Learning device, prediction device, method, and program

Abstract

A learning device according to one aspect comprises: an input unit that inputs input data including dialog data of a dialog aimed to determine an action plan of a subject, and a correct answer for an action continuation degree indicating a degree to which the subject continues to implement an action according to the action plan; a prediction unit that predicts a prediction value of the action continuation degree, on the basis of a feature amount of the input data and a prediction model that uses the feature amount as an input, and outputs the prediction value of the action continuation degree; and an update unit that updates a learnable parameter of the prediction model, on the basis of the correct answer for the action continuation degree and the prediction value of the action continuation degree.

Description

Learning device, prediction device, method, and program

[0001] The present disclosure relates to a learning device, a prediction device, a method, and a program.

[0002] In a dialogue aimed at determining a target person's action plan (e.g., a dialogue during a health guidance interview), it is important to understand the degree to which the target person continues to act in accordance with the action plan. As a related technique, a technique for predicting the user's intention and satisfaction level of a dialogue-based recommendation system is known (e.g., Non-Patent Document 1).

[0003] Wanling Cai, Li Chen, Predicting User Intents and Satisfaction with Dialogue-based Conversational Recommendations, In Proceedings of the 28th ACM Conference on User Modeling, Adaptation and Personalization. 33-42.

[0004] However, conventional techniques including the technique described in Non-Patent Document 1 have not been able to predict the degree to which a subject will continuously perform behavior in accordance with a behavior plan.

[0005] The present disclosure has been made in consideration of the above points, and provides a technology that can predict the continuity of behavior according to a behavior plan.

[0006] A learning device according to one aspect of the present disclosure includes an input unit that inputs input data including dialogue data of a dialogue intended to determine a behavior plan for a subject and a correct answer for behavior continuity that represents the degree to which the subject continuously performs behavior in accordance with the behavior plan, a prediction unit that predicts a predicted value of the behavior continuity based on features of the input data and a prediction model that inputs the features and outputs a predicted value of the behavior continuity, and an update unit that updates learnable parameters of the prediction model based on the correct answer for the behavior continuity and the predicted value of the behavior continuity.

[0007] It is possible to predict the degree of continuity of behavior according to an action plan.

[0008] FIG. 1 is a diagram illustrating an example of a hardware configuration of a prediction device according to the present embodiment; FIG. 2 is a diagram illustrating an example of a functional configuration of a learning phase of a prediction device according to the present embodiment; FIG. 3 is a flowchart illustrating an example of a learning process according to the present embodiment; FIG. 4 is a diagram illustrating an example of dialogue data; FIG. 5 is a diagram illustrating an example of dialogue data after preprocessing; FIG. 6 is a diagram illustrating an example of a functional configuration of an inference phase of a prediction device according to the present embodiment; FIG. 7 is a flowchart illustrating an example of a prediction process according to the present embodiment; FIG. 8 is a diagram illustrating an example of a graph representing a plan continuity degree.

[0009] An embodiment of the present invention will be described in detail below with reference to the drawings. In the following embodiment, a prediction device 10 will be described that can predict the degree to which a subject will continuously perform actions in accordance with an action plan when a dialogue is held between two or more users, including the subject, with the aim of determining an action plan for the subject. Hereinafter, an index value representing the degree to which a subject will continuously perform actions in accordance with an action plan will be referred to as a "plan continuity level." Furthermore, actions in accordance with an action plan will be referred to as "planned actions." The plan continuity level can be expressed in various formats, such as information representing whether or not a planned action will be performed after a predetermined number of days has passed using a discrete value; information representing the possibility of performing a planned action after a predetermined number of days has passed using a continuous value, probability, percentage, etc.; information representing the possibility of performing a planned action for each number of days that has elapsed until a predetermined number of days has passed using a continuous value, probability, percentage, etc.; and the like.

