Information processing device, information processing method, and program

Abstract

In order to solve the problem of improving the accuracy of relational knowledge data referenced as case examples by a language model, an information processing device (1) comprises: an acquisition unit (11) that acquires a first relational knowledge data set consisting of first relational knowledge data indicating a plurality of pieces of knowledge having a specific relationship, and also acquires information specifying knowledge context; a generation unit (12) that uses a language model to generate a second relational knowledge data set consisting of second relational knowledge data according to the context by referring to the first relational knowledge data set as case examples; and an update unit (13) that updates the first relational knowledge data set using second relational knowledge data having a specific relationship in the second relational knowledge data set.

Description

Information processing device, information processing method, and program

[0001] The present invention relates to a technique for constructing relational knowledge.

[0002] Multiple pieces of knowledge that are related to each other are called related knowledge. For example, related knowledge in a certain industry can be useful information for understanding the situation in that industry and making future predictions. Non-Patent Document 1 describes a technology that uses a language model that can execute a target language processing task by providing examples of the language processing task, and generates new related knowledge data by referring to a small amount of related knowledge data as examples.

[0003] Peter West et. al., “Symbolic Knowledge Distillation: from General Language Models to Commonsense Models” arXiv: 2110.07178v2 [cs.CL] 28 Nov 2022

[0004] However, the technology described in Non-Patent Document 1 has the problem that the generated relational knowledge data may be appropriate in one context but may not be appropriate in a different context. Also, there is the problem that the accuracy of the generated relational knowledge data may depend on the accuracy of the referenced cases. To solve these problems, it is important to improve the accuracy of the relational knowledge data referenced as cases by the language model.

[0005] One aspect of the present invention has been made in view of the above-mentioned problems, and one example of its purpose is to improve the accuracy of relational knowledge data that is referenced as cases by a language model.

[0006] An information processing device according to one aspect of the present invention includes: an acquisition means for acquiring a first relational knowledge dataset composed of first relational knowledge data indicating a plurality of pieces of knowledge having a specific relationship, and information specifying the context of the knowledge; a generation means for generating a second relational knowledge dataset composed of second relational knowledge data corresponding to the context by using a language model and referring to the first relational knowledge dataset as an example; and an update means for updating the first relational knowledge dataset by using second relational knowledge data having the specific relationship from the second relational knowledge dataset.

[0007] An information processing method according to one aspect of the present invention is an information processing method executed by one or more processors, and includes: acquiring a first relational knowledge dataset composed of first relational knowledge data indicating a plurality of pieces of knowledge having a specific relationship, and information specifying the context of the knowledge; generating a second relational knowledge dataset composed of second relational knowledge data corresponding to the context by referring to the first relational knowledge dataset as an example using a language model; and updating the first relational knowledge dataset using the second relational knowledge data having the specific relationship from the second relational knowledge dataset.

[0008] A program according to one aspect of the present invention causes a computer to function as: an acquisition means for acquiring a first relational knowledge dataset composed of first relational knowledge data indicating a plurality of pieces of knowledge having a specific relationship and information specifying the context of the knowledge; a generation means for generating a second relational knowledge dataset composed of second relational knowledge data corresponding to the context by referring to the first relational knowledge dataset as an example using a language model; and an update means for updating the first relational knowledge dataset by using the second relational knowledge data having the specific relationship from the second relational knowledge dataset.

[0009] According to one aspect of the present invention, it is possible to improve the accuracy of relational knowledge data that is referenced as a case by a language model.

[0010] FIG. 1 is a block diagram showing the configuration of an information processing device according to exemplary embodiment 1 of the present invention. FIG. 2 is a flow chart showing the flow of an information processing method according to exemplary embodiment 1 of the present invention. FIG. 3 is a block diagram showing the configuration of an information processing system according to exemplary embodiment 2 of the present invention. FIG. 4 is a flow chart showing the flow of an information processing method according to exemplary embodiment 2 of the present invention. FIG. 5 is a schematic diagram showing the transition of information generated in each step of the information processing method shown in FIG. 4. FIG. 6 is a schematic diagram continuing from FIG. 5. FIG. 7 is a diagram showing a specific example of a first relational knowledge dataset. FIG. 8 is a diagram showing a specific example of a context list. FIG. 9 is a diagram explaining a specific example of a first prompt. FIG. 10 is a diagram explaining a specific example of a second antecedent data list. FIG. 11 is a diagram explaining a specific example of a second prompt. FIG. 12 is a diagram showing a specific example of a second relational knowledge dataset. FIG. 13 is a diagram explaining a specific example of a third relational knowledge dataset. FIG. 14 is a diagram explaining a specific example of a third prompt. FIG. 15 is a block diagram showing an example of the hardware configuration of an information processing device according to each exemplary embodiment of the present invention.

[0011] [First Exemplary Embodiment] A first exemplary embodiment of the present invention will be described in detail with reference to the drawings. This exemplary embodiment is a basic form of the exemplary embodiments described below.

[0012] (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 an acquisition unit 11, a generation unit 12, and an update unit 13.

[0013] The acquisition unit 11 acquires a first relational knowledge dataset composed of first relational knowledge data indicating a plurality of pieces of knowledge having a specific relationship, and information specifying the context of the knowledge. The generation unit 12 uses a language model to refer to the first relational knowledge dataset as an example, thereby generating a second relational knowledge dataset composed of second relational knowledge data corresponding to the context. The update unit 13 updates the first relational knowledge dataset using second relational knowledge data having a specific relationship from the second relational knowledge dataset.

[0014] (Example of Implementation by Program) When the above-described information processing device 1 is configured by a computer having one or more processors and a memory, this exemplary embodiment can be realized by a program stored in the memory. The program causes the computer to function as an acquisition unit 11 that acquires a first relational knowledge dataset composed of first relational knowledge data indicating a plurality of pieces of knowledge having a specific relationship and information specifying the context of the knowledge, a generation unit 12 that uses a language model to generate a second relational knowledge dataset composed of second relational knowledge data according to the context by referring to the first relational knowledge dataset as an example, and an update unit 13 that updates the first relational knowledge dataset by using the second relational knowledge data having a specific relationship from the second relational knowledge dataset.

[0015] (Flow of information processing method S1) The information processing device 1 configured as described above executes the information processing method S1 according to this exemplary embodiment. When the information processing device 1 described above is configured by a computer including one or more processors and a memory, the information processing method S1 is realized, for example, by the one or more processors reading and executing the above-described program.

[0016] The flow of information processing method S1 will be described with reference to FIG. 2. FIG. 2 is a flow diagram showing the flow of information processing method S1. Information processing method S1 includes steps S11, S12, and S13. In step S11, the acquisition unit 11 acquires a first relational knowledge dataset composed of first relational knowledge data indicating a plurality of pieces of knowledge having a specific relationship, and information specifying the context of the knowledge. In step S12, the generation unit 12 uses a language model to refer to the first relational knowledge dataset as an example, thereby generating a second relational knowledge dataset composed of second relational knowledge data corresponding to the context. In step S13, the update unit 13 updates the first relational knowledge dataset using second relational knowledge data having a specific relationship from the second relational knowledge dataset.

