Learning device, inference device, policy planning assistance system, program, learning method, and inference method
The learning device and system address the challenge of predicting policy effectiveness by converting data into vectors and training a model to infer policy impact, enhancing the ability to assess measure performance beyond conventional recommendation methods.
Patent Information
- Application Number
- PCT/JP2023/018831
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2023-05-19
- Publication Date
- 2024-11-28
AI Technical Summary
Conventional policy planning devices cannot predict the effectiveness of measures, only determining the degree of recommendation based on past selections without forecasting future performance.
A learning device and system that converts reference data into learning and inference policy vectors, using these vectors to train a learning model that infers the effectiveness of policies by extracting and generating teacher and inference data, allowing for the prediction of measure effects.
Enables the prediction of measure effects by generating a learning model that can infer the impact of policies based on historical data, providing a more accurate assessment of policy effectiveness.
Smart Images

Figure JP2023018831_28112024_PF_FP_ABST
Abstract
Description
Learning device, inference device, policy planning support system, program, learning method, and inference method
[0001] The present disclosure relates to a learning device, an inference device, a policy planning support system, a program, a learning method, and an inference method.
[0002] Conventionally, devices have been provided to assist users in devising measures. For example, Patent Literature 1 describes a measure planning device that can access measure result information that stores measure results and event information that stores events. For each measure, the measure result is defined by explanatory variables related to the measure and objective variables related to the measure. An event is defined by the event period and attribute information of the measures implemented during the event period.
[0003] The policy planning device then executes a selection process to select a specific policy result corresponding to the event from the policy result information, a setting process to calculate an influence indicating the possibility of use in policy planning due to a change in situation based on the number of times the specific policy result has been selected and set it as the specific policy result, and a calculation process to calculate a recommendation level for the specific policy that resulted in the specific policy result based on the influence and the objective variable of the specific policy result and output the calculation result.
[0004] Japanese Patent Application Laid-Open No. 2021-117837
[0005] Conventional techniques determine the degree of recommendation of a policy based on the results of actually selecting a policy for a target event, but they cannot predict how effective that policy will be.
[0006] Therefore, one or more aspects of the present disclosure aim to make it possible to predict the effectiveness of measures.
[0007] A learning device according to one aspect of the present disclosure is characterized by comprising: an extraction unit that extracts learning measures, which are measures for use in learning, learning effects, which are the effects of the learning measures, and multiple learning elements, which are multiple elements related to the learning measures, from reference data, which is data to be learned, and generates a learning measure vector by converting the learning measures into a vector; a teacher data generation unit that generates teacher data using the learning measure vector and the multiple learning elements as learning data and the learning effects as correct answer data; and a learning unit that learns a learning model using the teacher data.
[0008] An inference device according to one aspect of the present disclosure is characterized in that it comprises an inference data acquisition unit that acquires inference data indicating an inference measure, which is a measure to be inferred, and a plurality of inference elements, which are a plurality of elements related to the inference measure; and an inference unit that generates an inference measure vector by converting the inference measure into a vector, and performs the inference of the effect of the inference measure by inputting input data including the inference measure vector and the plurality of inference elements into a learning model that has been trained using teacher data in which the training measure vector and the plurality of learning elements are used as training data and in which the training effect, which is the effect of the training measure, which is the measure converted into the training measure vector, is used as correct answer data.
[0009] A policy planning support system according to one aspect of the present disclosure comprises an extraction unit that extracts, from reference data that is data to be learned, learning measures that are measures for use in learning, learning effects that are the effects of the learning measures, and a plurality of learning elements that are a plurality of elements related to the learning measures, and generates a learning policy vector by converting the learning measures into a vector; a teacher data generation unit that generates teacher data that uses the learning policy vector and the plurality of learning elements as learning data and the learning effect as correct answer data; a learning unit that trains a learning model using the teacher data; an inference data acquisition unit that acquires inference data indicating an inference policy that is a policy to be inferred and a plurality of inference elements that are a plurality of elements related to the inference policy; and an inference unit that generates an inference policy vector by converting the inference policy into a vector, and infers the effect of the inference policy by inputting input data including the inference policy vector and the plurality of inference elements into the learning model.
