Information processing device, method, and program

Abstract

To generate a language model that ennobles date and time adjustment.SOLUTION: An information processing device is provided, comprising an input unit for receiving date and time, a processing unit configured to generate a response assuming that the current date and time is the above mentioned date and time using a trained generative AI model configured to output a response according to an input sentence, and an output unit for outputting the response.SELECTED DRAWING: Figure 3

Description

[Technical field]

[0001] The present invention relates to an information processing device, a method, and a program. [Background technology]

[0002] Conventionally, Retrieval augument is known as data augmentation using search (Non-Patent Document 1).

[0003] There is also a technology called AutoGPT that automatically selects appropriate tasks. [Prior art documents] [Non-patent literature]

[0004] [Non-Patent Document 1] Internet Search<URL:https: / / aclanthology.org / 2022.emnlp-main.294 / > Summary of the Invention [Problem to be solved by the invention]

[0005] An object of the present invention is to provide an information processing device, method, and program capable of generating a language model that enables date and time adjustment.

[0006] An object of the present invention is to provide an information processing device, method, and program capable of inferring a score representing the importance of at least one of a problem and a solution.

[0007] An object of the present invention is to provide an information processing device, method, and program capable of inferring a solution to a problem. [Means for solving the problem]

[0008] A first aspect of the present disclosure is an information processing device that includes an input unit that accepts a date and time, a processing unit that uses a trained generative AI model that outputs a response based on an input sentence to generate the response assuming that the current date and time is the date and time, and an output unit that outputs the response.

[0009] A second aspect of the present disclosure is an information processing device that includes an input unit that accepts data representing at least one of a problem and a solution, and an inference unit that uses a trained score inference model that outputs a score representing the importance of at least one of the problem and the solution to infer a score for the accepted data representing at least one of the problem and the solution.

[0010] A third aspect of the present disclosure is an information processing device that includes an input unit that accepts data representing at least one of a problem and a solution, and an inference unit that infers a score representing the importance of at least one of the problem and the solution based on the accepted data representing at least one of the problem and the solution.

[0011] A fourth aspect of the present disclosure is an information processing device that includes an input unit that accepts data representing a problem, and an inference unit that infers a solution for the accepted data representing the problem using a trained solution inference model that outputs a solution to the problem. Effect of the Invention

[0012] According to the disclosed technology, a language model that enables date and time adjustment can be generated.

[0013] The disclosed technology allows for the inference of a score representing the importance of at least one of a problem and a solution.

[0014] The disclosed technology makes it possible to infer solutions to problems. [Brief description of the drawings]

[0015] [Figure 1]1 is a block diagram showing a configuration of an information processing system according to an embodiment of the present invention; [Diagram 2] FIG. 2 is a schematic block diagram of an example of a computer that functions as a management server of the present embodiment. [Diagram 3] FIG. 2 is a block diagram showing a configuration of a management server according to the first embodiment. [Figure 4] FIG. 11 is a block diagram showing a configuration of a management server according to a second embodiment. [Diagram 5] FIG. 13 is a block diagram showing a configuration of a management server according to a third embodiment. [Figure 6] FIG. 13 is a diagram illustrating an example of generating a search query using a generative AI model and executing a search. [Figure 7] 13 is a flowchart showing the contents of a search processing routine of a management server according to the third embodiment. [Figure 8] FIG. 13 is a block diagram showing a configuration of a management server according to a fourth embodiment. [Figure 9] FIG. 13 is a block diagram showing the configuration of a management server according to a fifth embodiment. [Figure 10] FIG. 23 is a block diagram showing the configuration of a management server according to a sixth embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0016] Hereinafter, an embodiment of the present invention will be described in detail with reference to the drawings.

[0017] First Embodiment As shown in Fig. 1, the information processing system 100 according to the first embodiment includes a management server 10 installed on the side of a service management company, and a user terminal 24 operated by a user. The management server 10 is an example of an information processing device. For simplicity, Fig. 1 shows an example in which two user terminals 24 are provided, but three or more user terminals 24 may be provided.

[0018] The management server 10 and the user terminal 24 are connected via a network 26 such as the Internet.

[0019] The user terminal 24 may be a smartphone terminal, a mobile phone, a PDA (Personal Digital Assistant) terminal, or a notebook computer terminal.

[0020] FIG. 2 is a block diagram showing the hardware configuration of the management server 10 of this embodiment.

[0021] 2, the management server 10 includes a CPU (Central Processing Unit) 11, a ROM (Read Only Memory) 12, a RAM (Random Access Memory) 13, a storage 14, an input unit 15, a display unit 16, and a communication interface (I / F) 17. Each component is connected to each other via a bus 19 so as to be able to communicate with each other.