[0017] (Effects of the Present Exemplary Embodiment) As described above, the present exemplary embodiment employs a configuration in which a first relational knowledge dataset composed of first relational knowledge data indicating a plurality of pieces of knowledge having a specific relationship, and information specifying the context of the knowledge are acquired, and a language model is used to reference the first relational knowledge dataset as cases to generate a second relational knowledge dataset composed of second relational knowledge data according to the context, and the first relational knowledge dataset is updated using the second relational knowledge data having the specific relationship from among the second relational knowledge dataset. Therefore, according to the present exemplary embodiment, it is possible to obtain the effect of improving the accuracy of the relational knowledge data referenced as cases by the language model.

[0018] [Embodiment 2] An information processing system 100 including an information processing device 1A according to a third embodiment of the present invention will be described in detail with reference to the drawings. Note that components having the same functions as those described in embodiment 1 will be assigned the same reference numerals, and descriptions thereof will be omitted as appropriate.

[0019] Overview of Information Processing System 100 The information processing system 100 is a system that generates relational knowledge data having a specific relationship in a specified context using a language model 3A. Here, the relational knowledge data is composed of multiple pieces of knowledge having some kind of relationship. The relational knowledge data having a specific relationship is composed of multiple pieces of knowledge having a specific relationship. Furthermore, the knowledge is information expressed in a natural language sentence. Furthermore, the context is information that specifies the scope or conditions under which a specific relationship holds between multiple pieces of knowledge. Examples of context include, but are not limited to, industry type, country, person, and era. In this exemplary embodiment, a relationship in which a context is defined, such as a "causal relationship" or "temporal relationship," is applied as the "specific relationship." Among multiple pieces of knowledge having a specific relationship, knowledge in a "before" context is referred to as "antecedent knowledge." Furthermore, knowledge in a "after" context is referred to as "consequent knowledge." In other words, a relationship in which a context is defined is a relationship in which the original relationship does not hold if the antecedent knowledge and the consequent knowledge are swapped. For example, when applying "causal relationship" as a "specific relationship," knowledge of the antecedent indicates the cause, and knowledge of the consequent indicates the effect.

[0020] <Configuration of Information Processing System 100> Fig. 3 is a block diagram showing the configuration of the information processing system 100. As shown in Fig. 3, the information processing system 100 includes an information processing device 1A, a relational knowledge database 2A, and a language model 3A. The information processing device 1A is communicably connected to each of the relational knowledge database 2A and the language model 3A.

[0021] (Relational Knowledge Database 2A) The relational knowledge database 2A is a device that stores relational knowledge data having a specific relationship. The relational knowledge data in this exemplary embodiment includes antecedent data, consequent data, and context data. The antecedent data is a natural language sentence that indicates knowledge of the antecedent. The consequent data is a natural language sentence that indicates knowledge of the consequent. The context data is a natural language sentence that indicates a context related to knowledge of the antecedent and consequent.

[0022] (Language Model 3A) The language model 3A is a general-purpose large-scale language model trained to perform a language processing task by referencing examples of the task. Hereinafter, examples are also referred to as samples. For example, the language model 3A may be a model that receives a prompt as input and generates information corresponding to the prompt. The prompt and the generated information are each natural language sentences. The prompt includes instructions, samples, and queries. The instructions indicate content that instructs the language model 3A on the information to be generated by the language model 3A. The language model 3A generates information corresponding to the query by referencing samples as examples based on the instructions. The language model 3A may also output a score indicating the appropriateness of the information along with the generated information. An example of the language model 3A includes, but is not limited to, GTP3 (Generative Pre-trained Transformer 3).

[0023] (Information processing device 1A) As shown in Fig. 3, information processing device 1A includes a control unit 110 and a storage unit 120. Control unit 110 controls each unit of information processing device 1A. Storage unit 120 stores various data used by control unit 110. Control unit 110 includes an acquisition unit 11A, a generation unit 12A, a determination unit 16A, and an update unit 13A.

[0024] The acquiring unit 11A acquires a first relational knowledge dataset and information specifying a context, similar to the acquiring unit 11 in exemplary embodiment 1. The context list shown in Fig. 3 is an example of information specifying a context, and includes one or more contexts. In this exemplary embodiment, the first relational knowledge data constituting the first relational knowledge dataset includes first antecedent data indicating knowledge of an antecedent, and first consequent data indicating knowledge of a consequent having a specific relationship to the knowledge of the antecedent.

[0025] The generation unit 12A generates a second relational knowledge dataset according to the context list by referring to the first relational knowledge dataset as an example using the language model 3A. The generation unit 12A also generates a new second relational knowledge dataset by referring to the first relational knowledge dataset updated by the update unit 13A as an example.

[0026] For example, the generation unit 12A includes a first generation unit 14A and a second generation unit 15A. The first generation unit 14A generates second antecedent data corresponding to a context by referring to the first antecedent data as a case. The second generation unit 15A generates second consequent data corresponding to the second antecedent data by referring to the first relational knowledge data as a case. In this way, the generation unit 12A generates second relational knowledge data including the second antecedent data and the second consequent data.

[0027] The determination unit 16A determines whether the second relationship knowledge data has the different relationship by using the language model 3A and referring to a third relationship knowledge data set constituted by the third relationship knowledge data as an example. The third relationship knowledge data indicates a plurality of pieces of knowledge having a relationship different from the specific relationship.

[0028] The updating unit 13A is configured similarly to the updating unit 13 in the first exemplary embodiment, and in addition, identifies "second relationship knowledge data having a specific relationship" to be used for updating the first relationship knowledge dataset as follows. For example, the updating unit 13A updates the first relationship knowledge dataset by using, from the second relationship knowledge dataset, second relationship knowledge data that satisfies a condition based on the score calculated by the language model 3A as second relationship knowledge data having a specific relationship. Furthermore, the updating unit 13A updates the first relationship knowledge dataset by using, from the second relationship knowledge dataset, second relationship knowledge data selected by the user as second relationship knowledge data having a specific relationship.

[0029] Furthermore, the updating unit 13A updates the first relationship knowledge dataset by using, from the second relationship knowledge dataset, the second relationship knowledge data that has been determined not to have the above-mentioned "different relationship" as the second relationship knowledge data having the specific relationship. Note that, to identify the "second relationship knowledge data having the specific relationship" to be used in updating the first relationship knowledge dataset, the updating unit 13A may use any one of a score-based condition, a selection by the user, and a determination by the determining unit 16A, or may use a combination of some or all of them.

[0030] Furthermore, the updating unit 13A updates the third relationship knowledge dataset using the second relationship knowledge data that has been determined to have the above-mentioned "different relationship" from among the second relationship knowledge dataset. Note that the updating unit 13A may use a score-based condition, a user selection, or a combination of these, instead of or in addition to the determination by the determining unit 16A, to identify the "second relationship knowledge data having a different relationship" to be used in updating the third relationship knowledge dataset.

[0031] The storage unit 120 stores first instruction data, second instruction data, third instruction data, a second relational knowledge dataset, and a third relational knowledge dataset. Of the information stored in the storage unit 120, the first instruction data, the second instruction data, the third instruction data, and the third relational knowledge dataset are stored before execution of an information processing method S1A (described below) and are updated by execution of the information processing method S1A. The second relational knowledge dataset does not need to be stored before execution of the information processing method S1A, but is generated by execution of the information processing method S1A. In addition, the first relational knowledge dataset and a context list are input to the information processing device 1A.

