Information processing device, information processing method, and information processing program

Abstract

To provide a technology about efficient learning for identifying a building type from address information.SOLUTION: An information processing device: generates a prompt including address information as input data, and a building type of a building in the address information, as correct answer data with respect to the input data; and performs prompt learning of a natural language model which is learned in advance, by using one or more prompts, so as to execute a task which outputs the building type of the building in inputted object address information.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to a technique for identifying a building type of a building based on address information. [Background technology]

[0002] In logistics delivery, information on whether or not there is a customer in the building at the destination address, or whether a delivery box is installed in the building at the destination address, is important for efficient delivery. This information can be affected by the building type of the building at the destination address. For example, if the building type of the building at the destination address is a company building, there is a high probability that customers will be present on weekdays but not on weekends. Therefore, it is important to accurately identify the building type of the building at the address.

[0003] To identify the building type of a building at an address, a property database such as that described in Patent Document 1 can be used. The database includes address information and the building type of the building at the address. By using the database, the building type for each address can be identified. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Japanese Patent Publication No. 2022-118473 Summary of the Invention [Problem to be solved by the invention]

[0005] However, when using a database such as that described in the document, the database must be updated every time address information or building type information is newly added or changed. Therefore, using a database to identify the building type of a building in address information is not efficient. Therefore, a technology is needed to efficiently identify the building type of a building in address information using means other than a database.

[0006] One way to identify the building type of a building from address information without using a database is to use a machine learning learning model that is configured to input address information, identify the building type, and output it. However, there is a problem in that the cost of training a learning model to learn a new task is generally high.

[0007] The present invention has been made in view of the above-mentioned problems, and aims to provide a technology relating to efficient learning for identifying building types from address information. [Means for solving the problem]

[0008] In order to solve the above problem, one aspect of an information processing device according to the present invention has a generation unit that generates a prompt including address information as input data and a building type of a building in the address information as correct answer data for the input data, and a learning unit that performs prompt learning of a pre-trained natural language model using one or more of the prompts so as to execute a task of outputting the building type of a building in the input target address information.

[0009] In order to solve the above problem, one aspect of the information processing method according to the present invention is an information processing method executed by an information processing device, which includes: generating, by a generation unit, a prompt including address information as input data and a building type of a building in the address information as correct answer data for the input data; and performing prompt training of a pre-trained natural language model using one or more of the prompts by a learning unit so as to execute a task of outputting the building type of a building in the input target address information.

[0010] In order to solve the above problem, one aspect of the information processing program according to the present invention is an information processing program for causing a computer to execute information processing, the program being for causing a computer to execute processing including: a generation process for generating a prompt including address information as input data and the building type of the building in the address information as correct answer data for the input data; and a learning process for performing prompt learning of a pre-trained natural language model using one or more of the prompts so as to execute a task of outputting the building type of the building in the input target address information. [Effects of the Invention]

[0011] The present invention provides a technique for efficient learning to identify building types from address information. The above-mentioned objects, aspects, and advantages of the present invention, as well as other objects, aspects, and advantages of the present invention not described above, will be understood by those skilled in the art from the following detailed description of the invention by referring to the accompanying drawings and the claims. [Brief explanation of the drawings]

[0012] [Figure 1] FIG. 1 shows an example of the functional configuration of an information processing device according to the first embodiment. [Figure 2] FIG. 2 shows an example of the contents stored in the address / building type database. [Figure 3] Figure 3 shows an example of a prompt. [Figure 4A]FIG. 4A shows an example of a prompt containing fictitious data. [Figure 4B] FIG. 4B shows another example of a prompt containing fictitious data. [Figure 4C] FIG. 4C shows another example of a prompt containing fictitious data. [Figure 5] FIG. 5 is a flowchart of the process executed by the information processing device according to the first embodiment. [Figure 6] FIG. 6 shows an example of the functional configuration of an information processing device according to the second embodiment. [Figure 7] FIG. 7 is a conceptual diagram for explaining the setting process of the priority prompt by the setting unit. [Figure 8] FIG. 8 shows an example of the hardware configuration of an information processing device according to the embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0013] Hereinafter, with reference to the accompanying drawings, an embodiment for carrying out the present invention will be described in detail. Among the components disclosed below, those having the same function will be given the same reference numerals, and their description will be omitted. Note that the embodiment disclosed below is an example of a means for realizing the present invention, and should be appropriately modified or changed depending on the configuration of the device to which the present invention is applied and various conditions. The present invention is not limited to the following embodiment. Furthermore, not all of the combinations of features described in the present embodiment are necessarily essential to the solution of the present invention.

