q1 instructs a prediction of sales for a specific store, and it is difficult to generate an accurate answer to such a decomposition query Q1 using a general language model. Therefore, a prediction model for predicting sales for a specific store may be prepared in advance, and the allocation unit 102A may allocate the prediction model to a decomposition query instructing a prediction of sales for a specific store. This makes it possible to generate an accurate answer based on accurate prediction results. Note that the prediction model can be generated by machine learning using training data that is correlated with the prediction target (sales in the above example) (data correlated with sales at store X in the above example). Furthermore, in predictions using a prediction model, the generation control unit 103A may acquire data correlated with the prediction target and input it into the prediction model. The same applies when a classification model or the like is applied.

[0105] [Example of Software Implementation] Some or all of the functions of the information processing device 1, 1A may be implemented by hardware such as an integrated circuit (IC chip), or may be implemented by software.

[0106] In the latter case, the information processing device 1 or 1A is realized by, for example, a computer that executes instructions of a program, which is software that realizes each function. An example of such a computer (hereinafter referred to as computer C) is shown in Fig. 6. Fig. 6 is a block diagram showing the hardware configuration of computer C that functions as information processing device 1 or 1A.

[0107] The computer C includes at least one processor C1 and at least one memory C2. The memory C2 stores a program (generation control program) P for causing the computer C to operate as the information processing device 1 or 1A. In the computer C, the processor C1 reads and executes the program P from the memory C2, thereby realizing the functions of the information processing device 1 or 1A.

[0108] The processor C1 may be, for example, a central processing unit (CPU), a graphics processing unit (GPU), a digital signal processor (DSP), a micro processing unit (MPU), a floating point number processing unit (FPU), a physics processing unit (PPU), a tensor processing unit (TPU), a quantum processor, a microcontroller, or a combination thereof. The memory C2 may be, for example, a flash memory, a hard disk drive (HDD), a solid state drive (SSD), or a combination thereof.

[0109] The computer C may further include a RAM (Random Access Memory) for expanding the program P during execution and for temporarily storing various data. The computer C may also include a communication interface for transmitting and receiving data to and from other devices. The computer C may also include an input / output interface for connecting input / output devices such as a keyboard, a mouse, a display, and a printer.

[0110] The program P can also be recorded on a non-transitory, tangible recording medium M that can be read by the computer C. Such a recording medium M can be, for example, a tape, a disk, a card, a semiconductor memory, or a programmable logic circuit. The computer C can acquire the program P via such a recording medium M. The program P can also be transmitted via a transmission medium. Such a transmission medium can be, for example, a communication network or broadcast waves. The computer C can also acquire the program P via such a transmission medium.

[0111] Furthermore, each of the above functions of the information processing device 1 or 1A may be realized by a single processor provided in a single computer, by multiple processors provided in a single computer working together, or by multiple processors provided in each of multiple computers working together. Furthermore, the program for causing the information processing device 1 or 1A to realize each of the above functions may be stored in a single memory provided in a single computer, or may be distributed and stored in multiple memories provided in a single computer, or may be distributed and stored in multiple memories provided in each of multiple computers.

[0112] [Additional Notes] This disclosure includes the technologies described in the following supplementary notes. However, the present invention is not limited to the technologies described in the following supplementary notes, and various modifications are possible within the scope of the claims.

[0113] (Appendix A1) An information processing device comprising: a query decomposition unit that decomposes a query input by a user to generate a plurality of decomposed queries; an allocation unit that allocates each of the plurality of decomposed queries to one of a plurality of trained models that have been machine-learned to be able to generate an answer to the query; and a generation control unit that sends the decomposed query to each trained model in accordance with the allocation determined by the allocation unit, and causes the trained models to generate an answer.

[0114] (Appendix A2) An information processing device as described in Appendix A1, comprising a presentation control unit that integrates answers generated by the plurality of trained models and presents them to the user. (Appendix A3) An information processing device as described in Appendix A1 or A2, comprising a compatibility evaluation unit that evaluates the compatibility of each of the plurality of decomposed queries with the plurality of trained models, and the allocation unit allocates the decomposed queries based on the compatibility evaluation result by the compatibility evaluation unit.

[0115] (Supplementary Note A4) The information processing device according to Supplementary Note A3, wherein the compatibility evaluation unit evaluates compatibility between the decomposed query and the trained model using query feature information indicating features of the decomposed query and model feature information indicating features of the trained model.

[0116] (Supplementary Note A5) The information processing device according to Supplementary Note A4, wherein the model feature information is information generated using training data used in machine learning of the trained model.

[0117] (Supplementary Note A6) The information processing device according to Supplementary Note A4 or A5, wherein the query feature information is information generated from the decomposed query using a machine-learned feature extraction model, and further comprising an update unit that updates the feature extraction model based on an evaluation result of the answer.

[0118] (Supplementary Note A7) The information processing device according to any one of Supplementary Notes A1 to A6, wherein the allocation unit allocates the decomposition queries using load information indicating a load status in answer generation using each of the plurality of trained models.