[0032] <Flow of Information Processing Method S1A> The information processing device 1A configured as described above executes the information processing method S1A according to this exemplary embodiment. The information processing method S1A will be described with reference to FIGS. 4 to 6. FIG. 4 is a flow diagram showing the flow of the information processing method S1A. FIG. 5 is a schematic diagram showing the transition of information generated in each step of the information processing method S1A. FIG. 6 is a schematic diagram following FIG. 5.

[0033] (Step S101) In step S101, the acquisition unit 11A acquires a first relational knowledge data set and a context list.

[0034] (First relational knowledge dataset) The first relational knowledge dataset includes information to be input as a sample to the language model 3A. The information processing method S1A can be executed repeatedly, and the first relational knowledge dataset acquired during the first execution may be created by a person or by another device. Furthermore, the first relational knowledge dataset acquired after the second execution is updated by the previous execution of the information processing method S1A.

[0035] The first relational knowledge data set is composed of one or more first relational knowledge data, each of which has a specific relationship. In this exemplary embodiment, the first relational knowledge data includes first antecedent data, first consequent data, and first context data.

[0036] FIG. 7 is a diagram showing a specific example of a first relational knowledge data set. As shown in FIG. 7, the first relational knowledge data set includes first relational knowledge data D11, D12, D13, etc. When there is no need to particularly distinguish between them, each will also be referred to as first relational knowledge data D1. The first relational knowledge data D1 has a "causal relationship," which is an example of a "specific relationship." The first relational knowledge data D1 includes a first item "INDUSTRIY," a second item "CAUSE," and a third item "EFFECT." In the first relational knowledge data D1, the natural language sentence following the item name "INDUSTRIY" is an example of first context data and indicates the type of industry. The natural language sentence following the item name "CAUSE" is an example of first antecedent data and indicates knowledge of an antecedent that is a cause in the industry. The natural language sentence following the item name "EFFECT" is an example of first consequent data and indicates knowledge of a consequent that is a result of the cause.

[0037] For example, the first relational knowledge data D11 indicates that in the context "agriculture" indicated by the first context data, there is a causal relationship between the antecedent knowledge indicated by the first antecedent data, "There is a growing focus on functional fruits / Can clearly state the relationship between fruits and health." and the consequent knowledge indicated by the first consequent data, "Enhance and add a variety of nutrients to fruits."

[0038] 7, for the first relational knowledge data D12 and D13, meaningless character strings such as "xxxxx" are shown as the first context data, first antecedent data, and first consequent data, but these character strings are merely schematic representations of natural language sentences. In the following drawings, such meaningless character strings are also assumed to represent natural language sentences in a schematic manner.

[0039] (Context List) A context list is an example of information that specifies a context, and includes information that specifies one or more contexts. Hereinafter, each context indicated by the context list will also be referred to as a "specified context," and information indicating a specified context will also be referred to as "specified context data." The specified context data is input to the language model 3A as a query. FIG. 8 is a diagram showing a specific example of a context list. In the example of FIG. 8, the context list includes multiple specified context data expressed in natural language sentences, such as "agriculture," "fishing," "forestry," etc.

[0040] (Steps S102 to S103: Generation of Second Antecedent Data) Steps S102 to S103 will be described with reference again to Figures 4 and 5. The first generation unit 14A executes steps S102 to S103 for each context specified in the context list to generate a second antecedent data list made up of second antecedent data corresponding to each specified context. Specifically, first, in step S102, the first generation unit 14A generates a first prompt to be input to the language model 3A.

[0041] (First Prompt) The first prompt is information input to the language model 3A to generate second antecedent data according to the specified context. The configuration of the first prompt will be described with reference to Fig. 5. As shown in Fig. 5, the first prompt generated in step S102 includes first instruction data as an "instruction," first antecedent data and first context data as a "sample," and specified context data as a "query."

[0042] For example, the first generation unit 14A extracts some or all of the first relational knowledge data from the first relational knowledge dataset and includes the first context data and the first antecedent data included in each extracted first relational knowledge data as samples in the first prompt. Furthermore, for example, the first generation unit 14A includes the specified context data as a "query" in the first prompt. Furthermore, for example, the first generation unit 14A includes the first instruction data stored in the storage unit 120 as an "instruction" in the first prompt. The first instruction data is a natural language sentence indicating content that instructs the language model 3A on the information to be generated by the first prompt, and is stored in advance in the storage unit 120 as a template.

[0043] 9 is a diagram illustrating a specific example of a first prompt. In the example of FIG. 9, the first instruction data includes a natural language sentence that indicates that the information to be generated by the first prompt includes a first item "INDUSTRIY" and a second item "STEEP," and that explains the definition of each item. Note that the item name "STEEP" may be any item name that corresponds to the first antecedent data, and is not limited to the example described above.

[0044] In the example of FIG. 9 , the samples included in the first prompt include sample SMP11. Sample SMP11 includes the first context data in the first relational knowledge data D12 shown in FIG. 7 as the first item "INDUSTRIY" and the first antecedent data as the second item "STEEP." The first relational knowledge data D12 used to generate sample SMP11 may be extracted from the first relational knowledge dataset by the first generation unit 14A. The extraction process may be performed randomly, or samples that satisfy a predetermined condition may be extracted. For example, the predetermined condition may be a condition based on a score, a generation date and time, a past extraction history, or the like.

[0045] The first prompt generation process in step S102 may be performed individually for each context specified in the context list, or may be performed once for some or all of the multiple contexts. If performed individually, the samples included in the first prompts generated for each specified context may be different from each other. This reduces bias in the content of the samples and enables the generation of more diverse second antecedent data. Note that, although the example in FIG. 9 shows one sample SMP11 included in the first prompt, the first prompt may include multiple samples.

[0046] 9, the query included in the first prompt includes specified context data as the first item "INDUSTRIY," but includes only the item name as the second item "STEEP" without any natural language sentence following the item name. Such a query indicates a request to generate a natural language sentence indicating "STEEP" that corresponds to the context "fishing" indicated by the specified context data.

[0047] Next, in step S103, the first generation unit 14A generates second antecedent data according to the specified context by inputting the generated first prompt into the language model 3A. By performing the processes of steps S102 to S103 for each context specified by the context list, a second antecedent data list is generated as shown in FIG.

[0048] (Second Antecedent Data List) The second antecedent data list includes one or more pieces of second antecedent data generated for each specified context. FIG. 10 is a diagram showing a specific example of the second antecedent data list. As shown in FIG. 10, the second antecedent data list includes second antecedent data and second context data indicating the context of the knowledge indicated by the second antecedent data. The specified context data included as a query in the first prompt to generate the second antecedent data is applied as the second context data. The first line in FIG. 10 is generated, for example, by inputting the first prompt shown in FIG. 9 to the language model 3A. Specifically, based on the first instruction data shown in FIG. 9 and using a sample as an example, the language model 3A generates the second item "The use of robots for fishing purposes is becoming more widespread. Robots can be used for a variety of tasks." in response to the first item "fishing" included in the query. The natural language sentence generated as the second item data is an example of second antecedent data. In other words, by using the first antecedent data as a sample, "antecedent knowledge" that may be a cause of the specified context "fishing" is generated. Note that the first generation unit 14A may repeat step S103 multiple times for one context. As a result, multiple pieces of second antecedent data may be generated for one context.