[0010] A program according to a first aspect of the present disclosure causes a computer to function as an extraction unit that extracts learning measures, which are measures for use in learning, learning effects, which are the effects of the learning measures, and multiple learning elements, which are multiple elements related to the learning measures, from reference data, which is data to be learned, and generates a learning measure vector by converting the learning measures into a vector; a teacher data generation unit that generates teacher data using the learning measure vector and the multiple learning elements as learning data and the learning effect as correct answer data; and a learning unit that learns a learning model using the teacher data.
[0011] A program relating to a second aspect of the present disclosure is characterized in that it causes a computer to function as an inference data acquisition unit that acquires inference data indicating an inference measure, which is a measure to be inferred, and a plurality of inference elements, which are a plurality of elements related to the inference measure, and an inference unit that generates an inference measure vector by converting the inference measure into a vector, and infers the effect of the inference measure by inputting input data including the inference measure vector and the plurality of inference elements into a learning model that has been trained using teacher data in which the learning measure vector and the plurality of learning elements are used as learning data and the learning effect, which is the effect of the learning measure, which is the measure converted into the learning measure vector, is used as correct answer data.
[0012] A learning method according to one aspect of the present disclosure is characterized in that it extracts learning measures, which are measures for use in learning, learning effects, which are the effects of the learning measures, and multiple learning elements, which are multiple elements related to the learning measures, from reference data, which is the data to be learned; generates a learning measure vector by converting the learning measures into a vector; generates teacher data in which the learning measure vector and the multiple learning elements are used as learning data and the learning effects are correct answer data; and trains a learning model using the teacher data.
[0013] An inference method according to one aspect of the present disclosure is characterized in that it acquires inference data indicating an inference measure, which is a measure to be inferred, and a plurality of inference elements, which are a plurality of elements related to the inference measure, generates an inference measure vector by converting the inference measure into a vector, and infers the effect of the inference measure by inputting input data including the inference measure vector and the plurality of inference elements into a learning model trained using teacher data in which a training measure vector and the plurality of learning elements are used as training data and a training effect, which is the effect of the training measure, which is the measure converted into the training measure vector, is used as correct answer data.
[0014] According to one or more aspects of the present disclosure, the effectiveness of a policy can be predicted.
[0015] FIG. 1 is a block diagram schematically showing the configuration of a policy planning support device according to embodiments 1 and 2. FIG. 2 is a block diagram schematically showing the configuration of an inference unit according to embodiment 1. FIG. 3 is a block diagram schematically showing the configuration of a PC. FIG. 4 is a flowchart showing a learning operation in the policy planning support device according to embodiment 1. FIG. 5 is a flowchart showing the inference operation in the policy planning support device according to embodiment 1. FIG. 6 is a block diagram schematically showing the configuration of an inference unit according to embodiment 2. FIG. 7 is a flowchart showing the inference operation in the policy planning support device according to embodiment 2. FIG. 8 is a block diagram schematically showing the configuration of a policy planning support device according to embodiment 3. FIG. 9 is a block diagram schematically showing the configuration of an inference unit according to embodiment 3. FIG. 10 is a flowchart showing the learning operation in the policy planning support device according to embodiment 3. FIG. 11 is a flowchart showing the inference operation in the policy planning support device according to embodiment 3.
[0016] 1 is a block diagram showing a schematic configuration of a policy planning support device 100 according to embodiment 1. The policy planning support device 100 includes a reference data acquisition unit 101, a policy effect extraction unit 102, a teacher data generation unit 103, a learning unit 104, a learning model storage unit 105, an inference data acquisition unit 106, an inference unit 107, and an inference result output unit 113.
[0017] The reference data acquisition unit 101 acquires reference data, which is data to be learned. The acquired reference data is provided to the policy effect extraction unit 102. The reference data is assumed to be data that includes at least a document. For example, the reference data acquisition unit 101 may acquire the reference data from another device via a communication unit (not shown) provided in the policy planning support device 100 or via a network such as the Internet. Alternatively, the reference data acquisition unit 101 may acquire the reference data from a storage unit (not shown) provided in the policy planning support device 100.