[0022] The CPU 11 is a central processing unit, and executes various programs and controls each part. That is, the CPU 11 reads a program from the ROM 12 or the storage 14, and executes the program using the RAM 13 as a working area. The CPU 11 controls each of the above components and performs various arithmetic processing according to the program stored in the ROM 12 or the storage 14. In this embodiment, the ROM 12 or the storage 14 stores a program for performing various processing.

[0023] The ROM 12 stores various programs and various data. The RAM 13 temporarily stores programs or data as a working area. The storage 14 is configured with an HDD (Hard Disk Drive) or an SSD (Solid State Drive) and stores various programs including an operating system and various data.

[0024] The input unit 15 includes a pointing device such as a mouse, and a keyboard, and is used to perform various input operations.

[0025] The display unit 16 is, for example, a liquid crystal display, and displays various information. The display unit 16 may function as the input unit 15 by adopting a touch panel system.

[0026] The communication interface 17 is an interface for communicating with other devices, and uses standards such as Ethernet (registered trademark), FDDI, and Wi-Fi (registered trademark).

[0027] Next, a description will be given of the functional configuration of the management server 10. As shown in Fig. 3, the management server 10 functionally comprises an input unit 30, a processing unit 32, an output unit 34, and a model storage unit 52.

[0028] The input unit 30 accepts user input and date and time from the user terminal 24 .

[0029] Specifically, it accepts general date and time information, information indicating a certain point in time for a specific subject, and a trained generative AI model 52A.

[0030] The processing unit 32 uses a trained generative AI model 52A that outputs a response according to an input sentence to generate a response to a user input, assuming that the current time is the input date and time.

[0031] Specifically, a general large-scale language model (LLM) may be provided as an additional input, and the date and time may be specified by a prompt only, or an LLM with the function of allowing date and time adjustment may be obtained by directly learning it.

[0032] More specifically, when emulating a certain entity, the date and time information is used, or the memory (browsing cache log, discussion log, etc.) up to a certain point is used.

[0033] Here, the following three methods are shown as methods for realizing the generative AI model 52A of this embodiment, but the present invention is not limited to these.

[0034] Example 1: Controlled by prompts. A common LLM (e.g. chatGPT) is required as the input trained model.

[0035] Example 2: Save the model separately for each timestamp. As the input trained model, we need the LLM for each timestamp.

[0036] Example 3: Learning and modeling including time series information. Learning takes into account the time at which resources used during learning could be created (references to dates and times in text, or creation dates and times). Text data for LLM learning is required for the learning input, and large-scale language models (TMLLM: Time Machine Large Language Models) are obtained as the learning output assuming that the current time is the input date and time, and the TMLLM obtained as the learning output is required as the inference input.

[0037] The output unit 34 outputs a response generated in response to the user input to the user terminal. The output unit also outputs an LLM restricted to general date and time information or an LLM restricted to information indicating a certain point in time for a specific subject.

[0038] The model storage unit 52 stores the generative AI model 52A. The generative AI model 52A may be stored in an external device other than the management server 10. In this case, the management server 10 may transmit a prompt to the external device and acquire the output of the generative AI model 52A.

[0039] As described above, the information processing system according to this embodiment can generate a language model that allows date and time adjustment.

[0040] <Second embodiment> 1, an information processing system 200 according to the second embodiment includes a management server 210 installed on the side of a service management company, and a user terminal 24 operated by a user. The management server 210 is an example of an information processing device.

[0041] FIG. 2 is a block diagram showing the hardware configuration of the management server 210 of this embodiment.

[0042] Next, a description will be given of the functional configuration of the management server 210. As shown in Fig. 4, the management server 210 functionally comprises an input unit 230, an inference unit 232, and a model storage unit 252.

[0043] The management server 210 may manually specify the search target resource and score required for the learning data by adding a general LLM or the TMLLM output in the first embodiment as an input during learning.

[0044] The management server 210 outputs the score inference model as an output during learning.

[0045] The input unit 230 receives data representing at least one of a problem and a solution from the user terminal 24.

[0046] The inference unit 232 infers a score for data representing at least one of the received problem and solution using a trained score inference model 252A that outputs a score representing the importance of at least one of the problem and solution.

[0047] Specifically, the inference unit 232 infers the score using a trained generative AI model that outputs a response according to an input sentence as the trained score inference model 252A. Here, the generative AI model is additionally trained using a previously prepared pair of at least one of a problem and a solution, and a score.