[0049] (Steps S104 to S105: Generation of Second Relational Knowledge Data Set) Steps S104 to S105 will be described again with reference to Figures 4 and 5. The second generation unit 15A executes steps S104 to S105 for each piece of second antecedent data included in the second antecedent data list to generate a second relational knowledge data set made up of second relational knowledge data including each piece of second antecedent data. Specifically, first, in step S104, the first generation unit 14A generates a second prompt to be input to the language model 3A.

[0050] (Second Prompt) The second prompt is information input to the language model 3A in order to generate second consequent data corresponding to the specified second antecedent data. The configuration of the second prompt will be described with reference to Fig. 5. As shown in Fig. 5, the second prompt generated in step S104 includes second instruction data as an "instruction," first relational knowledge data as a "sample," and second context data and second antecedent data as a "query."

[0051] For example, in step S104, the first generation unit 14A extracts some or all of the first relational knowledge data from the first relational knowledge dataset and includes each extracted first relational knowledge data as a sample in the second prompt. Also, for example, the first generation unit 14A includes the second context data and the second antecedent data included in the second antecedent data list as a "query" in the second prompt. Also, for example, the second generation unit 15A includes the second instruction data stored in the storage unit 120 as an "instruction" in the second prompt. The second instruction data is a natural language sentence indicating content that instructs the language model 3A on the information to be generated by the second prompt, and is stored in advance in the storage unit 120 as a template.

[0052] 11 is a diagram illustrating a specific example of the second prompt. In the example of Fig. 11, the second instruction data includes natural language sentences that explain that the information to be generated by the second prompt includes the first item "INDUSTRIY," the second item "CAUSE," and the third item "EFFECT," as well as the definitions of each item.

[0053] In the example of FIG. 11 , the samples included in the second prompt include sample SMP21. Sample SMP21 includes the first relational knowledge data D13 shown in FIG. 7 as the first item "INDUSTRIY," the second item "CAUSE," and the third item "EFFECT." The first relational knowledge data D13 used to generate sample SMP21 may be extracted from the first relational knowledge dataset by the second generation unit 15A. The extraction process may be performed randomly, or may extract data that meets a predetermined condition. For example, the predetermined condition may be a condition based on the score, the generation date and time, past extraction history, etc.

[0054] The second prompt generation process in step S104 may be performed individually for each second antecedent data item included in the second antecedent data list, or may be performed once for some or all of the multiple second antecedent data items. If performed individually, the samples included in the second prompts generated for each second antecedent data item may be different from each other. This reduces bias in the content of the samples, allowing for the generation of more diverse second consequent data. Note that, although the example in FIG. 11 shows one sample SMP21 included in the second prompt, the number of samples included in the second prompt may be multiple.

[0055] 11, the query included in the second prompt includes second context data as the first item "INDUSTRIY" and second antecedent data as the second item "CAUSE." Meanwhile, the query includes only an item name as the third item "EFFECT" and does not include a natural language sentence following the item name. This query requests the generation of a natural language sentence indicating "EFFECT" that corresponds to the context "fishing" indicated by the second context data and the second antecedent data "The use of robots...."

[0056] Next, in step S105, the second generation unit 15A generates second consequent data corresponding to the second context data and the second antecedent data by inputting the generated second prompt into the language model 3A. By executing the processes of steps S104 to S105 for each second antecedent data included in the second antecedent data list, a second relational knowledge data set corresponding to the specified context is generated, as shown in FIG.

[0057] (Second Relational Knowledge Dataset) The second relational knowledge dataset includes one or more second antecedent data included in the second antecedent data list, second context data corresponding to the second antecedent data, and second consequent data corresponding to the second antecedent data. FIG. 12 is a diagram showing a specific example of the second relational knowledge dataset. As shown in FIG. 12, the second relational knowledge dataset includes second relational knowledge data D21, D22, .... When there is no need to particularly distinguish between them, each will also be referred to as second relational knowledge data D2. The second relational knowledge data D2 includes second context data, second antecedent data, and second consequent data. In other words, the second relational knowledge dataset is configured by adding a column for "second consequent data" to the second antecedent data list shown in FIG. 10. The second relationship knowledge data D2 is generated by the language model 3A using the first relationship knowledge data D1 as a sample, and therefore is likely to have a "specific relationship" like the first relationship knowledge data D1, but does not necessarily have a "specific relationship." Therefore, the subsequent steps S106 to S108 are executed to perform processing according to the relationship possessed by the second relationship knowledge data.

[0058] (Steps S106 to S108: Determining Relationships in Second Relation Knowledge Data) Steps S106 to S108 will be described with reference again to FIGS. 4 and 6. By executing these steps, the determination unit 16A determines, using the language model 3A, whether the relationship in each piece of second relation knowledge data is a "specific relationship," a "different relationship," or an "other relationship." Hereinafter, a plurality of pieces of knowledge having a "specific relationship" will also be referred to as positive examples, and a plurality of pieces of knowledge having a "different relationship" will also be referred to as negative examples. Specifically, first, in step S106, the determination unit 16A generates a third prompt to be input to the language model 3A. Furthermore, the third relation knowledge dataset is used in generating the third prompt.

[0059] (Third relational knowledge dataset) The third relational knowledge dataset includes information to be input as a sample to the language model 3A. The information processing method S1A can be executed repeatedly, and the third relational knowledge dataset stored in the storage unit 120 before the first execution may be created by a person or by another device. Furthermore, the third relational knowledge dataset used from the second time onwards is the one updated by the previous execution of the information processing method S1A.

[0060] The third relationship knowledge data set indicates one or more pieces of third relationship knowledge data. As described above, the third relationship knowledge data indicates a plurality of pieces of knowledge having a relationship different from the specific relationship. In this exemplary embodiment, the third relationship knowledge data includes third context data, third antecedent data, and third consequent data. The third antecedent data and the third consequent data have a "different relationship" in the context indicated by the third context data. For example, when a "causal relationship" is applied as the "specific relationship," the "different relationship" may be, but is not limited to, a "synonymous relationship," an "entailment relationship," or the like. Here, the "different relationship" may be a relationship in which a context is not defined, such as a "synonymous relationship" or an "entailment relationship." A relationship in which a context is not defined is a relationship in which the "different relationship" holds even if the third antecedent data and the third consequent data are interchanged. However, the "different relationship" is not limited to this and may also be a relationship in which a context is defined.

[0061] FIG. 13 is a diagram showing a specific example of the third relational knowledge dataset. As shown in FIG. 13, the third relational knowledge dataset includes third relational knowledge data D31, D32, D33, etc. When there is no need to particularly distinguish between them, each will also be referred to as third relational knowledge data D3. The third relational knowledge data D3 has a "synonymous relationship," which is an example of a "different relationship." The third relational knowledge data D3 includes a first item "INDUSTRIY," a second item "SENTENCE1," and a third item "SENTENCE2." The natural language sentence following the item name "INDUSTRIY" is an example of third context data and indicates the type of industry. The natural language sentence following the item name "SENTENCE1" is an example of third antecedent data and indicates one of multiple pieces of knowledge having a "different relationship." The natural language sentence following the item name "SENTENCE2" is an example of third consequent data, and indicates the other of a plurality of pieces of knowledge having "different relationships."