[0010] Here, the target person refers to a user for whom an action plan is to be established, and the action plan refers to a plan for the target person's future actions. A dialogue for the purpose of establishing an action plan takes place between two or more users. It is sufficient that at least the target person is a human. Users other than the target person may be humans, or may be machines, devices, programs, systems, etc. Examples of machines, devices, programs, systems, etc. include virtual people, characters, and interactive recommendation systems realized by technologies such as rule-based, knowledge graph, and artificial intelligence (AI). A dialogue between two or more users may also be called, for example, a "conversation" or a "discussion."

[0011] Furthermore, examples of dialogue aimed at establishing an action plan include dialogue between a subject and a specialist staff member (such as a public health nurse or a registered dietitian) during a health guidance interview. In the following, we will mainly consider dialogue during a health guidance interview as dialogue aimed at establishing an action plan, and refer to specialist staff members such as public health nurses and registered dietitians as "interviewers."

[0012] However, the dialogue in a health guidance interview is just one example, and there are various types of dialogues that aim to determine an action plan, such as dialogues in study guidance interviews at schools or cram schools, dialogues in personnel interviews at workplaces, dialogues in meetings at workplaces, etc.

[0013] The prediction device 10 according to this embodiment includes a learning phase in which a machine learning model is trained to output a plan continuity level using the feature quantities of input data including data representing the content of a dialogue between an interviewer and a subject in a dialogue aimed at establishing an action plan (hereinafter also referred to as "dialogue data"). The prediction device 10 according to this embodiment also includes an inference phase in which, when input data including dialogue data is provided, the trained machine learning model predicts the plan continuity level from the feature quantities of the input data. Hereinafter, a machine learning model that outputs a plan continuity level using the feature quantities of input data including dialogue data will be referred to as a "prediction model." Note that various machine learning models can be used as the prediction model, including, for example, a neural network, a Light Gradient Boosting Machine (LightGBM), etc.

[0014] <Example of Hardware Configuration of Prediction Device 10> An example of the hardware configuration of the prediction device 10 according to this embodiment is shown in Fig. 1. As shown in Fig. 1, the prediction device 10 according to this embodiment includes an input device 101, a display device 102, an external I / F 103, a communication I / F 104, a RAM (Random Access Memory) 105, a ROM (Read Only Memory) 106, an auxiliary storage device 107, and a processor 108. Each of these pieces of hardware is connected to each other via a bus 109 so as to be able to communicate with each other.

[0015] The input device 101 is, for example, a keyboard, a mouse, a touch panel, a physical button, etc. The display device 102 is, for example, a display, a display panel, etc. Note that the prediction device 10 does not necessarily have to include at least one of the input device 101 and the display device 102, for example.

[0016] The external I / F 103 is an interface with an external device such as a recording medium 103a. Examples of the recording medium 103a include a CD (Compact Disc), a DVD (Digital Versatile Disk), an SD memory card (Secure Digital memory card), and a USB (Universal Serial Bus) memory card.

[0017] The communication I / F 104 is an interface for connecting to a communication network. The RAM 105 is a volatile semiconductor memory (storage device) that temporarily stores programs and data. The ROM 106 is a non-volatile semiconductor memory (storage device) that can store programs and data even when the power is turned off. The auxiliary storage device 107 is a non-volatile storage device such as a hard disk drive (HDD), a solid state drive (SSD), or a flash memory. The processor 108 is one of various arithmetic devices such as a central processing unit (CPU) or a graphics processing unit (GPU).

[0018] 1 is an example, and the hardware configuration of the prediction device 10 is not limited to this. The prediction device 10 may have, for example, multiple auxiliary storage devices 107 or multiple processors 108, may not have some of the hardware shown in the figure, or may have various types of hardware other than the hardware shown in the figure.

[0019] [Learning Phase] The following describes the learning phase of the prediction device 10. Note that the prediction device 10 in the learning phase may be called, for example, a "learning device," a "training device," or a "prediction model creation device."