[0014] First Embodiment An information processing device according to this embodiment generates a training dataset including a plurality of prompts, each of which includes address information and a building type of a building in the address information. The information processing device then uses the training dataset to perform prompt training (equivalent to prompt tuning) of a natural language model so as to execute a task of outputting the building type of a building in input address information. Prompt training refers to learning a method for executing a task without updating the parameters of a pre-trained natural language model based on a training dataset including one or more small number of prompts (equivalent to training data). Unlike fine tuning, this method can reduce costs associated with data collection and training. Prompt training may be performed, for example, in the form of few-shot learning (few-shot prompting), one-shot learning (one-shot prompting), or chain-of-thought learning (chain-of-thought prompting). In this embodiment, a set of one or more prompts, equivalent to training data, is simply referred to as a training dataset. Furthermore, the information processing device predicts the building type of the building in the target address information using the natural language model that has undergone prompt learning.

[0015] [Example of functional configuration of information processing device according to first embodiment] 1 shows an example of the functional configuration of an information processing device 10 according to this embodiment. As an example of its functional configuration, the information processing device 10 includes a generation unit 101, a learning unit 102, an address information acquisition unit 103, a building type prediction unit 104, a learning model storage unit 110, a learning data storage unit 120, and an address / building type database 130. The learning model storage unit 110 is configured to be able to store a natural language model 111. The natural language model 111 is a Transformer-based pre-trained natural language processing model, such as a Generative Pre-trained Transformer (GPT)-based interactive large language model (LLM).

[0016] The information processing device 10 may be, for example, a device such as a desktop PC (Personal Computer), a notebook PC, a tablet, or a general-purpose computer. The entire information processing device 10 may not be provided in a single device, but may be provided in multiple devices. For example, at least a part of the information processing device 10 may be realized in an external server or a virtual server built in a cloud server. In this case, the functions described in this embodiment are realized by cooperation between the information processing device 10 and the server.

[0017] The generation unit 101 generates a training dataset including address information and a plurality of prompts (also referred to as training data or teacher data) including the building type of the building in the address information. The training unit 102 trains the natural language model 111 so that, when address information is input to the natural language model 111, the natural language model 111 predicts and identifies the building type of the building in the address information. In this embodiment, the training unit 102 trains the natural language model 111 using a short-shot training format. Therefore, the prompts generated by the generation unit 101 are configured to include one or more fewer prompts for causing the natural language model 111 to identify each of a plurality of expected building types. Prompt training according to this embodiment is preferably performed using one prompt, preferably three or fewer prompts, preferably five or fewer prompts, preferably seven or fewer prompts, preferably ten or fewer prompts, and preferably twenty or fewer prompts. The address information acquisition unit 103 acquires the target address information (hereinafter also referred to as target address information), and the building type prediction unit 104 predicts the building type of the building in the target address information using the natural language model 111 learned by the learning unit 102. Hereinafter, the functions of the generation unit 101, the learning unit 102, the address information acquisition unit 103, and the building type prediction unit 104 will be specifically described.

[0018] The generation unit 101 generates a prompt including address information and the building type of the building in the address information. In this embodiment, the generation unit 101 generates the prompt using the address / building type database 130. The address / building type database 130 is a database that stores an address information list, which is a list of address information that actually exists, and the building types of the addresses in the address information included in the list.

[0019] FIG. 2 shows an example of the contents stored in the address / building type database 130. The address / building type database 130 stores an address information list 20, which is a list of address information that actually exists, and a building type list 21, which is a list of building types for addresses in the address information included in the list. Address information and the building type corresponding to the address information are associated with each other. That is, address information and the building type corresponding to the address information form one pair. In this embodiment, three building types are assumed: a "business building," a "unit building," and a "house" (i.e., there are three building types). A "business building" is a building primarily used as an office, including offices and company buildings. A "unit building" is a building primarily used as a residence for multiple households, including buildings, condominiums, and apartments (residences) with multiple residential sections (individually occupied spaces). A "house" is a building primarily used as a residence for a single household, including a private home and a detached house. The building type of an address in address information may change depending on the architecture of the building, etc. Therefore, the address information list 20 and the building type list 21 in the address / building type database 130 can be updated at regular intervals.