[0018] The policy effect extraction unit 102 is an extraction unit that extracts policies, the effects of the policies, and multiple elements related to the policies from the reference data. The multiple elements to be extracted may be predetermined, such as the implementation date and time of the policies, the targets of the policies, etc. The policies, effects, and elements extracted here are also referred to as learning policies, learning effects, and learning elements, respectively.
[0019] For example, the policy effect extraction unit 102 may use a known technique such as argument mining to extract policies, their effects, and multiple elements related to the policies from the documents in the reference data. Specifically, the policy effect extraction unit 102 may use a known natural language processing technique to extract policies and their effects by detecting keywords corresponding to the policies and words related to the keywords that indicate effects from the documents in the reference data.
[0020] The policy effect extraction unit 102 generates a policy vector by converting the extracted policies into a vector. The converted policy vector is also called a training policy vector. For example, the policy effect extraction unit 102 may convert from language to a vector using a known technique such as embedding. Then, the policy effect extraction unit 102 provides the policy vector, the effect, and multiple elements to the training data generation unit 103.
[0021] The teacher data generation unit 103 generates teacher data from the policy vector, the effect, and multiple elements. The generated teacher data is provided to the learning unit 104. For example, the teacher data generation unit 103 generates teacher data in which the policy vector and multiple elements are used as learning data and the effect is used as correct answer data.
[0022] The learning unit 104 uses the training data to learn a learning model for inferring the effectiveness of a policy from the policy vector and multiple elements. The learning model is stored in the learning model storage unit 105.
[0023] The learning model storage unit 105 stores the learning model learned by the learning unit 104 .
[0024] The inference data acquisition unit 106 acquires inference data, which is data for inferring effects. The inference data is provided to the inference unit 107. The inference data indicates a policy to be inferred and multiple elements related to the policy. The policy and elements indicated by the inference data are also referred to as an inference policy and an inference element, respectively. For example, the inference data acquisition unit 106 may acquire the inference data by receiving input of the inference data via an input unit (not shown) provided in the policy planning support device 100. The inference data acquisition unit 106 may also acquire the inference data from another device via a communication unit (not shown) provided in the policy planning support device 100.
[0025] The inference unit 107 infers the effect of a policy from the policy and multiple elements indicated by the inference data. Fig. 2 is a block diagram showing a schematic configuration of the inference unit 107 in embodiment 1. The inference unit 107 includes a policy effect inference unit 108 and an output data generation unit 109.
[0026] The policy effect inference unit 108 generates a policy vector by converting the policies indicated in the inference data into a vector. The method for generating the policy vector here may be the same as the method used by the policy effect extraction unit 102. The policy vector converted here is also referred to as an inference policy vector.
[0027] The policy effect inference unit 108 generates input data including the converted policy vector and a plurality of elements represented by the inference data, and infers the effect of the policy by inputting the input data to the learning model stored in the learning model storage unit 105. The inferred effect is provided to the output data generation unit 109.
[0028] The output data generating unit 109 generates output data indicating the inferred effect. The output data is provided to the inference result output unit 113.
[0029] The inference result output unit 113 outputs the effect inferred by the inference unit 107. For example, the inference result output unit 113 displays the effect indicated by the output data. The inference result output unit 113 may also transmit the output data to another device.
[0030] The policy planning support device 100 described above can be realized by, for example, a computer such as a PC (Personal Computer) 140 shown in Fig. 3. The PC 140 includes an auxiliary storage device 141 such as an HDD (Hard Disk Drive) and an SSD (Solid State Drive), a memory 142, a processor 143 such as a CPU (Central Processing Unit), a communication I / F (Interface) 144 such as a NIC (Network Interface Card), an input I / F 145 such as a keyboard and a mouse, and a display 146.
[0031] For example, the learning model storage unit 105 can be realized by the auxiliary storage device 141 or the memory 142. The reference data acquisition unit 101, the policy effect extraction unit 102, the teacher data generation unit 103, the learning unit 104, the inference data acquisition unit 106, and the inference unit 107 can be realized by loading a program stored in the auxiliary storage device 141 into the memory 142 and having the processor 143 execute the program.