[0048] The generative AI model is a general LLM or the TMLLM output in the first embodiment.

[0049] The inference unit 232 may output the score together with the reason why the score is high / low.

[0050] The management server 210 operates, for example, in accordance with steps 1 to 4 below.

[0051] In step 1, in order to prepare learning data, the definition of the search target resource and score for the "search target (patent, paper, press release, product, business, etc.)" is given manually or automatically. Multiple search targets may be used.

[0052] In the case of manual search, for example, if the search target is "research," "papers" are specified as the search target resource, and the score is specified as the number of citations.

[0053] In the case of automatic search, for example, the search target is input as "research," "papers" are the search target resources, and the score is estimated as the number of citations.

[0054] AutoGPT is an example of prior art that enables automatic estimation. Actions are determined by asking the LLM where to search to achieve a certain goal. Another example is a patent, where a collection of patent documents is gathered (such as Google Patent), and the score could be the performance of the company division that holds the patent or the sales of the product.

[0055] In step 2, training data is created.

[0056] For example, in the case of "research," individual papers are entered one by one from the "papers" that are the "resources to be searched" determined in step 1, and the "problem" and "solution" are extracted, and the number of citations of the paper is given as a "score" (for press, the number of citations, for services or products, the sales, etc.).

[0057] In addition, when using TMLLM, the estimated creation date and time of the resource is added (for patents, the patent application date and time; for press releases, the publication date, etc.).

[0058] In step 3, the model is trained.

[0059] The model learning guideline is based on the hypothesis that the more the "solution" matches the "problem," the higher the score will be, and this is achieved by learning a model that estimates the score using only the "problem" and "solution" as input.

[0060] In addition, when using TMLLM, it is based on the hypothesis that the more up-to-date the "problem" and "solution" are, the higher the score will be.

[0061] When using TMLLM, this is achieved by backdating the estimated creation date and time added to the training data and then training a model to estimate the score.

[0062] In this case, the TMLLM itself may have the function of directly estimating scores through supervised fine tuning, or a score estimator may be added by externally connecting it to the information output by the TMLLM (such as a text summary or embedding information regarding the reason why the score of the pair is estimated to be high / low).

[0063] In step 4, inference is performed. The problem-solution pair is used as input, and a score is inferred using the score inference model via the LLM or TMLLM. The reason why the score is estimated to be high / low may be output from the LLM / TMLLM.

[0064] When TMLLM is used, the date and time of TMLLM is regarded as the current date and time.

[0065] In addition, the same result can be achieved with just the "problem" alone, rather than the pair of "problem" and "solution," by ignoring the "solution" in step 2.

[0066] On the other hand, when evaluating the "solution" alone, this can be achieved by subtracting the score of the "challenge" alone from the score of the "challenge" / "solution" pair.

[0067] The model storage unit 252 stores the score inference model 252A. The score inference model 252A may be stored in an external device other than the management server 210. In this case, the management server 210 may transmit a prompt to the external device and acquire the output of the score inference model 252A.

[0068] The second embodiment and the first embodiment may be combined. In this case, the input unit 230 further receives a date and time, and the inference unit 232 infers a score when it is assumed that the current time is the received date and time. Specifically, the score is inferred using the TMLLM when it is assumed that the input date and time.

[0069] <Third embodiment> 1, an information processing system 300 according to the third embodiment includes a management server 310 installed on the side of a service management company, and a user terminal 24 operated by a user. The management server 310 is an example of an information processing device.

[0070] FIG. 2 is a block diagram showing the hardware configuration of the management server 310 of this embodiment.

[0071] Next, a description will be given of the functional configuration of the management server 310. As shown in Fig. 5, the management server 310 functionally comprises an input unit 330, an extraction unit 332, a keyword reception unit 334, a subset creation unit 336, a sentence generation unit 338, an evaluation unit 340, a subset extraction unit 342, a subset reception unit 344, a search unit 346, an inference unit 348, and a model storage unit 352.

[0072] The input unit 330 accepts an input sentence representing at least one of a problem and a solution from the user terminal 24. Specifically, at least one of a problem and a solution, and a generative AI model which is a general LLM or the TMLLM output in the first embodiment are accepted.

[0073] The extraction unit 332 extracts a plurality of keywords from an input sentence that represents at least one of a received problem and a solution.

[0074] For example, the extraction unit 332 uses the generative AI model 52A to extract multiple keywords K from the input sentence S. As an example, a prompt such as "Please guess three to four keywords necessary to explain this input" is input to the generative AI model 352A, and multiple keywords K are obtained from the output of the generative AI model 352A (see FIG. 6).