[0119] (Appendix A8) The information processing device according to Appendix A2, wherein the plurality of trained models include a plurality of language models, the allocation unit allocates one of the decomposed queries to the plurality of language models, and the presentation control unit generates an answer by combining candidate outputs predicted by each of the plurality of language models for one of the decomposed queries.

[0120] (Appendix B1) A generation control method in which at least one processor executes a query decomposition process in which the processor decomposes a query input by a user to generate a plurality of decomposed queries; an allocation process in which the processor allocates each of the plurality of decomposed queries to one of a plurality of trained models that have been machine-learned to be able to generate an answer to the query; and a generation control process in which the processor sends the decomposed query to each trained model in accordance with the allocation determined by the allocation process, causing the trained model to generate an answer.

[0121] (Appendix B2) A generation control method as described in Appendix B1, in which the at least one processor includes a presentation control process that integrates answers generated by the multiple trained models and presents them to the user. (Appendix B3) A generation control method as described in Appendix B1 or B2, in which the at least one processor includes a compatibility evaluation process that evaluates the compatibility of each of the multiple decomposed queries with the multiple trained models, and in the allocation process, the at least one processor allocates the decomposed queries based on the compatibility evaluation results of the compatibility evaluation process.

[0122] (Supplementary Note B4) In the generation control method described in Supplementary Note B3, in the compatibility evaluation process, the at least one processor evaluates the compatibility between the decomposed query and the trained model using query feature information indicating features of the decomposed query and model feature information indicating features of the trained model.

[0123] (Supplementary Note B5) The generation control method according to Supplementary Note B4, wherein the model feature information is information generated using training data used in machine learning of the trained model.

[0124] (Appendix B6) The generation control method according to Appendix B4 or B5, wherein the query feature information is information generated from the decomposed query using a machine-learned feature extraction model, and the at least one processor includes an update process for updating the feature extraction model based on an evaluation result of the answer.

[0125] (Appendix B7) The generation control method according to any one of Appendices B1 to B6, wherein in the allocation process, the at least one processor allocates the decomposed queries using load information indicating the load situation in answer generation using each of the plurality of trained models.

[0126] (Appendix B8) The generation control method described in Appendix B2, wherein the plurality of learned models include a plurality of language models, and the at least one processor, in the allocation process, allocates one of the decomposed queries to the plurality of language models, and in the presentation control process, generates an answer by combining candidate outputs predicted by each of the plurality of language models for one of the decomposed queries.

[0127] (Appendix C1) A generation control program that causes a computer to function as a query decomposition unit that decomposes a query input by a user to generate a plurality of decomposed queries, an allocation unit that allocates each of the plurality of decomposed queries to one of a plurality of trained models that have been machine-learned to be able to generate answers to the query, and a generation control unit that sends the decomposed queries to each trained model in accordance with the allocation determined by the allocation unit, causing the trained models to generate answers.

[0128] (Supplementary Note C2) The generation control program according to Supplementary Note C1, which causes the computer to function as a presentation control unit that integrates answers generated by the plurality of trained models and presents the integrated answers to the user.

[0129] (Appendix C3) The generation control program according to Appendix C1 or C2, wherein the computer is caused to function as a compatibility evaluation unit that evaluates compatibility with the plurality of trained models for each of the plurality of decomposition queries, and the allocation unit allocates the decomposition queries based on a compatibility evaluation result by the compatibility evaluation unit.

[0130] (Supplementary Note C4) The generation control program according to Supplementary Note C3, wherein the compatibility evaluation unit evaluates compatibility between the decomposed query and the trained model using query feature information indicating features of the decomposed query and model feature information indicating features of the trained model.

[0131] (Supplementary Note C5) The generation control program according to Supplementary Note C4, wherein the model feature information is information generated using training data used in machine learning of the trained model.

[0132] (Appendix C6) The generation control program according to Appendix C4 or C5, wherein the query feature information is information generated from the decomposed query using a machine-learned feature extraction model, and the computer is caused to function as an update unit that updates the feature extraction model based on an evaluation result of the answer.

[0133] (Appendix C7) The generation control program according to any one of Appendices C1 to C6, wherein the allocation unit allocates the decomposition query using load information indicating a load situation in answer generation using each of the plurality of trained models.

[0134] (Appendix C8) The generation control program according to Appendix C2, wherein the plurality of trained models include a plurality of language models, the allocation unit allocates one of the decomposed queries to the plurality of language models, and the presentation control unit generates an answer by combining candidate outputs predicted by each of the plurality of language models for one of the decomposed queries.

[0135] (Appendix D1) An information processing device comprising at least one processor, the at least one processor executing a query decomposition process that decomposes a query input by a user to generate a plurality of decomposed queries; an allocation process that allocates each of the plurality of decomposed queries to one of a plurality of trained models that have been machine-learned to be able to generate an answer to the query; and a generation control process that sends the decomposed query to each trained model in accordance with the allocation determined by the allocation unit, causing the trained model to generate an answer.

[0136] The information processing device may further include a memory, and the memory may store a program for causing the at least one processor to execute each of the processes.