[0032] The inference result output unit 113 can be realized by the communication I / F 144 or the display 146 .
[0033] The program may be downloaded to the auxiliary storage device 141 from a recording medium (not shown) via a reader / writer (not shown) or from a network via the communication I / F 144, and then loaded onto the memory 142 and executed by the processor 143. Alternatively, the program may be directly loaded onto the memory 142 from a recording medium via the reader / writer or from a network via the communication I / F 144, and then executed by the processor 143. In other words, the program may be provided by a program product such as a recording medium.
[0034] 4 is a flowchart showing the learning operation in the policy planning support device 100 according to the first embodiment. First, the reference data acquisition unit 101 acquires reference data (S10). The acquired reference data is provided to the policy effect extraction unit 102.
[0035] The measure effect extraction unit 102 extracts measures, the effects of the measures, and multiple elements related to the measures from the reference data (S11). For example, when learning about measures to reduce overtime hours of workers, the effect of the measures will be time-series data on overtime hours before and after the measures, and the multiple elements will be the day of the week, month of the measures, period of the measures, number of holidays taken by workers, number of development projects, target departments, etc.
[0036] The measure effect extraction unit 102 converts the extracted measures into vectors to generate a measure vector (S12). The measure vector, the effect, and a plurality of elements are provided to the training data generation unit 103.
[0037] The training data generating unit 103 generates training data from the policy vector, the effect, and a plurality of elements (S13). The generated training data is provided to the learning unit 104.
[0038] The learning unit 104 uses the training data to learn a learning model that infers the effect of a policy from the policy vector and multiple elements (S14). The learning model is stored in the learning model storage unit 105.
[0039] 5 is a flowchart showing the inference operation of the policy planning support device 100 in the first embodiment. First, the inference data acquisition unit 106 acquires inference data, which is data for inferring effects (S20). The inference data is provided to the inference unit 107. Here, when inferring the effects of a measure to reduce workers' overtime hours, the inference data includes the measure to reduce workers' overtime hours and multiple elements such as the day of the week on which the measure is implemented, the month in which it is implemented, the number of holidays taken by workers, the number of development projects, and the target departments.
[0040] The measure effect inference unit 108 generates a measure vector by converting the measure indicated by the inference data into a vector (S21).
[0041] The policy effect inference unit 108 generates input data including the converted policy vector and a plurality of elements represented by the inference data, and infers the effect of the policy by inputting the input data to the learning model stored in the learning model storage unit 105 (S22). The inferred effect is provided to the output data generation unit 109.
[0042] The output data generation unit 109 generates output data indicating the inferred effect (S23). The output data is provided to the inference result output unit 113, which outputs the inferred result.
[0043] As described above, according to embodiment 1, a learning model is generated by learning the effectiveness of a measure from reference data, and the effectiveness of the measure can be inferred using the learning model.
[0044] Embodiment 2 As shown in FIG. 1 , a policy planning support device 200 according to embodiment 2 includes a reference data acquisition unit 101, a policy effect extraction unit 102, a teacher data generation unit 103, a learning unit 104, a learning model storage unit 105, an inference data acquisition unit 106, an inference unit 207, and an inference result output unit 113.
[0045] The reference data acquisition unit 101, the policy effect extraction unit 102, the teacher data generation unit 103, the learning unit 104, the learning model storage unit 105, the inference data acquisition unit 106, and the inference result output unit 113 of the policy planning support device 200 according to embodiment 2 are similar to the reference data acquisition unit 101, the policy effect extraction unit 102, the teacher data generation unit 103, the learning unit 104, the learning model storage unit 105, the inference data acquisition unit 106, and the inference result output unit 113 of the policy planning support device 100 according to embodiment 1.
[0046] The inference unit 207 infers the effect of a measure from the measure and multiple elements indicated in the inference data, similar to the inference unit 107 in the first embodiment. Furthermore, in the second embodiment, the inference unit 207 identifies the basis for the inference. For example, the inference unit 207 identifies, as the basis for the inference, at least the inference element that has the greatest influence on the inference among the multiple inference elements.