[0020] The generation unit 101 first generates (acquires) a prompt from address information and building types stored in the address / building type database 130 as shown in FIG. 2. Specifically, the generation unit 101 generates a prompt including, as input data, a pair of one piece of address information in the address information list 20 and, as correct answer data (correct answer label), a building type corresponding to the address information in the building type list 21. In addition to the pair of address information and building type, the prompt may include text corresponding to a task definition indicating a prediction of the building type from the address. In FIG. 2, an example prompt 22 corresponds to a prompt including, as input data, "abc Co., Ltd., 1-2-3, D-machi, ABC-ku, Tokyo" and, as correct answer data, "business building." The generation unit 101 stores the generated prompt in the training data storage unit 120. The generation unit 101 may generate a prompt that reflects the address information, building type, and task definition according to a predetermined format.

[0021] In this embodiment, when the prompt learning is performed in the chain-of-sorts learning format described above, the generation unit 101 may further include in the prompt, in addition to the address information, building type, and task definition, text indicating a rule corresponding to the basis for the building type corresponding to the address information. For example, the generation unit 101 may include in the prompt text indicating a rule corresponding to the basis, such as indicating that the building type corresponds to a "business building" because the address information includes "corporation." The rule may be logic for generating fictitious data, which will be described later, or logic for a predetermined rule-based process.

[0022] In this embodiment, three building types are assumed: a "business building," a "unit building," and a "house." The learning unit 102 trains the natural language model 111 so that it can identify each of these building types. Accordingly, the generation unit 101 generates at least three prompts, each having the building types to be identified, i.e., a "business building," a "unit building," and a "house," as correct answer data. Referring to FIG. 2 , as an example, the generation unit 101 generates prompts 22, 23, and 24. Prompt 22 is a prompt that includes input data such as "abc Corporation, 1-2-3, D-machi, ABC-ku, Tokyo" and correct answer data such as "business building." Prompt 23 is a prompt that includes input data such as "Residence abc303, 1-2-3, D-machi, ABC-shi, Osaka" and correct answer data such as "unit building." Prompt 24 is a prompt that includes input data such as "1234, D-machi, ABC-shi, Chiba" and correct answer data such as "house." Thus, prompt 22 has "business building" as correct answer data, prompt 23 has "unit building" as correct answer data, and prompt 24 has "house" as correct answer data.

[0023] FIG. 3 shows a training dataset 30 including prompts 22, 23, and 24 in FIG. 2 as examples of one or more prompts. In this embodiment, the training dataset 30 is configured to include prompts equal to the number of building types (=3) so that the natural language model 111 can distinguish between "business building," "unit building," and "house," i.e., so that the natural language model 111 can learn the relationship between address information and "business building," "unit building," and "house" using a short-shot learning format. However, the number of prompts included in the training dataset 30 is not limited to the number of building types. For example, the generation unit 101 may include prompts including a small number (e.g., five or less) of pairs of correct answer data and address information for each building type in the training dataset 30.

[0024] Alternatively, or in addition, the generation unit 101 may include in the training dataset a prompt including a building type pair that does not generally correspond to address information, among pairs of address information and building types included in the address / building type database 130. That is, the generation unit 101 may be configured to include a prompt including a building type that is easily mistaken for address information. Referring to FIG. 2 , the address information "FG1-2-303, DE Ward, ABC City, Fukuoka Prefecture" includes the code "303" at the end, and therefore is generally considered to be a "unit building" including a residence, but the corresponding building type is "house." Therefore, the generation unit 101 may include prompt 25, which includes "FG1-2-303, DE Ward, ABC City, Fukuoka Prefecture" as input data and "house" as correct answer data, in the training dataset 30 shown in FIG. 3 . In this case, the generation unit 101 may add prompt 25 to prompts 22, 23, and 24, or may replace prompt 24, which has "house" as correct answer data, with prompt 25. Such a pair of address information and a building type that is likely to be mistaken for the address information can be specified and input by, for example, an operator (user).