[0075] The method of extracting keywords is not limited to the method using the generative AI model, and other extraction methods may be used.

[0076] The keyword receiving unit 334 presents a screen displaying the multiple keywords extracted by the extraction unit 332 to the user terminal 24, and receives from the user terminal 24 a request to delete or add a keyword.

[0077] The subset creating unit 334 creates a plurality of keyword subsets based on the result of deleting or adding received keywords from the extracted plurality of keywords.

[0078] Specifically, the subset creation unit 334 mechanically combines the extracted keywords to create a subset Ks(n) (see FIG. 6). For example, as all combinations for the four keywords, 14 (=4C4+4C3+4C2+4C1) subsets Ks(1) to Ks(14) are created.

[0079] Here, when there are a large number of keywords, etc., restrictions may be imposed not on all combinations of keywords but on an arbitrary number of keywords, a threshold value, etc. For example, multiple subsets may be created with the number of keywords included in the subset within a predetermined range (for example, 3 to 4, or 3 or less).

[0080] The sentence generation unit 336 inputs a prompt to the generative AI model 352A for each of a plurality of subsets, instructing the generative AI model 352A to generate a sentence from the subset without including keywords that are not included in the subset, and obtains the generated sentence from the output of the generative AI model 352A.

[0081] Specifically, each subset is used as input, and a subset generation sentence S'(n) is generated that better represents the original input sentence and does not use keywords that are not included in the subset (see Figure 6). For example, for subsets Ks(1) to Ks(14), subset generation sentences S'(1) to S'(14) are generated.

[0082] Here, the prompt may indicate that if generating a sentence from the subset is difficult, then no sentence will be generated.

[0083] For example, a prompt such as "Based on a subset of keywords, generate a sentence that is as close as possible to the original input sentence. However, the generated sentence must be based only on the subset of keywords and cannot include other keywords. If it is difficult to generate a sentence based only on the subset of keywords, output 'N / A'" is input to the generative AI model 352A, and the generated sentence is obtained from the output of the generative AI model 352A.

[0084] The evaluation unit 340 evaluates the relevance between the input sentence S and the generated sentence S'(n) for each of the plurality of subsets.

[0085] The subset extraction unit 342 extracts at least one subset as an important keyword set in descending order of relevance.

[0086] Specifically, the subset extraction unit 342 extracts, in descending order of the score indicating the relevance, subsets whose scores indicating the relevance are equal to or greater than a threshold as important keyword sets.

[0087] The subset receiving unit 344 presents to the user terminal 24 a screen displaying the subsets extracted as important keyword sets by the subset extracting unit 342, and receives requests to delete or add subsets, or edit the order of the subsets.

[0088] The search unit 346 executes a search in the search target database using the set of important keywords extracted in descending order of relevance as a search query.

[0089] Specifically, the search unit 346 performs a search using a search engine with a set of important keywords extracted in descending order of relevance score as a search query, and a database of the search engine as a search target database. A database storing patent documents may be used as the search target database.

[0090] The inference unit 348 infers a score representing the importance of at least one of the problem and the solution based on the search result by the search unit 346.

[0091] Specifically, the inference input is matched to the search results by the search unit 346 using the LLM and its prompts, and scores for importance and novelty are assigned to the search results. The inference output may include the score as well as the reason why the score is high or low. For example, if the search results by the search unit 346 include at least one of the accepted problem and solution, a low score is assigned as the importance and novelty score.

[0092] The search result output unit 350 presents the search results by the search unit 346, divided into each subset extracted as important keywords. By presenting the results divided into each subset in this way, it becomes easier to understand which viewpoint the search results are for. In addition, when search results overlap between subsets, the search results may be integrated and presented.

[0093] The model storage unit 352 stores the generation AI model 352A. The generation AI model 352A may be stored in an external device other than the management server 310. In this case, the management server 310 may transmit a prompt to the external device and acquire the output of the generation AI model 352A.

[0094] Here, an example of the generating AI model 352A is ChatGPT (Internet search<URL: https: / / openai.com / blog / chatgpt> ), Gemini (Internet search <url: https: gemini.google.com ?hl="ja">) and other generative AIs. The generative AI model 352A is obtained by performing deep learning on a neural network. A prompt including an instruction is input to the generative AI model 352A, and inference data such as voice data indicating a voice, text data indicating a text, and image data indicating an image is input. The generative AI model 352A infers from the input inference data according to the instruction indicated by the prompt, and outputs the inference result in a data format such as voice data and text data. Here, inference refers to, for example, analysis, classification, prediction, and / or summarization.