[0020] <Example of Functional Configuration of Prediction Device 10 in the Learning Phase> FIG. 2 shows an example of the functional configuration of the prediction device 10 in the learning phase according to this embodiment. As shown in FIG. 2 , the prediction device 10 in the learning phase includes an input unit 201, a preprocessing unit 202, a feature generation unit 203, a prediction unit 204, an update unit 205, and a termination determination unit 206. These units are realized, for example, by a process in which one or more programs installed in the prediction device 10 are executed by the processor 108 or the like. The prediction device 10 in the learning phase also includes a prediction model storage unit 301 and a training data storage unit 302. These storage units are realized, for example, by a storage area of ​​the auxiliary storage device 107 or the like. However, for example, at least one of the prediction model storage unit 301 and the training data storage unit 302 may be realized by a storage area of ​​a storage device (e.g., a storage device provided in a database server) communicatively connected to the prediction device 10.

[0021] The input unit 201 inputs the learning data stored in the learning data storage unit 302. The learning data is data for learning a prediction model, and is composed of input data including dialogue data and training data. The training data is data that represents the correct answer to the degree of plan continuity output when the feature values ​​of the input data (more precisely, the feature values ​​of the preprocessed input data after preprocessing the input data) are input to the prediction model. Note that the input data included in the learning data may be referred to as "explanatory variable data" or simply "explanatory variables," and the training data may be referred to as "objective variable data" or simply "objective variables."

[0022] For simplicity, unless otherwise specified, input data will be considered to include only dialogue data, and input data and dialogue data will be considered to be the same. Here, dialogue data refers to data representing the content of a dialogue intended to determine an action plan, such as speaker-labeled text data obtained by converting the voice data of the dialogue into text with a speaker label using a voice recognition technology capable of speaker recognition. Note that a speaker label is label information for identifying which speaker spoke a certain text.

[0023] However, the input data may include, in addition to the dialogue data, data representing various information related to the dialogue intended to determine an action plan, such as data representing personal information such as the age and gender of at least one of the interviewer and the interviewee (so-called demographic data), and data representing the interview format (e.g., online interview / offline interview, etc.).

[0024] The preprocessing unit 202 performs predetermined preprocessing on input data included in the training data input by the input unit 201. Specifically, the preprocessing unit 202 performs data cleansing, such as correcting typographical errors in text represented by the dialogue data, deleting fillers, deleting speaker labels, etc., on dialogue data, which is input data included in the training data. However, preprocessing on input data is not essential; for example, if the dialogue data does not contain typographical errors, fillers, speaker labels, etc., preprocessing on the dialogue data may not be performed.

[0025] The feature generation unit 203 generates features of the input data after preprocessing by the preprocessing unit 202. Specifically, the feature generation unit 203 divides the text represented by the dialogue data into token units, and generates an embedding vector representing the embedded expression of each token as a feature. Note that a token refers to a word, character, subword, a certain coherent character string, etc.

[0026] In addition, when the input data includes data representing various pieces of information related to the dialogue, the feature generating unit 203 may also use the features obtained by digitizing or vectorizing such information as the features of the preprocessed input data.

[0027] The prediction unit 204 predicts the degree of plan continuity by using the prediction model stored in the prediction model storage unit 301 and the feature quantities generated by the feature quantity generation unit 203 and inputting the feature quantities into the prediction model. Hereinafter, the degree of plan continuity predicted by the prediction unit 204 will also be referred to as the "plan continuity prediction value."

[0028] The update unit 205 updates the learnable parameters of the prediction model using the error between the teacher data included in the learning data input by the input unit 201 and the plan continuity prediction value predicted by the prediction unit 204. In other words, the update unit 205 updates the learnable parameters of the prediction model so as to minimize the objective function including the error by a known optimization method.