[0062] For example, the third relational knowledge data D31 indicates that in the context "Nursing Care Business" indicated by the third context data, the knowledge indicated by the third antecedent data, "The use of robots in nursing care is becoming more widespread." and the knowledge indicated by the third consequent data, "Recently the use of robots in nursing care is very popular." have a synonymous relationship.

[0063] (Third Prompt) The third prompt includes information to be input to the language model 3A in order to determine whether the second relational knowledge data is a negative example having a "different relationship." The configuration of the third prompt will be described with reference to Fig. 6. As shown in Fig. 6, the third prompt generated in step S106 includes third instruction data as "instruction," third relational knowledge data as "negative example samples," first relational knowledge data as "non-negative example samples," and second relational knowledge data as "query."

[0064] For example, the determination unit 16A extracts some or all of the third relational knowledge data from the third relational knowledge dataset and includes each extracted third relational knowledge data as a negative example sample in the third prompt. The determination unit 16A also extracts some or all of the first relational knowledge data from the first relational knowledge dataset and includes each extracted first relational knowledge data as a non-negative example sample in the third prompt. For example, the determination unit 16A also includes each second relational knowledge data included in the second relational knowledge dataset as a "query" in the third prompt. For example, the determination unit 16A also includes third instruction data stored in the storage unit 120 as an "instruction" in the third prompt. The third instruction data is a natural language sentence indicating content that instructs the language model 3A on the information to be generated by the third prompt, and is stored in advance in the storage unit 120 as a template.

[0065] 14 is a diagram illustrating a specific example of the third prompt. In the example of FIG. 14, the third instruction data indicates that the information to be generated by the third prompt includes the first item "INDUSTRIY," the second item "SENTENCE1," the third item "SENTENCE2," and the fourth item "Label," and includes natural language sentences explaining the definitions of each item. In this example, the third instruction data includes a natural language sentence explaining that, as the definition of the fourth item "Label," "e" indicates a different relationship and "x" indicates that the relationship is not different.

[0066] In the example of FIG. 14, the samples included in the third prompt include a negative example sample SMP31 and a non-negative example sample SMP32. The negative example sample SMP31 includes the third relationship knowledge data D31 shown in FIG. 13 as the first to third items, and includes "e" as the fourth item "Label." Here, "e" is a label indicating a "different relationship." The non-negative example sample SMP32 includes the first relationship knowledge data D12 shown in FIG. 7 as the first to third items, and includes "x" as the fourth item "Label." Here, "x" is a label indicating a "not different relationship."

[0067] Here, the third relational knowledge data D31 for generating the negative example sample SMP31 may be extracted from the third relational knowledge dataset by the determination unit 16A. Furthermore, the first relational knowledge data D12 for generating the non-negative example sample SMP32 may be extracted from the first relational knowledge dataset by the determination unit 16A. The extraction process may be performed randomly, or data that satisfies a predetermined condition may be extracted. For example, the predetermined condition may be a condition based on the score, the generation date and time, past extraction history, etc.

[0068] The third prompt generation process in step S106 may be performed individually for each piece of second relationship knowledge data included in the second relationship knowledge data set, or may be performed once for some or multiple pieces of second relationship knowledge data. When performed individually, the samples included in the third prompts generated for each piece of second relationship knowledge data may be different from each other. This reduces bias in the content of the samples and enables more accurate determination of whether or not a sample is a negative example. While the example in FIG. 14 shows one negative example sample SMP31 included in the third prompt, the third prompt may contain multiple negative example samples. Furthermore, while the third prompt shows one non-negative example sample SMP32 included in the third prompt, the third prompt may contain multiple non-negative example samples.

[0069] 14, the query included in the third prompt includes the second relational knowledge data as the first to third items, but includes only the item name as the fourth item, "Label," without any natural language sentence following the item name. Such a query indicates a request to generate a natural language sentence indicating "Label" that corresponds to the second relational knowledge data included in the query.

[0070] Next, in step S107, the determination unit 16A inputs the generated third prompt into the language model 3A, thereby determining whether each piece of second relationship knowledge data included in the second relationship knowledge data set is a negative example. For example, the determination unit 16A determines that second relationship knowledge data in which "e" is generated as the "Label" is a negative example (i.e., has a "different relationship"). Also, for example, the determination unit 16A determines that second relationship knowledge data in which "x" is generated as the "Label" is not a negative example (i.e., does not have a different relationship).

[0071] Next, in step S108, the determination unit 16A determines whether the second relationship knowledge data determined not to be a negative example in step S107 is a positive example or a negative example based on further input from the user. Note that the user may input annotation information indicating whether the second relationship knowledge data determined not to be a negative example is a positive example having a specific relationship, a negative example having a different relationship, or something else. This makes it possible to identify "positive examples having a specific relationship" with greater accuracy than when all second relationship knowledge data determined not to be "negative examples" by the language model 3A are treated as positive examples.

[0072] In this way, as shown in steps S107 to S108 of Figure 6, each piece of second relationship knowledge data included in the second relationship knowledge data set is classified as a positive example or a negative example. Second relationship knowledge data classified as a positive example is determined not to be a negative example by the language model 3A and selected by the user as a positive example. Second relationship knowledge data classified as a negative example includes data determined to be a negative example by the language model 3A. Furthermore, second relationship knowledge data classified as a negative example may include data determined not to be a negative example by the language model 3A but selected by the user as a negative example.

[0073] (Steps S109 to S111: Updating) Steps S109 to S111 will be described with reference again to Figures 4 and 6. Steps S109 to S111 are steps for updating the first relationship knowledge dataset, the third relationship knowledge dataset, and the relationship knowledge database 2A. As shown in steps S109 to S111 in Figure 6, the second relationship knowledge data classified as a positive example is used to update the first relationship knowledge dataset and the relationship knowledge database 2A. Furthermore, the second relationship knowledge data classified as a negative example is used to update the third relationship knowledge dataset.

[0074] In step S109, the update unit 13A updates the third relationship knowledge dataset using the second relationship knowledge data classified as a negative example from the second relationship knowledge dataset. "Updating the third relationship knowledge dataset" may mean, for example, adding the second relationship knowledge data classified as a negative example to the third relationship knowledge dataset as new third relationship knowledge data. Furthermore, "updating the third relationship knowledge dataset" may mean, for example, replacing part or all of the third relationship knowledge dataset with the second relationship knowledge data classified as a negative example as new third relationship knowledge data.

[0075] Furthermore, the language model 3A may output a score indicating the appropriateness of the fourth item "Label" generated for each piece of second relationship knowledge data in response to the third prompt. In this case, the update unit 13A may update the third relationship knowledge data set using, from among the second relationship knowledge data classified as negative examples, the score that satisfies a predetermined condition (e.g., the score is equal to or greater than a threshold, the score is within a predetermined rank from the top, or is within a predetermined percentage, etc.). Furthermore, the update unit 13A may update the third relationship knowledge data set using, from among the second relationship knowledge data classified as negative examples, the second relationship knowledge data that is selected by the user.

[0076] By executing step S109, the first relational knowledge data set to be referenced the next time the information processing method S1A is repeated is updated, thereby improving the accuracy of the second prompts generated based on the first relational knowledge data set, thereby improving the accuracy of the generated second relational knowledge data as positive examples.