[0047] 6 is a block diagram showing a schematic configuration of the inference unit 207 according to the second embodiment. The inference unit 207 includes a policy effect inference unit 108, an output data generation unit 209, and an inference basis specification unit 210.
[0048] The measure effect inference unit 108 of the inference unit 207 in the second embodiment is similar to the measure effect inference unit 108 of the inference unit 107 in the first embodiment.
[0049] The inference basis identification unit 210 identifies the basis for the inference made by the policy effect inference unit 108. For example, the inference basis identification unit 210 uses a publicly known AI (Artificial Intelligence) explanation technique such as SHAP to identify factors that have a significant impact on the inference (here, the effect). SHAP calculates a value indicating how each feature variable affects the prediction. This allows the inference basis identification unit 210 to identify multiple elements that have a significant impact on the inferred effect as the basis for the inference.
[0050] Then, the inference basis specification unit 210 generates inference basis explanation information in which the plurality of elements are arranged in descending order of their influence on the inference, and provides the inference basis explanation information to the output data generation unit 209 .
[0051] The output data generation unit 209 generates output data indicating the inferred effect and the elements indicated in the inference basis explanation information. The output data is provided to the inference result output unit 113. Note that the output data may indicate all elements indicated in the inference basis explanation information, or may indicate a predetermined number of elements in descending order of their influence on the inference.
[0052] Fig. 7 is a flowchart showing the inference operation in the policy planning support device 200 according to the second embodiment. In the flowchart shown in Fig. 7, the same processes as those in the flowchart shown in Fig. 5 are denoted by the same reference numerals as those in Fig. 5.
[0053] The processing in steps S20 to S22 in the flowchart shown in Fig. 7 is the same as the processing in steps S20 to S22 in the flowchart shown in Fig. 5. However, in Fig. 7, after the processing in step S22, the processing proceeds to step S33.
[0054] In step S33, the inference basis specification unit 210 specifies the basis of the inference made by the policy effect inference unit 108. The basis here is an element that has a large influence on the inference. The specified basis is provided to the output data generation unit 209.
[0055] The output data generation unit 209 generates output data indicating the inferred effect and the identified basis (S34). The output data is provided to the inference result output unit 113, which outputs the inference result.
[0056] As described above, according to the second embodiment, a learning model is generated by learning the effects of measures from reference data, and the learning model can be used to infer the effects of measures and provide the basis for that inference.
[0057] 8 is a block diagram showing a schematic configuration of a policy planning support device 300 according to embodiment 3. The policy planning support device 300 includes a reference data acquisition unit 101, a teacher data generation unit 303, a learning unit 304, a learning model storage unit 105, an inference data acquisition unit 106, an inference unit 307, an inference result output unit 113, and a policy effect / demerit extraction unit 320.
[0058] The reference data acquisition unit 101, learning model storage unit 105, inference data acquisition unit 106, and inference result output unit 113 of the policy planning support device 300 according to embodiment 3 are similar to the reference data acquisition unit 101, learning model storage unit 105, inference data acquisition unit 106, and inference result output unit 113 of the policy planning support device 100 according to embodiment 1.
[0059] The policy effect / disadvantage extraction unit 320 is an extraction unit that extracts policies, their effects, their disadvantages, and multiple elements related to the policies from the reference data. The disadvantages extracted here are also called learning disadvantages. In the third embodiment, for example, the policy effect / disadvantage extraction unit 320 may use a known technique such as argument mining to extract policies, their effects, their disadvantages, and multiple elements related to the policies.
[0060] The policy effect / demerit extraction unit 320 generates a policy vector by converting the extracted policies into vectors, as in the first embodiment. The policy effect / demerit extraction unit 320 then provides the policy vector, effects, demerits, and multiple elements to the teacher data generation unit 303.
[0061] The teacher data generation unit 303 generates teacher data from the policy vector, the effect, the demerit, and multiple elements. The generated teacher data is provided to the learning unit 304. For example, the teacher data generation unit 303 generates teacher data in which the policy vector and multiple elements are used as learning data and the effect and demerit are used as correct answer data. In other words, the teacher data generation unit 303 in the third embodiment also includes the learning demerit in the correct answer data.