[0029] Various error values ​​can be used depending on the format of the plan continuity. For example, if the plan continuity is information that represents, as a discrete value, whether or not a planned behavior will be implemented after a predetermined number of days has passed, predicting the plan continuity is a classification task, and therefore cross-entropy error, etc., can be used. On the other hand, if the plan continuity is information that represents the possibility of implementing a planned behavior after a predetermined number of days has passed as a continuous value, probability, percentage, etc., or information that represents the possibility of implementing a planned behavior for each number of days until a predetermined number of days has passed as a continuous value, probability, percentage, etc., predicting the plan continuity is a regression task, and therefore root mean square error, mean square error, etc. can be used. Furthermore, the objective function may be, for example, the error, or may be an objective function obtained by adding a regularization term (e.g., an L1 regularization term, an L2 regularization term, etc.) to the error.

[0030] The termination determination unit 206 determines whether to terminate the update of the training parameters of the prediction model. For example, if a predetermined termination condition is met, the termination determination unit 206 determines to terminate the update of the training parameters of the prediction model; otherwise, the termination determination unit 206 determines not to terminate the update of the training parameters of the prediction model. Examples of the predetermined termination condition include the number of updates of the training parameters of the prediction model exceeding a predetermined threshold, the training parameters of the prediction model converging, and the number of epochs exceeding a predetermined threshold. The number of epochs represents how many times one piece of training data has been used to update the training parameters. For example, the number of epochs is initialized to 0 and is incremented by 1 each time all of the training data stored in the training data storage unit 302 is used to update the training parameters.

[0031] The prediction model storage unit 301 stores a prediction model and its learning target parameters. Note that the prediction model storage unit 301 may also store hyperparameters, such as a learning rate used to update the learning target parameters and a coefficient for a regularization term, in addition to the prediction model and its learning target parameters.

[0032] The learning data storage unit 302 stores learning data. The learning data may be called, for example, "training data." A data set consisting of all the learning data stored in the learning data storage unit 302 may be called, for example, a "learning data set."

[0033] <Learning Process> The learning process according to this embodiment will be described below with reference to Fig. 3. It is assumed below that the trainable parameters of the prediction model are initialized to appropriate values. Note that, as an example, the case where the prediction model is trained by online learning will be described below. However, this is merely an example, and the prediction model may also be trained by, for example, mini-batch learning, batch learning, or the like.

[0034] The input unit 201 inputs, from among the learning data stored in the learning data storage unit 302, learning data that has not yet been input in the current epoch number (step S101).

[0035] The preprocessing unit 202 performs predetermined preprocessing on the input data included in the training data input in step S101 (step S102). That is, the preprocessing unit 202 performs data cleansing on the dialogue data, which is the input data included in the training data, as preprocessing, such as correcting typos in the text represented by the dialogue data, deleting fillers, deleting speaker labels, etc.

[0036] As an example, a case will be described in which filler deletion and speaker label deletion are performed on the dialogue data 1000 shown in FIG. 4 . The text represented by the dialogue data 1000 shown in FIG. 4 includes a filler 1011, "um," and a filler 1012, "um,". The text represented by the dialogue data 1000 shown in FIG. 4 also includes speaker labels 1021 to 1023 representing subjects and speaker labels 1031 to 1032 representing interviewers. In this case, the preprocessing unit 202 deletes the filler 1011, the filler 1012, the speaker labels 1021 to 1023, and the speaker labels 1031 to 1032 from the text represented by the dialogue data 1000 shown in FIG. 4 . As a result, dialogue data 1100 shown in FIG. 5 is obtained as preprocessed dialogue data.

[0037] The feature generation unit 203 generates features of the input data after the preprocessing in step S102 (step S103). That is, the feature generation unit 203 divides the text represented by the dialogue data, which is the input data after the preprocessing in step S102, into tokens, and generates an embedding vector representing the embedded expression for each token as a feature. As a result, a series of embedding vectors is obtained as the features of the input data after the preprocessing.

[0038] In addition, when the input data contains data representing various information related to the dialogue (e.g., personal information of the interviewer or interviewee, interview format, etc.), for example, the information can be quantified or vectorized, or a series of such information, and a series of embedding vectors can be obtained as features of the input data after preprocessing.