[0095] <Function of information processing system> Next, the operation of the information processing system 300 according to this embodiment will be described.

[0096] When the management server 310 receives an input sentence indicating at least one of a problem and a solution from the user terminal 24, it executes a search processing routine shown in Fig. 7. The search processing routine is an example of an information processing method.

[0097] In step S300, the input unit 330 acquires an input sentence representing at least one of a problem and a solution received from the user terminal 24.

[0098] In step S302, the extraction unit 332 extracts a plurality of keywords from the input sentence.

[0099] In step S304, the keyword receiving unit 334 presents a screen displaying the multiple keywords extracted by the extraction unit 332 to the user terminal 24, and receives from the user terminal 24 a command to delete or add a keyword.

[0100] In step S306, the subset creating unit 336 creates a plurality of keyword subsets based on the result of deleting or adding the received keywords from the extracted plurality of keywords.

[0101] In step S308, the sentence generation unit 338 inputs a prompt to the generative AI model 352A for each of the multiple subsets, instructing the generative AI model 352A to generate a sentence from the subset without including keywords that are not included in the subset, and obtains the generated sentence from the output of the generative AI model 352A.

[0102] In step S310, the evaluation unit 340 evaluates the relevance between the input sentence S and the generated sentence S'(n) for each of the plurality of subsets.

[0103] In step S312, the subset extraction unit 342 extracts at least one subset as an important keyword set in descending order of relevance.

[0104] In step S314, the subset receiving unit 344 presents to the user terminal 24 a screen displaying the subsets extracted by the subset extraction unit 342 as important keyword sets, and accepts deletion or addition of subsets, or editing of the order of the subsets.

[0105] In step S316, the search unit 346 performs a search on the search target database using the extracted set of important keywords in descending order of relevance as a search query, reflecting the received deletion or addition of a subset, or editing of the order of the subsets.

[0106] In step S318, the inference unit 348 infers a score representing the importance of the input sentence representing at least one of the problem and the solution, based on the search result by the search unit 346.

[0107] In step S320, the search result output unit 350 outputs to the user terminal 24 a screen presenting the search results by the search unit 346, divided into each subset extracted as an important keyword set, and ends the search processing routine.

[0108] As described above, the information processing system according to this embodiment can infer a score that indicates the importance of at least one of a problem and a solution.

[0109] In addition, multiple keyword subsets are created from multiple keywords extracted from an input sentence that represents at least one of a problem and a solution, and for each subset, a sentence is generated using a generative AI model so as not to include keywords not included in the subset, and a search is performed on a search target database using the subsets in descending order of relevance to the original input sentence as important search queries. In this way, a search query can be created from an input sentence using the generative AI model, and a search can be performed.

[0110] The present invention is not limited to the above-described embodiment, and various modifications and applications are possible without departing from the spirit and scope of the present invention.

[0111] For example, in the above embodiment, the case where the relevance between the input sentence S and the generated sentence S'(n) is evaluated is described as an example, but the present invention is not limited to this. The sentence generation unit 338 may obtain the embedding of the generated sentence from the output of the generative AI model 352A, and the evaluation unit 340 may evaluate the relevance between the embedding of the input sentence and the embedding of the subset for each of the multiple subsets. In this way, the sentence generation unit 338 does not necessarily need to generate S' as a sentence, and an embedding may be used as long as it is a method that allows a comparison between Ks(n) and the original S.

[0112] The evaluation unit 340 may further evaluate the coverage of the input sentence for a set of sentences generated for each of the subsets whose relevance score is equal to or greater than the threshold. When the coverage score is less than the threshold, the evaluation unit 340 evaluates the coverage of the input sentence for a result of adding a sentence generated for a subset whose relevance score is less than the threshold to a set of sentences generated for each of the subsets whose relevance score is equal to or greater than the threshold. In addition, a subset is identified when the score indicating coverage is equal to or greater than the threshold after adding a sentence generated for a subset whose relevance score is less than the threshold, and the search unit 346 performs a search by giving priority to the identified subset. In this way, coverage may be taken into account. When S'' obtained from the set of adopted sentences S' does not cover the input sentence S and is covered by adding S'(n), S'(n) can be adopted as the search query with priority.

[0113] Moreover, the third embodiment and the first embodiment may be combined. In this case, the input unit 330 may further receive a date and time, and the inference unit 348 may infer a score when it is assumed that the current time is the received date and time. Specifically, the score may be inferred using the TMLLM when it is assumed that the input date and time.