[0062] The learning unit 304 uses the training data to learn a learning model that infers the effects and disadvantages of a policy from the policy vector and multiple elements. Therefore, the learning model in the third embodiment is a model trained using data that also indicates the learning disadvantages, which are the disadvantages of the learning policy, as correct answer data. The learning model is stored in the learning model storage unit 105.
[0063] The inference unit 307 infers the effects and disadvantages of a policy from the policy and multiple elements indicated in the inference data, and identifies the basis for the inference. In the third embodiment, the inference unit 307 also infers the disadvantages of the policy by receiving output of the disadvantages of the policy to be inferred from the learning model. Figure 9 is a block diagram showing a schematic configuration of the inference unit 307 in the third embodiment. The inference unit 307 includes an output data generation unit 309, a policy effect / disadvantage inference unit 311, and an inference basis explanation unit 312.
[0064] The measure effect / demerit inference unit 311 generates a measure vector by converting the measures indicated by the inference data into a vector. The method for generating the measure vector here may be the same as the method used by the measure effect / demerit extraction unit 320.
[0065] The policy effect / demerit inference unit 311 generates input data including the converted policy vector and a plurality of elements indicated by the inference data, and infers the effects and demerits of the policy by inputting the input data into the learning model stored in the learning model storage unit 105. The inferred effects and demerits are provided to the output data generation unit 309.
[0066] The inference basis explanation unit 312 identifies, as the basis for the inference, at least the inference element that has the greatest influence on the inference among the multiple inference elements. Then, the inference basis explanation unit 312 generates an explanation sentence that explains the basis for the inference in natural language using the identified inference element.
[0067] For example, the inference basis explanation unit 312 identifies the basis of the inference made by the policy effect / demerit inference unit 311. Here, the inference basis explanation unit 312 identifies the basis by a process similar to the process performed by the inference basis identification unit 210 in the second embodiment.
[0068] In the third embodiment, the inference basis explanation unit 312 generates an inference basis explanation that explains a predetermined number of identified grounds that have a large influence on the inference. For example, the inference basis explanation unit 312 may generate an inference basis explanation that explains the grounds of the inference in natural language using a known technology such as a language generation model such as GPT-3 (Generative Pre-trained Transformer 3). The generated inference basis explanation is provided to the output data generation unit 309.
[0069] The output data generating unit 309 generates output data indicating the inferred advantages and disadvantages and an explanation of the inference basis. The output data is provided to the inference result output unit 113.
[0070] The above-described policy planning support device 300 can also be realized by, for example, a PC 140 as shown in Fig. 3. The policy effect / demerit extraction unit 320 can also be realized by loading a program stored in an auxiliary storage device 141 into a memory 142 and having a processor 143 execute the program.
[0071] Fig. 10 is a flowchart showing the learning operation in the policy planning support device 300 according to the third embodiment. In the flowchart shown in Fig. 10, the same processes as those in the flowchart shown in Fig. 4 are denoted by the same reference numerals as those in Fig. 4.
[0072] The process of step S10 in the flowchart shown in Fig. 10 is the same as the process of step S10 in the flowchart shown in Fig. 4. However, in Fig. 10, after the process of step S10, the process proceeds to step S41.
[0073] In step S41, the measure effect / demerit extraction unit 320 extracts measures, their effects, their demerits, and multiple elements related to the measures from the reference data (S41). Then, the process proceeds to step S12.
[0074] The process of step S12 in the flowchart shown in Fig. 10 is the same as the process of step S12 in the flowchart shown in Fig. 4. However, in Fig. 10, after the process of step S12, the process proceeds to step S43.
[0075] In step S43, the training data generating unit 303 generates training data from the policy vector, the effect, the demerit, and a plurality of elements. The generated training data is provided to the learning unit 304.
[0076] The learning unit 304 uses the training data to learn a learning model that infers the effects and disadvantages of a policy from the policy vector and multiple elements (S44). The learning model is stored in the learning model storage unit 105.