[0137] (Appendix D2) An information processing device described in Appendix D1, wherein the at least one processor executes a presentation control process that integrates answers generated by the multiple trained models and presents them to the user. (Appendix D3) An information processing device described in Appendix D1 or D2, wherein the at least one processor executes a compatibility evaluation process that evaluates the compatibility of each of the multiple decomposed queries with the multiple trained models, and in the allocation process, the at least one processor allocates the decomposed queries based on the compatibility evaluation results of the compatibility evaluation process.

[0138] (Supplementary Note D4) The information processing device according to Supplementary Note D3, wherein in the compatibility evaluation process, the at least one processor evaluates compatibility between the decomposed query and the trained model using query feature information indicating features of the decomposed query and model feature information indicating features of the trained model.

[0139] (Supplementary Note D5) The information processing device according to Supplementary Note D4, wherein the model feature information is information generated using training data used in machine learning of the trained model.

[0140] (Appendix D6) The information processing device according to appendix D4 or D5, wherein the query feature information is information generated from the decomposed query using a machine-learned feature extraction model, and the at least one processor executes an update process to update the feature extraction model based on an evaluation result of the answer.

[0141] (Appendix D7) The information processing device according to any one of Appendices D1 to D6, wherein in the allocation process, the at least one processor allocates the decomposition query using load information indicating the load status in answer generation using each of the plurality of trained models.

[0142] (Appendix D8) The information processing device described in Appendix D2, wherein the plurality of learned models include a plurality of language models, and wherein the at least one processor, in the allocation process, allocates one of the decomposed queries to the plurality of language models, and, in the presentation control process, generates an answer by combining candidate outputs predicted by each of the plurality of language models for one of the decomposed queries.

[0143] (Appendix E) A non-transitory recording medium having recorded thereon a generation control program that causes a computer to execute a query decomposition process that decomposes a query input by a user to generate multiple decomposed queries, an allocation process that allocates each of the multiple decomposed queries to one of multiple trained models that have been machine-learned to be able to generate answers to the query, and a generation control process that sends the decomposed queries to each trained model in accordance with the allocation determined by the allocation process, causing the trained models to generate answers.

[0144] REFERENCE SIGNS LIST 1 Information processing device 101 Query decomposition unit 102 Allocation unit 103 Generation control unit 1A Information processing device 101A Query decomposition unit 102A Allocation unit 103A Generation control unit 104A Presentation control unit 106A Matching evaluation unit 108A Update unit 21A, 22A Generative model (trained model)

Claims

1. An information processing device comprising: a query decomposition unit that decomposes a query input by a user to generate a plurality of decomposed queries; an allocation unit that allocates each of the plurality of decomposed queries to one of a plurality of trained models that have been machine-learned to generate an answer to the query; and a generation control unit that transmits the decomposed query to each trained model in accordance with the allocation determined by the allocation unit, thereby causing the trained models to generate an answer.

2. The information processing device according to claim 1, further comprising a presentation control unit that integrates answers generated by the multiple trained models and presents them to the user.

3. An information processing device as described in claim 1 or 2, further comprising a compatibility evaluation unit that evaluates compatibility of each of the plurality of decomposition queries with the plurality of trained models, and the allocation unit allocates the decomposition queries based on the compatibility evaluation result by the compatibility evaluation unit.

4. The information processing device described in claim 3, wherein the compatibility evaluation unit evaluates the compatibility between the decomposition query and the trained model using query feature information indicating the features of the decomposition query and model feature information indicating the features of the trained model.

5. The information processing device according to claim 4, wherein the model feature information is information generated using training data used in machine learning of the learned model.

6. The information processing device according to claim 4, wherein the query feature information is information generated from the decomposed query using a machine-learned feature extraction model, and further comprising an update unit that updates the feature extraction model based on an evaluation result for the answer.

7. An information processing device as described in claim 1 or 2, wherein the allocation unit allocates the decomposition queries using load information indicating the load situation in answer generation using each of the multiple trained models.

8. The information processing device of claim 2, wherein the plurality of learned models include a plurality of language models, the allocation unit allocates one of the decomposed queries to the plurality of language models, and the presentation control unit generates an answer by combining candidate outputs predicted by each of the plurality of language models for one of the decomposed queries.

9. A generation control method in which at least one processor executes the following: a query decomposition process for decomposing a query input by a user to generate a plurality of decomposed queries; an allocation process for allocating each of the plurality of decomposed queries to one of a plurality of trained models that have been machine-learned to be able to generate an answer to the query; and a generation control process for sending the decomposed query to each trained model in accordance with the allocation determined by the allocation process to generate an answer.

10. A generation control program that causes a computer to function as: a query decomposition unit that decomposes a query input by a user to generate a plurality of decomposed queries; an allocation unit that allocates each of the plurality of decomposed queries to one of a plurality of trained models that have been machine-learned to be able to generate an answer to the query; and a generation control unit that sends the decomposed queries to each trained model in accordance with the allocation determined by the allocation unit to generate an answer.

Images