[0114] <Fourth embodiment> 1, an information processing system 400 according to the fourth embodiment includes a management server 410 installed on the side of a service management company, and a user terminal 24 operated by a user. The management server 410 is an example of an information processing device.

[0115] FIG. 2 is a block diagram showing the hardware configuration of the management server 410 of this embodiment.

[0116] Next, a description will be given of the functional configuration of the management server 410. As shown in Fig. 8, the management server 410 functionally comprises an input unit 430, an inference unit 432, and a model storage unit 452.

[0117] The input unit 430 receives data representing the assignment from the user terminal 24 .

[0118] The inference unit 432 uses a trained solution inference model 452A that outputs a solution to a problem to infer the solution for data representing the received problem.

[0119] Specifically, the inference unit 432 infers a solution using a trained generative AI model that outputs a response according to an input sentence as the trained solution inference model 452A. Here, the generative AI model is additionally trained using a previously prepared pair of a problem and a solution.

[0120] More specifically, the system inputs a "problem" using LLM and scores such as importance, and outputs a "solution." For example, the "problem" could be "Japanese-English translation doesn't work well unless the subject is included in the machine translation," and the "solution" could be "automatically complete the Japanese subject." Note that images can also be used instead of text.

[0121] The input during learning is learning data that is a collection of multiple pairs of "solutions" for past "problems" and "LLM or TMLLM" as a generative AI model. A solution inference model that generates "solutions" from the input "problems" is trained.

[0122] For example, in the case of papers, similarly to the second embodiment, a problem and a solution are extracted from each paper, paired, and supervised learning is performed in the form of outputting the solution for the problem.

[0123] In variant 1, by adding the resource creation date and time when the learning data is created, the TMLLM is returned to the time of resource creation and supervised learning is performed by outputting solutions to problems.

[0124] In the second modification, the score used in the second embodiment may be further added to the learning data, which makes it possible to estimate data with a high score during inference.

[0125] The inputs for inference are the "problem", the solution inference model, and the "LLM or TMLLM" as a generation AI model. The inference unit generates a "solution" as the inference output based on the LLM and the score.

[0126] In the first modified example, when TMLLM is used, the specified date and time is set to the present.

[0127] In the second modification, by including the score during learning, it is possible to instruct and generate something with a higher score.

[0128] In the third modification, even if the score is not taken into account during learning, the output results may be reranked or scored during decoding using the score inference model of the second embodiment.

[0129] In the fourth modification, the output candidates are subjected to a search to evaluate their novelty score, and the novelty score is used for reranking the output results (the smaller the search count, the higher the novelty).

[0130] The model storage unit 452 stores a solution inference model 452A. The solution inference model 452A may be stored in an external device other than the management server 410. In this case, the management server 410 may transmit a prompt to the external device and acquire the output of the solution inference model 452A.

[0131] As described above, the information processing system according to this embodiment makes it possible to infer a solution to a problem.

[0132] <Fifth embodiment> 1, an information processing system 500 according to the fifth embodiment includes a management server 510 installed on the side of a service management company, and a user terminal 24 operated by a user. The management server 510 is an example of an information processing device.

[0133] FIG. 2 is a block diagram showing the hardware configuration of the management server 510 of this embodiment.

[0134] Next, a description will be given of the functional configuration of the management server 510. As shown in Fig. 9, the management server 510 functionally comprises an input unit 530, an inference unit 532, and a model storage unit 552.

[0135] The input unit 530 accepts data representing a seed representing an experience or a viewpoint from the user terminal 24 .

[0136] The inference unit 532 infers problems and solutions for data representing the received seed, using a trained problem and solution inference model 552A that outputs combinations of problems and solutions for the seed.

[0137] Specifically, the inference unit 532 infers a problem and a solution using a trained generative AI model that outputs a response according to an input sentence as the trained problem solution inference model 552A. Here, the generative AI model is additionally trained using a previously prepared pair of a seed and a combination of a problem and a solution.

[0138] More specifically, using the TMLLM of the first embodiment, a "seed representing experience or perspective" is input, and at least one of a "problem" and a "solution" is output. The input during learning is "learning data" which is a collection of multiple pairs of "problems and solutions" for past "seeds representing experience or perspectives," and the "LLM or TMLLM."

[0139] For example, in the case of papers, similar to the second embodiment, problems and solutions are extracted from each paper, and a "seed representing the experience or viewpoint that led to the paper" is added to the paper to create learning data, and supervised learning is performed. Specifically, a case is assumed in which a pair exists between a user's browser browsing history and a paper written by the user.