[0077] In step S110, the update unit 13A updates the first relationship knowledge dataset using the second relationship knowledge data classified as a positive example from the second relationship knowledge dataset. "Updating the first relationship knowledge dataset" may mean, for example, adding the second relationship knowledge data classified as a positive example to the first relationship knowledge dataset as new first relationship knowledge data. Furthermore, "updating the first relationship knowledge dataset" may mean, for example, replacing part or all of the first relationship knowledge dataset with the second relationship knowledge data classified as a positive example as new first relationship knowledge data. The updated first relationship knowledge dataset may be stored in the storage unit 120.

[0078] Furthermore, the generation unit 12A may output a score indicating the appropriateness of the second item "STEEP" (i.e., the second relational knowledge data) generated for each context in response to the second prompt. In this case, the update unit 13A may update the first relational knowledge dataset using, among the second relational knowledge data classified as positive examples, the score of which satisfies a predetermined condition (e.g., the score is equal to or greater than a threshold, the score is within a predetermined rank from the top, or is within a predetermined percentage, etc.). Furthermore, the update unit 13A may update the first relational knowledge dataset using, among the second relational knowledge data classified as positive examples, the second relational knowledge data selected by the user.

[0079] By executing step S109, the third relational knowledge data set to be referenced the next time the information processing method S1A is repeated is updated, thereby improving the accuracy of the third prompts generated based on the third relational knowledge data set. As a result, the accuracy of determining negative examples in the second relational knowledge data set is improved. This makes it possible to accurately exclude negative examples from the second relational knowledge data set.

[0080] In step S111, the update unit 13A updates the relationship knowledge database 2A by using the second relationship knowledge data classified as a positive example from the second relationship knowledge data set. For example, the update unit 13A adds the second relationship knowledge data classified as a positive example to the relationship knowledge database 2A as new “relationship knowledge data having a specific relationship.”

[0081] The information processing device 1A then repeats the information processing method S1A. In the next iteration of the information processing method S1A, in step S101, the acquisition unit 11A acquires an updated first relational knowledge dataset. In steps S102 and S104, the generation unit 12A generates a first prompt and a second prompt using the updated first relational knowledge dataset. As a result, in step S105, a new second relational knowledge dataset is generated that references the updated first relational knowledge dataset as a case. In step S106, the determination unit 16A generates a third prompt using the updated first relational knowledge dataset and the updated third relational knowledge dataset.

[0082] By repeating the information processing method S1A in this manner, it is possible to generate a second relational knowledge data set with higher accuracy as positive examples according to each specified context, and to more accurately determine and exclude negative examples from the second relational knowledge data set, thereby accurately updating the relational knowledge database 2A.

[0083] <Effects of the Present Exemplary Embodiment> As described above, the present exemplary embodiment employs a configuration in which a new second relationship knowledge dataset is generated by referring to the updated first relationship knowledge dataset as a case. Therefore, according to the present exemplary embodiment, since the first relationship knowledge dataset, which has higher accuracy as a dataset having a specific relationship, can be referred to as a case, it is possible to generate, with high accuracy, relationship knowledge data having a specific relationship to be added to the relationship knowledge database 2A in accordance with the specified context.

[0084] Furthermore, the relationship knowledge data accumulated in the relationship knowledge database 2A according to this exemplary embodiment can be utilized for judgments, analyses, etc. in a specified context. In this case, such judgments, analyses, etc. can be performed with high accuracy. For example, when a "causal relationship" is applied as the specific relationship and an "industry type" is applied as the context, the relationship knowledge data accumulated in the relationship knowledge database 2A can be utilized for management judgments, industry analyses, etc. in the specified industry. In this case, changes / transformations in the industry can be accurately grasped, and future changes accompanying such changes can be accurately predicted.

[0085] Furthermore, this exemplary embodiment employs a configuration in which, of the second relationship knowledge dataset, second relationship knowledge data that satisfies a condition based on the score calculated by the language model 3A is used as second relationship knowledge data having a specific relationship, and the first relationship knowledge dataset is updated. Therefore, according to this exemplary embodiment, the first relationship knowledge dataset can be updated using second relationship knowledge data that is highly likely to have a specific relationship based on the score, and therefore the accuracy of the first relationship knowledge dataset that is referenced as cases having a specific relationship can be improved.

[0086] Furthermore, this exemplary embodiment employs a configuration in which the second relationship knowledge data set selected by the user from the second relationship knowledge data set is used as the second relationship knowledge data having a specific relationship to update the first relationship knowledge data set. Therefore, according to this exemplary embodiment, the first relationship knowledge data set can be updated using the second relationship knowledge data selected by the user as having a specific relationship, thereby improving the accuracy of the first relationship knowledge data set that is referenced as cases having a specific relationship.

[0087] Furthermore, this exemplary embodiment employs a configuration in which, by using a language model 3A, a third relationship knowledge dataset composed of third relationship knowledge data indicating multiple pieces of knowledge having a relationship different from the specific relationship is referenced as an example to determine whether the second relationship knowledge data has a different relationship, and the second relationship knowledge data determined not to have a different relationship in the second relationship knowledge dataset is used as the second relationship knowledge data having the specific relationship to update the first relationship knowledge dataset. Therefore, according to this exemplary embodiment, the same language model 3A is used to exclude negative examples from the second relationship knowledge dataset generated using the language model 3A, thereby ensuring accuracy. Furthermore, since there is no need to use a different model to exclude negative examples from the second relationship knowledge dataset generated using the language model 3A, negative examples can be more easily excluded.

[0088] Furthermore, this exemplary embodiment employs a configuration in which the third relationship knowledge dataset is updated using second relationship knowledge data determined to have a different relationship from the second relationship knowledge dataset. Therefore, according to this exemplary embodiment, the accuracy of the third relationship knowledge dataset having a relationship different from the specific relationship, which is referenced as a case to exclude negative examples from the second relationship knowledge dataset generated using the language model 3A, is improved, and as a result, it is possible to accurately extract relationship knowledge data having a specific relationship to be added to the relationship knowledge database 2A.

[0089] In addition, in this exemplary embodiment, the first relational knowledge data includes first antecedent data indicating knowledge of an antecedent and first consequent data indicating knowledge of a consequent having a specific relationship to the antecedent knowledge, and the second relational knowledge data including the second antecedent data and the second consequent data is generated by using the language model 3A to generate second antecedent data according to the context by referring to the first antecedent data as a case, and second consequent data according to the generated second antecedent data by referring to the first relational knowledge data as a case. Therefore, according to this exemplary embodiment, the second antecedent data according to the specified context is first created, and then the second consequent data according to the second antecedent data is generated, so that the second relational knowledge data according to the specified context can be generated more accurately.

[0090] [Modification 1] The above-described exemplary embodiment 2 can be modified to apply multiple relationships as relationships different from the specific relationship. In this modification, an example will be described in which a first relationship and a second relationship are applied as the different relationships.

[0091] In this modification, the third relationship knowledge data set includes third relationship knowledge data having a first relationship different from the specific relationship, and third relationship knowledge data having a second relationship different from both the specific relationship and the first relationship. For example, when a "causal relationship" is applied as the "specific relationship," a "synonymous relationship" may be applied as the "first relationship," and an "implication relationship" may be applied as the second relationship. However, the first relationship and the second relationship are not limited to these examples.