[0025] Alternatively or additionally, the generation unit 101 may generate fictitious data indicating a pair of fictitious address information as input data and a fictitious building type assigned to the fictitious address information as correct answer data, and include a prompt including the fictitious data in the training data set. The generation unit 101 may generate a prompt including the fictitious data according to a predetermined program or rule-based processing, or by an operation by an operator (user).

[0026] For example, the generation unit 101 may generate fictitious data by modifying a pair of address information and a building type included in the address / building type database 130. Specifically, the generation unit 101 may generate fictitious data including additional address information obtained by adding some information to address information (i.e., real address information) included in the address / building type database 130, and a fictitious building type for the additional address information. Alternatively or additionally, the generation unit 101 may generate fictitious data including deleted address information obtained by deleting some information from address information included in the address / building type database 130, and a fictitious building type for the deleted address information. Alternatively or additionally, the generation unit 101 may generate fictitious data including changed address information obtained by changing some information from address information included in the address / building type database 130, and a fictitious building type for the changed address information.

[0027] FIG. 4A illustrates, as an example of a prompt including fictitious data, a training dataset 40A including a prompt including fictitious data consisting of additional address information obtained by adding some information to real address information and a fictitious building type for the additional address information. The generation unit 101 generates prompt 41, which includes additional address information obtained by adding "abc Co., Ltd." to "1234 D-machi, ABC City, Chiba Prefecture," which is the address information of prompt 24 in FIG. 2, and includes a fictitious building type of "business building." The generation unit 101 also generates prompt 42, which includes additional address information obtained by adding "residence abc" to the address information of prompt 24 and includes a fictitious building type of "unit building." The generation unit 101 then generates training dataset 40A including prompts 41, 42, and 24. While training dataset 40A is configured to include prompts 41, 42, and 24, it may be configured to include other prompts described in this embodiment instead of or in addition to the prompts.

[0028] FIG. 4B illustrates another example of a prompt including fictitious data: a training dataset 40B including deleted address information in which some information has been deleted from real address information, and a prompt including fictitious data consisting of a fictitious building type for the deleted address information. The generation unit 101 generates prompt 43, which includes deleted address information obtained by deleting "Inc." from the address information of prompt 22 in FIG. 2 , "abc Corporation, 1-2-3 D-machi, ABC-ku, Tokyo," and a fictitious building type, "unit building." The generation unit 101 also generates prompt 44, which includes deleted address information obtained by deleting "abc Corporation" from the address information of prompt 22, and a fictitious building type, "house." The generation unit 101 then generates training dataset 40B including prompts 22, 43, and 44. While training dataset 40B is configured to include prompts 22, 43, and 44, it may alternatively or additionally include other prompts (training data) described in this embodiment.

[0029] FIG. 4C illustrates another example of a prompt containing fictitious data: training dataset 40C, which includes prompts containing fictitious data consisting of changed address information obtained by partially modifying real address information and a fictitious building type for the changed address information. The generation unit 101 generates prompt 45, which includes changed address information obtained by modifying "abc Co., Ltd., 1-2-3 D-machi, ABC-ku, Tokyo" (the address information of prompt 22 in FIG. 2 ) to "residence abc" and a fictitious building type "unit building." The generation unit 101 then generates training dataset 40C, which includes prompts 22, 45, and 44. While training dataset 40C is configured to include prompts 22, 45, and 44, it may alternatively or additionally include other prompts described in this embodiment.

[0030] In this embodiment, the generation unit 101 generates the prompt by using the address information and building type stored in the address / building type database 130. Alternatively or additionally, the generation unit 101 may be configured to generate the prompt based on input information received through a user operation, communication with an external device, or the like.

[0031] The learning unit 102 performs prompt training on the natural language model 111 using the training dataset (corresponding to one or more prompts) generated by the generation unit 101. For example, the learning unit 102 trains the natural language model 111 in a short-shot training format (short-shot training). Specifically, the learning unit 102 trains the natural language model 111 using the training dataset so as to execute a task of predicting and outputting the building type of a building in address information input as a prompt. The natural language model 111 trained using the training dataset is updated in the learning model storage unit 110.