[0077] Fig. 11 is a flowchart showing the inference operation in the policy planning support device 300 according to the third embodiment. In the flowchart shown in Fig. 11, the same processes as those in the flowchart shown in Fig. 5 are denoted by the same reference numerals as those in Fig. 5.
[0078] The processing in steps S20 and S21 in the flowchart shown in Fig. 11 is the same as the processing in steps S20 and S21 in the flowchart shown in Fig. 5. However, in Fig. 11, after the processing in step S21, the processing proceeds to step S52. Also, in Fig. 11, the processing in step S21 is performed by the policy effect / demerit inference unit 311.
[0079] In step S52, the policy effect / demerit inference unit 311 generates input data including the converted policy vector and a plurality of elements represented by the inference data, and infers the effects and demerits of the policy by inputting the input data into the learning model stored in the learning model storage unit 105. The inferred effects and demerits are provided to the output data generation unit 309.
[0080] The inference basis explanation unit 312 identifies the basis of the inference made by the policy effectiveness / demerit inference unit 311, and generates an inference basis explanation sentence that explains the identified basis (S53). The generated inference basis explanation sentence is provided to the output data generation unit 309.
[0081] The output data generation unit 309 generates output data indicating the inferred effect and the inference basis explanation (S54). The output data is provided to the inference result output unit 113, which outputs the inference result.
[0082] As described above, according to the third embodiment, a learning model is generated by learning the effects and disadvantages of a policy from reference data, and the effects and disadvantages of the policy can be inferred using the learning model. Furthermore, in the third embodiment, the basis for the inference is explained in natural language, making it easier to recognize the basis.
[0083] In the above-described first to third embodiments, an example is shown in which the policy planning support devices 100 to 300 are realized by a single PC 140, but the first to third embodiments are not limited to such an example. For example, the processes performed by the policy planning support devices 100 to 300 may be distributed and performed by multiple PCs connected to a network such as the Internet. In other words, the first to third embodiments may be realized as a policy planning support system that performs the same functions as the policy planning support devices 100 to 300. For example, a learning device that learns a learning model, a learning model storage device that stores the learning model, and an inference device that performs inference using the learning model may be connected via a network. In this case, the learning device may include a reference data acquisition unit 101, a policy effect extraction unit 102 or a policy effect disadvantage extraction unit 320, a teacher data generation unit 103, 303, a learning unit 104, 304, and a communication unit realized by a communication I / F (Interface) not shown, the learning model storage device may include a learning model storage unit 105 and a communication unit, and the inference device may include an inference data acquisition unit 106, an inference unit 107, 207, 307, an inference result output unit 113, and a communication unit.
[0084] In addition, in embodiment 2, by providing an inference basis explanation unit 312 shown in Figure 9 instead of the inference basis identification unit 210 shown in Figure 6, an explanatory text explaining the basis of the effect in natural language may be generated.
[0085] 100, 200, 300 Policy planning support device, 101 Reference data acquisition unit, 102 Policy effect extraction unit, 103, 303 Teacher data generation unit, 104, 304 Learning unit, 105 Learning model storage unit, 106 Inference data acquisition unit, 107, 207, 307 Inference unit, 108 Policy effect inference unit, 109, 209, 309 Output data generation unit, 210 Inference basis identification unit, 311 Policy effect disadvantage inference unit, 312 Inference basis explanation unit, 113 Inference result output unit, 320 Policy effect disadvantage extraction unit.
Claims
1. A learning device comprising: an extraction unit that extracts learning measures, which are learning measures, learning effects, which are the effects of the learning measures, and a plurality of learning elements, which are a plurality of elements related to the learning measures, from reference data, which is data to be learned, and generates a learning measure vector by converting the learning measures into a vector; a teacher data generation unit that generates teacher data using the learning measure vector and the plurality of learning elements as learning data and the learning effects as correct answer data; and a learning unit that learns a learning model using the teacher data.
2. The learning device described in claim 1, characterized in that the extraction unit also extracts learning demerits, which are demerits of the learning measures, from the reference data, and the teacher data generation unit also includes the learning demerits in the correct answer data.