[0092] In this modification, the determination unit 16A determines whether each piece of second relationship knowledge data included in the second relationship knowledge data set has either the first relationship or the second relationship. Hereinafter, a plurality of pieces of knowledge having the first relationship will also be referred to as first negative examples, and a plurality of pieces of knowledge having the second relationship will also be referred to as second negative examples.

[0093] For example, in this modification, the samples included in the third prompt are modified as follows: Specifically, the determination unit 16A extracts some or all of the third relationship knowledge data having the first relationship from the third relationship knowledge dataset, and includes each extracted piece of third relationship knowledge data as a first negative example sample in the third prompt; and the determination unit 16A extracts some or all of the third relationship knowledge data having the second relationship from the third relationship knowledge dataset, and includes each extracted piece of third relationship knowledge data as a second negative example sample in the third prompt.

[0094] Furthermore, for example, in the third instruction data as the "instruction" included in the third prompt, the definition of the fourth item "Label" is transformed as follows: For example, the definition of the fourth item "Label" is transformed into a natural language sentence that explains that "e" indicates the first negative example, "a" indicates the second negative example, and "x" indicates neither the first nor the second negative example.

[0095] In this way, the determination unit 16A determines whether each piece of second relationship knowledge data included in the second relationship knowledge data set is a first negative example, a second negative example, or another, by inputting the third prompt into the language model 3A. Furthermore, the determination unit 16A may further determine whether the second relationship knowledge data determined by the language model 3A to be neither a first negative example nor a second negative example is a "positive example," a "first negative example," a "second negative example," or "other," based on a user input.

[0096] In this modification, the update unit 13A updates the first relational knowledge dataset by using the second relational knowledge data determined to be neither the first nor the second negative example as the second relational knowledge data having a specific relationship. Specifically, the update unit 13A updates the first relational knowledge dataset by using the second relational knowledge data selected by the user as a "positive example" from the second relational knowledge data determined to be neither the first nor the second negative example. The update unit 13A also updates the third relational knowledge dataset by using the second relational knowledge data selected by the user as a "first negative example" from the second relational knowledge data determined to be neither the first nor the second negative example. The update unit 13A also updates the third relational knowledge dataset by using the second relational knowledge data selected by the user as a "second negative example" from the second relational knowledge data determined to be neither the first nor the second negative example.

[0097] According to this modification, multiple types of negative examples are determined from the second relation knowledge data set as not having a specific relationship and are to be excluded, so that negative examples not having a specific relationship can be excluded with higher accuracy. Note that although an example in which there are two types of negative examples has been described here, there may be three or more types of negative examples.

[0098] [Modification 2] The above-described modification 2 can be further modified so that the determining unit 16A determines in stages whether an example is a first negative example and whether an example is a second negative example.

[0099] In this modification, the determination unit 16A uses a first third prompt and a second third prompt. The first third prompt can be similarly explained by replacing "different relationships" with "first relationships" in the description of the third prompt in exemplary embodiment 2. The second third prompt can be similarly explained by replacing "different relationships" with "second relationships" in the description of the third prompt in exemplary embodiment 2.

[0100] The determination unit 16A first determines whether each piece of second relationship knowledge data is a first negative example by inputting a first third prompt into the language model 3A. Furthermore, the determination unit 16A determines whether each piece of second relationship knowledge data determined not to be a first negative example is a second negative example by inputting a second third prompt into the language model 3A. This allows the determination unit 16A to determine whether each piece of second relationship knowledge data is a first negative example, a second negative example, or something else.

[0101] According to this modified example, multiple types of negative examples that are excluded from the second relational knowledge dataset as not having a specific relationship are determined in stages, so that only a binary determination is required in one stage, allowing for more accurate determination of negative examples.

[0102] Although the example in which the determination is performed in two stages has been described above, if there are three or more types of negative examples, the number of determination stages may also be three or more. Furthermore, the number of types of negative examples does not necessarily have to match the number of stages. For example, the determination unit 16A may determine whether each piece of second relation knowledge data is a first negative example at one determination stage, and whether it is a second or third negative example at another determination stage.

[0103] [Other Modifications] In the above-described exemplary embodiment 2, an example has been described in which "user selection," "score-based conditions," and "exclusion of negative examples by the language model 3A" are all applied in combination to identify second relationship knowledge data of positive examples to be used for updating the first relationship knowledge dataset. However, it is also possible to modify the example so that second relationship knowledge data of positive examples is identified by applying some of these and not applying others.

[0104] In the above-described exemplary embodiment 2, an example has been described in which "user selection," "score-based conditions," and "determination of negative examples by the language model 3A" are all applied in combination to identify second relationship knowledge data as negative examples to be used for updating the third relationship knowledge dataset. However, it is also possible to modify this example so that second relationship knowledge data as negative examples are identified by applying some of these and not applying others.

[0105] In the above-described exemplary embodiment 2, the first relational knowledge data, the second relational knowledge data, and the third relational knowledge data are each described as data indicating a relationship between two pieces of knowledge. However, these relational knowledge data may be data indicating a relationship between three or more pieces of knowledge.

[0106] [Example of Software Implementation] Some or all of the functions of the information processing devices 1 and 1A (hereinafter, each will be referred to as the device) may be implemented by hardware such as an integrated circuit (IC chip), or by software.

[0107] In the latter case, the device is realized, for example, by 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. 15. Computer C includes at least one processor C1 and at least one memory C2. Memory C2 stores a program P for operating computer C as the device. In computer C, processor C1 reads and executes program P from memory C2, thereby realizing each function of the device.

[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] [Additional Note 1] The present invention is not limited to the above-described embodiments, and various modifications are possible within the scope of the claims. For example, embodiments obtained by appropriately combining the technical means disclosed in the above-described embodiments are also included in the technical scope of the present invention.

[0112] [Additional Note 2] Part or all of the above-described embodiment can also be described as follows: However, the present invention is not limited to the following described aspects.

[0113] (Supplementary Note 1) An information processing device comprising: an acquisition means for acquiring a first relational knowledge dataset constituted by first relational knowledge data indicating a plurality of pieces of knowledge having a specific relationship, and information specifying a context of the knowledge; a generation means for generating a second relational knowledge dataset constituted by second relational knowledge data corresponding to the context by referring to the first relational knowledge dataset as an example using a language model; and an update means for updating the first relational knowledge dataset by using second relational knowledge data having the specific relationship from among the second relational knowledge dataset.

[0114] (Supplementary Note 2) The information processing device according to Supplementary Note 1, wherein the generating means generates the new second relational knowledge dataset by referring to the updated first relational knowledge dataset as a case.

[0115] (Supplementary Note 3) The information processing device according to Supplementary Note 1 or 2, wherein the update means updates the first relationship knowledge dataset by using, from the second relationship knowledge dataset, second relationship knowledge data that satisfies a condition based on the score calculated by the language model as second relationship knowledge data having the specific relationship.

[0116] (Supplementary Note 4) The information processing device according to any one of Supplementary Notes 1 to 3, wherein the update means updates the first relationship knowledge dataset using second relationship knowledge data selected by a user from the second relationship knowledge dataset as second relationship knowledge data having the specific relationship.