[0032] The address information acquisition unit 103 acquires target address information for which a building type is to be predicted from input information to the address information acquisition unit 103. The address information acquisition unit 103 acquires the target address information from input information input by an operator (user), for example. Alternatively, the address information acquisition unit 103 may accept input information set in advance in the information processing device 10 according to a predetermined program, and acquire the target address information.

[0033] When the input information to the address information acquisition unit 103 consists of only address information, the address information acquisition unit 103 acquires the input information itself as target address information. When the input information includes address information and other information such as image information, the address information acquisition unit 103 may extract and acquire the target address information from the input information. When the input information is audio information including address information, the address information acquisition unit 103 converts the audio information into text information (i.e., transcribes it) and acquires the converted information as target address information. The conversion from audio information to text information can be performed using a known voice recognition function or the like.

[0034] The building type prediction unit 104 predicts the building type of the building in the target address information from the target address information acquired by the address information acquisition unit 103 and the natural language model 111 trained according to this embodiment. Specifically, the building type prediction unit 104 inputs the target address information into the natural language model 111 and predicts the output building type as the building type of the building in the target address information. The building type prediction unit 104 can output information on the predicted building type to the outside.

[0035] [Processing flow] Next, the processing executed by the information processing device 10 according to the present embodiment will be described with reference to Fig. 5. Fig. 5 is a flowchart of the processing executed by the information processing device 10. In S51, the generation unit 101 generates a training dataset (corresponding to one or more prompts) for training the natural language model 111. The procedure for generating the training dataset is as described above, and the generation unit 101 generates prompts, for example, by using the address / building type database 130. The training dataset is configured to include at least a prompt to which each building type to be identified for an address is assigned as correct answer data. As in this embodiment, if the building types to be identified are three types, "business building," "unit building," and "house," the training dataset includes three prompts to which these building types are assigned as correct answer data.

[0036] In S52, the training unit 102 performs prompt training of the natural language model 111 using a training dataset including one or more prompts generated in S51. In this embodiment, the training unit 102 trains the natural language model 111 using a short-shot training format. As a result, the natural language model 111 is adjusted to perform the task of predicting and outputting the building type of a building in address information input as a prompt.

[0037] In S53, the address information acquisition unit 103 acquires target address information for which a building type is to be predicted. The address information acquisition unit 103 can acquire the target address information from input information by an operator (user) or a predetermined program. In S54, the building type prediction unit 104 predicts the building type of the building in the target address information by inputting the target address information acquired in S53 into the natural language model 111. The building type prediction unit 104 may output the predicted building type information to the outside. The output may be any output process, and may be output (distribution) to an external device via communication, or may be output on a display unit.

[0038] In this way, the information processing device 10 according to the present embodiment generates a training dataset including one or more small number of prompts in order to perform prompt training of the natural language model 111. The training dataset may include prompts including pairs of address information and building types that are easily confused, or may include prompts including pairs of fictitious address information and fictitious building types. This enables the natural language model 111 to effectively perform prompt training. Then, the information processing device 10 predicts the building type for target address information using the natural language model 111 trained using the training dataset. Through such training and prediction, it becomes possible to identify the building type of a building in address information from the address information more efficiently than in the past.

[0039] Second Embodiment The information processing device according to this embodiment is configured to use a plurality of prompts generated according to the first embodiment to set prompts to be preferentially included in a training data set (hereinafter also referred to as priority prompts). In other words, the information processing device according to this embodiment is configured to use a plurality of generated prompts to set prompts (priority prompts) that the natural language model 111 should preferentially learn. The processing according to this embodiment corresponds to the processing included in the processing of S51 in FIG. 5 described in the first embodiment. Hereinafter, a description of the same configurations and features as those of the first embodiment will be omitted.

[0040] [Example of functional configuration of information processing device according to second embodiment] 6 shows an example of the functional configuration of an information processing device 60 according to this embodiment. Compared to the functional configuration of the information processing device 10 according to the first embodiment shown in FIG. 1, the information processing device 60 further includes a setting unit 105 and further includes a building type prediction model 112 stored in a learning model storage unit 110. The building type prediction model 112 is a Transformer-based natural language processing model, and is, for example, a natural language model fine-tuned to a pre-trained BERT (Bidirectional Encoder Representations from Transformers) to perform the task of outputting the building type of a building in input address information.