[0140] In variant 1, by adding the resource creation date and time when the learning data is created, the TMLLM is returned to the time of resource creation and performs supervised learning by outputting solutions to problems.

[0141] In the second modification, the score used in the second embodiment may be further added to the learning data, which makes it possible to estimate data with a high score during inference.

[0142] The inference unit 532 receives as input a "seed representing experience or perspective," a "problem-solving solution inference model," and an "LLM or TMLLM" as a generating AI model. Examples of the experience include browser browsing history and chat logs. Examples of the perspective include industry, business type, technology field, etc.

[0143] The output of the inference unit 532 is at least one of a problem and a solution. At this time, the output result of the score inference model 252A of the second embodiment may be reranked, or a score may be output.

[0144] In the first modified example, when TMLLM is used, the specified date and time is set to the present.

[0145] In the second modification, by including the score during learning, it is possible to instruct and generate something with a higher score.

[0146] In the third modification, even if the score is not taken into account during learning, the output results may be reranked or scored during decoding using the score inference model of claim 2.

[0147] In the fourth modification, the output candidates are subjected to a search to evaluate their novelty scores, and the novelty scores are used for reranking the output results (the smaller the search count, the higher the novelty).

[0148] In the fourth and fifth embodiments, novelty scores are assigned by a separate search because the update frequency of the LLM is higher than the update frequency of the search system.

[0149] If the update frequency of the LLM were to be the same as that of the search system, "novelty" would be included in the score of variant 2.

[0150] For example, when making inference, if the LLM is a model from one year ago and the search is real-time, if we only trust the LLM, there is a possibility that there will be overlaps with resources from within the past year, even if they are new from the state of the model from one year ago.

[0151] If the LLM could be updated in real time, these differences would disappear and novelty could be checked simultaneously using the TMLLM during learning.

[0152] The model storage unit 552 stores a solution to the problem inference model 552A. The solution to the problem inference model 552A may be stored in an external device other than the management server 510. In this case, the management server 510 may transmit a prompt to the external device and acquire the output of the solution to the problem inference model 552A.

[0153] As described above, the information processing system according to this embodiment can infer problems and solutions for seeds that represent experiences or perspectives.

[0154] Sixth embodiment 1, an information processing system 600 according to the sixth embodiment includes a management server 610 installed on the side of a service management company, and a user terminal 24 operated by a user. The management server 610 is an example of an information processing device.

[0155] FIG. 2 is a block diagram showing the hardware configuration of the management server 610 of this embodiment.

[0156] Next, a description will be given of the functional configuration of the management server 610. As shown in Fig. 10, the management server 610 functionally comprises an input unit 630, an inference unit 632, and a model storage unit 652.

[0157] The input unit 630 receives, from the user terminal 24, data indicating the discussion history of a project planned in the community.

[0158] The inference unit 632 infers a plan or check items for the data representing the received discussion history using a trained PDCA inference model 652A that outputs check items for checking a plan or the progress of a plan according to the discussion history.

[0159] Specifically, the inference unit 632 infers a plan or check item using a trained generative AI model that outputs a response according to an input sentence as the trained PDCA inference model 652A. Here, the generative AI model is additionally trained using a previously prepared pair of a discussion history and a plan or check item.

[0160] More specifically, during learning, the past discussion history and “LLM or TMLLM” as the generative AI model are input, and the PDCA inference model 652A is trained.

[0161] For example, there is a discussion history as the past "experience" of multiple people. For example, the purpose is to read the data of the discussion content and uncover what issues are posing and what is going well.

[0162] When a certain problem and solution are given, the management server 610 presents detailed targets including KPIs for implementing the plan as well as a proposed process for the plan.

[0163] The management server 610 accepts communication (discussion content and DAO) and numerical results for goals as execution Do from the user terminal. For example, input by the user or API linkage is used.

[0164] The management server 610 performs a numerical GAP comparison against the KPI and derives a ranking of importance as a check item Check. It also derives the cause (why it went well or not) from the contents of the discussion. Based on that, it presents a proposal for an improvement process.

[0165] The user takes an action to improve and is then observed again in the discussion as a result.

[0166] During learning, the management server 610 inputs the past discussion history and "LLM or TMLLM" and learns the PDCA inference model 652A.

[0167] In learning a plan, it is assumed that specific goals are added to the past discussion history in the plan. In addition, information indicating whether each discussion was ultimately successful or unsuccessful is also added. At this time, supervised learning is performed to present the specific goals of the plan after taking into account the situation at the time the plan is set and internally inferring what level of goal is most likely to be successful.