[0117] (Supplementary Note 5) The information processing device according to any one of Supplementary Notes 1 to 4, further comprising: a determination means for determining whether the second relationship knowledge data has the different relationship by using the language model and referring as an example to a third relationship knowledge data set constituted by third relationship knowledge data indicating a plurality of pieces of knowledge having a relationship different from the specific relationship; and the update means updates the first relationship knowledge data set by using, from the second relationship knowledge data set, second relationship knowledge data determined not to have the different relationship as second relationship knowledge data having the specific relationship.

[0118] (Supplementary Note 6) The information processing device according to Supplementary Note 5, wherein the update means updates the third relationship knowledge dataset using the second relationship knowledge data determined to have the different relationship among the second relationship knowledge dataset.

[0119] (Supplementary Note 7) The information processing device according to Supplementary Note 5 or 6, wherein the third relationship knowledge dataset includes third relationship knowledge data having a first relationship different from the specific relationship, and third relationship knowledge data having a second relationship different from both the specific relationship and the first relationship, the determination means determines whether the second relationship knowledge data has either the first relationship or the second relationship, and the update means updates the first relationship knowledge dataset by using the second relationship knowledge data determined to have neither the first relationship nor the second relationship as second relationship knowledge data having the specific relationship.

[0120] (Supplementary Note 8) The information processing device according to any one of Supplementary Notes 1 to 7, wherein the first relational knowledge data includes first antecedent data indicating knowledge of an antecedent, and first consequent data indicating knowledge of a consequent having the specific relationship to the antecedent knowledge, and the generating means includes: first generating means for generating second antecedent data corresponding to the context by referring to the first antecedent data as a case using the language model; and second generating means for generating second consequent data corresponding to the second antecedent data by referring to the first relational knowledge data as a case, and generates the second relational knowledge data including the second antecedent data and the second consequent data.

[0121] (Supplementary Note 9) An information processing method executed by one or more processors, comprising: acquiring a first relational knowledge dataset constituted by first relational knowledge data indicating a plurality of pieces of knowledge having a specific relationship, and information specifying a context of the knowledge; generating a second relational knowledge dataset constituted by second relational knowledge data according to the context by referring to the first relational knowledge dataset as an example using a language model; and updating the first relational knowledge dataset by using the second relational knowledge data having the specific relationship from among the second relational knowledge dataset.

[0122] (Supplementary Note 10) A program that causes a computer to function as: an acquisition means that acquires a first relational knowledge dataset composed of first relational knowledge data indicating a plurality of pieces of knowledge having a specific relationship, and information that specifies the context of the knowledge; a generation means that uses a language model to generate a second relational knowledge dataset composed of second relational knowledge data according to the context by referring to the first relational knowledge dataset as an example; and an update means that updates the first relational knowledge dataset by using the second relational knowledge data having the specific relationship from among the second relational knowledge dataset.

[0123] [Additional Note 3] Part or all of the above-described embodiment can also be expressed as follows.

[0124] An information processing apparatus comprising at least one processor, the processor executing: an acquisition process for acquiring a first relational knowledge dataset constituted by first relational knowledge data indicating a plurality of pieces of knowledge having a specific relationship, and information specifying a context of the knowledge; a generation process for generating a second relational knowledge dataset constituted by second relational knowledge data corresponding to the context by referring to the first relational knowledge dataset as an example using a language model; and an update process for updating the first relational knowledge dataset by using second relational knowledge data having the specific relationship from among the second relational knowledge dataset.

[0125] The information processing device may further include a memory that stores a program for causing the processor to execute the acquisition process, the generation process, and the update process. The program may be recorded on a computer-readable, non-transitory, tangible recording medium.

[0126] 1, 1A Information processing device 2A Relational knowledge database 3A Language model 11, 11A Acquisition unit 12, 12A Generation unit 13, 13A Update unit 14A First generation unit 15A Second generation unit 16A Determination unit 100 Information processing system 110 Control unit 120 Storage unit C1 Processor C2 Memory

Claims

1. an acquisition means for acquiring a first relational knowledge dataset constituted by first relational knowledge data indicating a plurality of pieces of knowledge having a specific relationship, and information specifying the context of the knowledge; a generating means for generating a second relational knowledge data set composed of second relational knowledge data corresponding to the context by referring to the first relational knowledge data set as an example using a language model; an updating means for updating the first relational knowledge data set by using second relational knowledge data having the specific relationship among the second relational knowledge data set; An information processing device comprising:

2. the generating means generates a new second relational knowledge dataset by referring to the updated first relational knowledge dataset as a case. The information processing device according to claim 1 .

3. the updating means updates the first relational knowledge data set by using, from the second relational knowledge data set, second relational knowledge data that satisfies a condition based on the score calculated by the language model as second relational knowledge data having the specific relationship.

3. The information processing device according to claim 1.

4. the updating means updates the first relationship knowledge data set by using the second relationship knowledge data selected by the user from the second relationship knowledge data set as the second relationship knowledge data having the specific relationship. The information processing device according to claim 1 .

5. a determination means for determining whether the second relational knowledge data has the different relationship by using the language model and referring to a third relational knowledge data set as an example, the third relational knowledge data set being made up of third relational knowledge data indicating a plurality of pieces of knowledge having a relationship different from the specific relationship; further comprising the updating means updates the first relationship knowledge data set by using the second relationship knowledge data determined not to have the different relationship among the second relationship knowledge data set as the second relationship knowledge data having the specific relationship. The information processing device according to claim 1 .

6. the updating means updates the third relationship knowledge data set by using the second relationship knowledge data determined to have the different relationship among the second relationship knowledge data set. The information processing device according to claim 5 .

7. the third relationship knowledge data set includes third relationship knowledge data having a first relationship different from the specific relationship, and third relationship knowledge data having a second relationship different from both the specific relationship and the first relationship, the determining means determines whether the second relation knowledge data has either the first relation or the second relation; the updating means updates the first relationship knowledge data set by using the second relationship knowledge data determined to have neither the first relationship nor the second relationship as second relationship knowledge data having the specific relationship. The information processing device according to claim 5 .

8. the first relational knowledge data includes first antecedent data indicating knowledge of an antecedent, and first consequent data indicating knowledge of a consequent having the specific relationship with the knowledge of the antecedent; The generating means uses the language model to a first generating means for generating second antecedent data according to the context by referring to the first antecedent data as a case; a second generating means for generating second consequent data corresponding to the second antecedent data by referring to the first relational knowledge data as a case; generating the second relational knowledge data including the second antecedent data and the second consequent data; The information processing device according to claim 1 .

9. 1. An information processing method executed by one or more processors, comprising: Acquiring a first relational knowledge dataset composed of first relational knowledge data indicating a plurality of pieces of knowledge having a specific relationship, and information specifying a context of the knowledge; generating a second relational knowledge dataset composed of second relational knowledge data corresponding to the context by referring to the first relational knowledge dataset as an example using a language model; updating the first relational knowledge data set using second relational knowledge data having the specific relationship among the second relational knowledge data set; An information processing method, including:

10. Computer, an acquisition means for acquiring a first relational knowledge dataset constituted by first relational knowledge data indicating a plurality of pieces of knowledge having a specific relationship, and information specifying the context of the knowledge; a generating means for generating a second relational knowledge data set composed of second relational knowledge data corresponding to the context by referring to the first relational knowledge data set as an example using a language model; an updating means for updating the first relational knowledge data set by using second relational knowledge data having the specific relationship among the second relational knowledge data set; A program that functions as a