[0041] The setting unit 105 sets a priority prompt using the learning data storage unit 120, which stores prompts generated according to the first embodiment. Specifically, the setting unit 105 first selects one prompt from the learning data storage unit 120. For example, the setting unit 105 may select one prompt according to a predetermined rule. Then, the setting unit 105 acquires the address information and building type included in the selected prompt as reference address information and reference building type. The setting unit 105 inputs the reference address information into the building type prediction model 112 and predicts the building type of the building in the reference address information. The building type predicted using the building type prediction model 112 will be referred to as the predicted building type hereinafter.

[0042] When the setting unit 105 acquires the predicted building type, it compares it with the reference building type. If the reference building type and the predicted building type match, the setting unit 105 does not set the priority prompt for the selected prompt. On the other hand, if the reference building type and the predicted building type are different, the setting unit 105 sets a priority prompt. This is because, when the reference building type and the predicted building type are different, that is, when the building types identified by two different means are different, the reference address information can be said to be address information that makes it difficult to identify the correct building type, and a prompt including the reference address information may be a prompt that the natural language model 111 should train. In this embodiment, the setting unit 105 sets a prompt including the reference address information and the predicted building type as a priority prompt. The reason for adopting the predicted building type as correct answer data is based on the assumption that the pre-trained and fine-tuned building type prediction model 112 has high prediction accuracy and correctly predicts the building type. Next, the setting unit 105 notifies the generation unit 101 of the set priority prompt. In response, the generation unit 101 includes the priority prompt in a training dataset (corresponding to one or more prompts). Here, the generation unit 101 may add the priority prompt to the training data set, or may delete one prompt from the training data set and add the priority prompt (that is, may replace one prompt).

[0043] The process of setting a priority prompt by the setting unit 105 will be described using a specific example with reference to FIG. 7 . FIG. 7 is a conceptual diagram for explaining the process of the setting unit 105. The prompt 24 is a prompt selected by the setting unit 105 from multiple prompts stored in the training data storage unit 120. In FIG. 3 described in the first embodiment, the prompt 24 is a prompt included in the training data set 30. The prompt 24 includes "1234, D-machi, ABC-shi, Chiba Prefecture" as address information (input data) and "house" as building type (correct answer data), which correspond to reference address information 70 and reference building type 71, respectively, in this embodiment. The setting unit 105 inputs the reference address information 70 to the building type prediction model 112, predicts the building type of the building in the reference address information 70, and acquires the predicted building type as the predicted building type 72.

[0044] The setting unit 105 compares the reference building type 71 with the predicted building type 72 and determines that they are different. In this case, the setting unit 105 sets a prompt including the reference address information 70 and the predicted building type 72 as the priority prompt 73. Then, the setting unit 105 notifies the generation unit 101 of the priority prompt 73. In response to this, the generation unit 101 includes the priority prompt 73 in the training dataset. Here, the priority prompt 73 may be added to the training dataset, or one prompt may be deleted from the training dataset and the priority prompt 73 may be added. For example, in the case of the training dataset 30 of FIG. 3 described in the first embodiment, prompt 24 is included, and it is estimated that the correct answer data for prompt 24 is "business building," predicted by the building type prediction model 112. Therefore, the generation unit 101 replaces prompt 24 with priority prompt 73.

[0045] In this way, the information processing device 60 according to this embodiment sets a prompt that the natural language model 111 should preferentially learn as a priority prompt. This allows the natural language model 111 to effectively perform prompt training using the priority prompt that should be preferentially learned. Also, as in the first embodiment, the information processing device 60 predicts the building type for target address information using the natural language model 111 trained using the training dataset. Through such learning and prediction, it becomes possible to identify the building type of the building in the address information from the address information more efficiently than in the past.

[0046] Note that, in the present embodiment, an example has been described in which a priority prompt is set using the building type prediction model 112, but the setting unit 105 may set a priority prompt by other means. For example, the setting unit 105 may set a priority prompt using a prediction result by the building type prediction unit 104. Specifically, for example, when the setting unit 105 determines, based on monitoring and instructions from the user, that the building type predicted for the target address information is different from the correct building type, the setting unit 105 may set a prompt having the target address information and the correct building type as the priority prompt.