[0168] In the study of the check item Check, the past discussion history is assumed to include specific goals in Plan and the results of GAP analysis in Check. In addition, information indicating whether each discussion was ultimately successful or unsuccessful is also assumed to be added.

[0169] At this time, at the time of the check item "Check," the system takes into account the situation and performs supervised learning to internally infer whether the GAP can be estimated in accordance with the supervised information and what kind of improvement process would be most likely to be successful, and then presents specific goals for the Plan.

[0170] When using TMLLM, the training data is reviewed for its state at each phase before learning.

[0171] At the time of inference, the inference unit 632 receives as input the discussion history at any time (before P or before C), "LLM or TMLLM", and the PDCA inference model 652A, and generates and presents the contents of Plan or Check.

[0172] The model storage unit 652 stores the PDCA inference model 652A. The PDCA inference model 652A may be stored in an external device other than the management server 610. In this case, the management server 610 may transmit a prompt to the external device and acquire the output of the PDCA inference model 652A.

[0173] As described above, according to the information processing system of this embodiment, it is possible to infer check items for checking a plan or the progress of a plan based on the discussion history. [Explanation of symbols]

[0174] 10, 210, 310, 410, 510, 610 Management Server 11 CPU 14. Storage 15 Input section 16 Display 17 Communication Interface 24 User terminals 30, 230, 330, 430, 530, 630 Input section 32 Processing section 34 Output section 52, 252, 352, 452, 552, 652 model memory 52A, 352A Generative AI Model 100, 200, 300, 400, 500, 600 Information Processing Systems 232, 348, 432, 532, 632 Reasoning part 252A score inference model 332 Extraction part 334 Subset Creation Unit 334 Keyword Reception Department 336 Sentence generation section 336 Subset Creation Unit 338 Sentence generation part 340 Evaluation Department 342 Subset extraction part 344 Subset Reception Department 346 Search Department 350 Search result output section 452A Solution Reasoning Model 552A Problem-solving reasoning model 652A Inference Model< / url:>

Claims

1. An input unit that accepts data representing at least one of a problem and a solution; an inference unit that infers a score for data representing at least one of the received problem and the solution by using a trained score inference model that outputs a score representing the importance of at least one of the problem and the solution; An information processing device comprising:

2. The information processing device according to claim 1 , wherein the inference unit infers the score using a trained generative AI model that outputs a response according to an input sentence as the trained score inference model.

3. The information processing device according to claim 2 , wherein the generative AI model is additionally learned using a pair of at least one of a problem and a solution and the score that has been prepared in advance.

4. The input unit further receives a date and time, The information processing apparatus according to claim 1 , further comprising: a processor for inferring the score assuming that the current time is the date and time.

5. An input unit that accepts data representing at least one of a problem and a solution; an inference unit that infers a score representing the importance of at least one of the problem and the solution based on data representing at least one of the received problem and the solution; An information processing device comprising:

6. an extraction unit for extracting a plurality of keywords from the data; a subset creation unit that creates a plurality of keyword subsets from the extracted plurality of keywords; an evaluation unit that evaluates, for each of the plurality of subsets, a relevance between the data and the subset; a subset extraction unit that extracts at least one of the subsets as an important keyword set in descending order of the relevance; a search unit that performs a search using the extracted set of important keywords as a search query; The information processing apparatus according to claim 5 , wherein the inference unit infers the score based on a search result by the search unit.

7. The input unit further receives a date and time, The information processing apparatus according to claim 5 , further comprising: a step of inferring the score assuming that the current time is the date and time.

8. an input section for accepting data representing a task; an inference unit that infers a solution for the received problem using a trained solution inference model that outputs a solution to the problem; An information processing device comprising:

9. The information processing device according to claim 8 , wherein the inference unit infers the solution using a trained generative AI model that outputs a response according to an input sentence as the trained solution inference model.

10. The information processing device according to claim 9 , wherein the generative AI model is incrementally learned using pairs of problems and solutions prepared in advance.

11. an input unit that receives data representing at least one of a problem and a solution; The inference unit infers a score for the received data representing at least one of the problem and the solution using a trained score inference model that outputs a score representing the importance of at least one of the problem and the solution. Information processing methods.

12. an input unit that accepts data representing at least one of a problem and a solution; The inference unit infers a score representing the importance of at least one of the problem and the solution based on the received data representing at least one of the problem and the solution. Information processing methods.

13. An input unit accepts data representing a task; An inference unit infers a solution for the received problem using a trained solution inference model that outputs a solution to the problem. Information processing methods.

14. An information processing program for causing a computer to function as the information processing device according to any one of claims 1 to 10.

Images