3. An inference device comprising: an inference data acquisition unit that acquires inference data indicating an inference measure, which is a measure to be inferred, and a plurality of inference elements, which are a plurality of elements related to the inference measure; and an inference unit that generates an inference measure vector by converting the inference measure into a vector, and inputs input data including the inference measure vector and the plurality of inference elements into a learning model that is trained using teacher data in which a learning measure vector and the plurality of learning elements are used as learning data, and a learning effect, which is the effect of the learning measure, which is the measure converted into the learning measure vector, is used as correct answer data, thereby performing the inference of the effect of the inference measure.
4. The inference device described in claim 3, characterized in that the learning model is a model trained using data that also indicates learning disadvantages, which are disadvantages of the learning measures, as the correct answer data, and the inference unit also receives output of the disadvantages of the inference measures from the learning model.
5. An inference device as claimed in claim 3 or 4, characterized in that the inference unit at least identifies, as the basis for the inference, the inference element having the greatest influence on the inference among the plurality of inference elements.
6. The inference device according to claim 5, wherein the inference unit generates an explanatory sentence in natural language that explains the grounds using the identified inference elements.
7. A policy planning support system comprising: an extraction unit that extracts learning measures, which are learning measures, learning effects, which are the effects of the learning measures, and a plurality of learning elements, which are a plurality of elements related to the learning measures, from reference data, which is data to be learned, and generates a learning policy vector by converting the learning measures into a vector; a teacher data generation unit that generates teacher data using the learning policy vector and the plurality of learning elements as learning data and the learning effect as correct answer data; a learning unit that learns a learning model using the teacher data; an inference data acquisition unit that acquires inference data indicating an inference policy, which is a policy to be inferred, and a plurality of inference elements, which are a plurality of elements related to the inference policy; and an inference unit that generates an inference policy vector by converting the inference policy into a vector, and infers the effect of the inference policy by inputting input data including the inference policy vector and the plurality of inference elements into the learning model.
8. A program causing a computer to function as: an extraction unit that extracts learning measures, which are learning measures, learning effects, which are the effects of the learning measures, and a plurality of learning elements, which are a plurality of elements related to the learning measures, from reference data, which is data to be learned, and generates a learning measure vector by converting the learning measures into a vector; a teacher data generation unit that generates teacher data using the learning measure vector and the plurality of learning elements as learning data and the learning effects as correct answer data; and a learning unit that learns a learning model using the teacher data.
9. A program for causing a computer to function as: an inference data acquisition unit that acquires inference data indicating an inference measure, which is a measure to be inferred, and a plurality of inference elements, which are a plurality of elements related to the inference measure; and an inference unit that generates an inference measure vector by converting the inference measure into a vector, and inputs input data including the inference measure vector and the plurality of inference elements into a learning model that has been trained using teacher data in which a learning measure vector and the plurality of learning elements are used as learning data, and a learning effect, which is the effect of the learning measure, which is the measure converted into the learning measure vector, is used as correct answer data, thereby inferring the effect of the inference measure.
10. A learning method comprising: extracting, from reference data, which is the data to be learned, a learning measure, which is a learning measure, a learning effect, which is the effect of the learning measure, and a plurality of learning elements, which are a plurality of elements related to the learning measure; generating a learning measure vector by converting the learning measure into a vector; generating teacher data using the learning measure vector and the plurality of learning elements as learning data and the learning effect as correct answer data; and learning a learning model using the teacher data.
11. An inference method comprising: acquiring inference data indicating an inference measure, which is a measure to be inferred, and a plurality of inference elements, which are a plurality of elements related to the inference measure; generating an inference measure vector by converting the inference measure into a vector; and inputting input data including the inference measure vector and the plurality of inference elements into a learning model trained using teacher data in which a learning measure vector and the plurality of learning elements are used as learning data, and a learning effect, which is the effect of the learning measure, which is the measure converted into the learning measure vector, is used as correct answer data, thereby inferring the effect of the inference measure.
Citation Information
Patent Citations
Computer system
JP2021067971A
Computer system and intervention effect prediction method
JP2023004225A
Cited By
system
JP2026155198A