[0047] Note that prompt training according to this embodiment may be performed using only priority prompts. Furthermore, the generator 101 may delete and add prompts to increase the proportion of priority prompts in one or more prompts (corresponding to the training data set), delete and add prompts to decrease the number of prompts in one or more prompts, or delete prompts that are not set as priority prompts.

[0048] When the prompt learning according to this embodiment is performed in the chain-of-sorts learning format described above, the generation unit 101 may further include in the prompt a logic corresponding to the reason why an arbitrary prompt is set as a priority prompt.

[0049] [Hardware configuration of information processing device] 8 is a block diagram showing an example of the hardware configuration of the information processing device (information processing device 10, 60) according to the above embodiment. The information processing device can be implemented on a single or multiple computers, mobile devices, or any other processing platform. 8, an example in which the information processing device is implemented in a single computer is shown, but the information processing device according to this embodiment may be implemented in a computer system including multiple computers, which may be connected to each other via a wired or wireless network so as to be able to communicate with each other.

[0050] 8, the information processing device may include a CPU (Central Processing Unit) 81, a ROM (Read Only Memory) 82, a RAM (Random Access Memory) 83, an HDD (Hard Disk Drive) 84, an input unit 85, a display unit 86, a communication I / F 87, and a system bus 88. The information processing device may also include an external memory. The CPU 81 controls the overall operation of the information processing device, and controls each component (82 to 87) via a system bus 88, which is a data transmission path.

[0051] The ROM 82 is a non-volatile memory that stores control programs and the like necessary for the CPU 81 to execute processing. Note that the programs may be stored in a non-volatile memory such as the HDD 84 or an SSD (Solid State Drive) or an external memory such as a removable storage medium (not shown). The RAM 83 is a volatile memory and functions as the main memory, work area, etc. of the CPU 81. That is, when executing a process, the CPU 81 loads necessary programs, etc. from the ROM 82 into the RAM 83 and executes the programs, etc. to realize various functional operations. The learning model storage unit 110, the learning data storage unit 120, and the address / building type database 130 (when the address / building type database 130 is provided in the information processing device) shown in FIGS. 1 and 6 can be configured in the RAM 83.

[0052] The HDD 84 stores, for example, various data and information required when the CPU 81 performs processing using a program. The HDD 84 also stores, for example, various data and information obtained when the CPU 81 performs processing using a program. The input unit 85 is composed of a keyboard and a pointing device such as a mouse. The display unit 86 is configured by a monitor such as a liquid crystal display (LCD), etc. The display unit 86 may be configured in combination with the input unit 85 to function as a GUI (Graphical User Interface).

[0053] The communication I / F 87 is an interface that controls communication between the information processing device and an external device. The communication I / F 87 provides an interface with a network and executes communication with the external device via the network. Various data, parameters, and the like are transmitted and received between the information processing device and the external device via the communication I / F 87. In this embodiment, the communication I / F 87 may execute communication via a wired LAN (Local Area Network) or a dedicated line that conforms to a communication standard such as Ethernet (registered trademark). However, the network that can be used in this embodiment is not limited to this and may be configured as a wireless network. This wireless network includes wireless PANs (Personal Area Networks) such as Bluetooth (registered trademark), ZigBee (registered trademark), and UWB (Ultra Wide Band). It also includes wireless LANs (Local Area Networks) such as Wi-Fi (Wireless Fidelity) (registered trademark) and wireless MANs (Metropolitan Area Networks) such as WiMAX (registered trademark). It also includes wireless WANs (Wide Area Networks) such as 4G and 5G defined by 3GPP (Third Generation Partnership Project) (registered trademark). The network may be any network that connects devices to each other so that they can communicate with each other, and the communication standard, scale, and configuration are not limited to those described above.

[0054] At least some of the functions of the information processing device can be realized by the CPU 81 in the information processing device executing a program. However, at least some of the functions of the information processing devices 10 and 60 shown in Figures 1 and 6 may be operated as dedicated hardware. In this case, the dedicated hardware may operate under the control of the CPU 81.

[0055] Although certain embodiments have been described above, these embodiments are merely examples and are not intended to limit the scope of the present invention. The devices and methods described herein may be embodied in forms other than those described above. Furthermore, appropriate omissions, substitutions, and modifications may be made to the above-described embodiments without departing from the scope of the present invention. Such omissions, substitutions, and modifications are included within the scope of the claims and their equivalents, and belong to the technical scope of the present invention.

[0056] The disclosure of this embodiment includes the following configuration. [1] An information processing device having: a generation unit that generates a learning dataset including prompts including address information as input data and the building type of the building in the address information as correct answer data for the input data; and a learning unit that performs prompt learning of a pre-trained natural language model using one or more of the prompts so as to execute a task of outputting the building type of the building in the input target address information.

[0057] [2] The information processing device described in [1], wherein the building type is one of a predetermined plurality of building types, and the generation unit includes each of the plurality of building types in the prompt as at least correct answer data.

[0058] [3] The information processing device described in [1] or [2], wherein the generation unit includes in one or more of the prompts a prompt including fictitious address information generated from real address information and a building type of the fictitious building in the fictitious address information.

[0059] [4] The information processing device according to [3], wherein the fictitious address information is deleted address information obtained by deleting part of the real address information.

[0060] [5] The information processing device according to [3] or [4], wherein the fictitious address information is additional address information in which some information is added to the real address information.

[0061] [6] The information processing device according to any one of [3] to [5], wherein the fictitious address information is modified address information obtained by modifying part of the real address information.

[0062] [7] The information processing device according to any one of [1] to [6], further comprising a setting unit that sets a prompt that the natural language model should preferentially learn as a priority prompt.

[0063] [8] An information processing device as described in [8], further comprising: an acquisition unit that selects one prompt from one or more of the prompts and acquires the address information and building type included in the selected prompt as reference address information and reference building type, respectively; and a prediction unit that predicts the building type of the building in the reference address information as a predicted building type by inputting the reference address information into a building type prediction model trained to predict building type from address information, wherein the setting unit sets a prompt including the reference address information and the predicted building type as the priority prompt when the reference building type and the predicted building type are different. [Explanation of symbols]

[0064] 10;60: Information processing device, 101: Generation unit, 102: Learning unit, 103: Address information acquisition unit, 104: Building type prediction unit, 105: Setting unit, 110: Learning model storage unit, 111: Natural language model, 112: Building type prediction model, 120: Learning data storage unit, 130: Address / building type database

Claims

1. a generation unit that generates a prompt including address information as input data and a building type of the building in the address information as correct answer data for the input data; a training unit that performs prompt training of a pre-trained natural language model using one or more of the prompts to execute a task of outputting a building type of a building in input target address information; An information processing device having the above.

2. The building type is one of a plurality of predetermined building types, The information processing device according to claim 1 , wherein the generating unit includes each of the plurality of building types as at least correct answer data in the prompt.

3. The information processing device according to claim 1 , wherein the generating unit includes in one or more of the prompts a prompt including fictitious address information generated from real address information and a building type of a fictitious building in the fictitious address information.

4. The information processing device according to claim 3 , wherein the fictitious address information is deleted address information obtained by deleting part of the real address information.

5. The information processing device according to claim 3 , wherein the fictitious address information is additional address information obtained by adding some information to the real address information.

6. The information processing device according to claim 5 , wherein the fictitious address information is modified address information obtained by modifying part of the real address information.

7. The information processing apparatus according to claim 1 , further comprising a setting unit that sets a prompt that the natural language model should preferentially learn as a priority prompt.

8. an acquisition unit that selects one prompt from the one or more prompts and acquires address information and a building type included in the selected prompt as reference address information and a reference building type, respectively; a prediction unit that predicts the building type of the building in the reference address information as a predicted building type by inputting the reference address information into a building type prediction model that has been trained to predict a building type from address information, The information processing device according to claim 7 , wherein the setting unit sets a prompt including the reference address information and the predicted building type as the priority prompt when the reference building type and the predicted building type are different.

9. An information processing method executed by an information processing device, generating, by a generating unit, a prompt including address information as input data and a building type of the building in the address information as correct answer data for the input data; performing prompt training of the pre-trained natural language model using one or more of the prompts by a training unit to execute a task of outputting a building type of a building in input target address information; An information processing method, including:

10. An information processing program for causing a computer to execute information processing, the program including: a generation process for generating a prompt including address information as input data and a building type of the building in the address information as correct answer data for the input data; a training process for performing prompt training of a pre-trained natural language model using one or more of the prompts so as to execute a task of outputting a building type of a building in input target address information, Information processing program.

Images