[0039] The prediction unit 204 predicts the plan continuity prediction value by using the prediction model stored in the prediction model storage unit 301 and the feature values ​​generated in step S103 (step S104). That is, the prediction unit 204 predicts the plan continuity prediction value by inputting the feature values ​​generated in step S103 into the prediction model.

[0040] The update unit 205 updates the learnable parameters of the prediction model using the error between the teacher data (i.e., the correct answer for the plan continuity prediction value) contained in the learning data input in the above step S101 and the plan continuity prediction value predicted in the above step S104 (step S105).

[0041] The termination determination unit 206 determines whether a predetermined termination condition is met (step S106).

[0042] If it is determined in step S106 that the predetermined termination condition is satisfied, the update unit 205 terminates the learning process.

[0043] On the other hand, if it is not determined in step S106 that the predetermined termination condition is satisfied, the update unit 205 returns to step S101. At this time, if there is no learning data that has not yet been input for the current epoch number among the learning data stored in the learning data storage unit 302, the update unit 205 increments the epoch number by 1 and returns to step S101.

[0044] Through the above learning process, the learnable parameters of the prediction model are learned. Hereinafter, the learned learnable parameters will also be referred to as "learned parameters." Furthermore, a prediction model whose learned parameters are learned will also be referred to as a "learned prediction model."

[0045] [Inference Phase] The inference phase of the prediction device 10 will be described below.

[0046] <Example of Functional Configuration of Prediction Device 10 in Inference Phase> FIG. 6 shows an example of the functional configuration of the prediction device 10 in the inference phase according to this embodiment. As shown in FIG. 6 , the prediction device 10 in the inference phase includes an input unit 201, a preprocessing unit 202, a feature generation unit 203, a prediction unit 204, and an output unit 207. These units are realized, for example, by a process in which one or more programs installed in the prediction device 10 are executed by the processor 108 or the like. The prediction device 10 in the inference phase also includes a prediction model storage unit 301. The prediction model storage unit 301 is realized, for example, by a storage area of ​​the auxiliary storage device 107 or the like. However, for example, the prediction model storage unit 301 may also be realized by a storage area of ​​a storage device (e.g., a storage device provided in a database server) communicatively connected to the prediction device 10.

[0047] When input data including dialogue data for which the plan continuity degree is to be predicted (hereinafter also referred to as "target input data") is given, the input unit 201 inputs the target input data.

[0048] For simplicity, unless otherwise specified, the target input data will be considered to include only dialogue data, and the target input data and dialogue data will be regarded as the same thing. However, the target input data may also include data other than dialogue data that represents various information related to dialogue aimed at determining an action plan.

[0049] The preprocessing unit 202 performs predetermined preprocessing on the target input data input by the input unit 201. The preprocessing method by the preprocessing unit 202 is the same as that in the learning phase.

[0050] The feature generating unit 203 generates feature amounts of the target input data after preprocessing by the preprocessing unit 202. The method of generating feature amounts by the feature generating unit 203 is the same as in the learning phase.

[0051] The prediction unit 204 uses the trained prediction model stored in the prediction model memory unit 301 and the features generated by the feature generation unit 203 to predict the plan continuity by inputting the features into the trained prediction model.

[0052] The output unit 207 outputs the plan continuity predicted by the prediction unit 204 to a predetermined output destination. Examples of the output destination include the display device 102 such as a display, a storage area such as the auxiliary storage device 107, and a display provided on a terminal connected in a communicable manner.

[0053] The prediction model storage unit 301 stores a trained prediction model and its trained parameters.

[0054] <Prediction Process> The prediction process according to this embodiment will be described below with reference to Fig. 7. In the following, it is assumed that target input data is provided to the prediction device 10.

[0055] The input unit 201 receives given target input data (step S201).

[0056] The preprocessing unit 202 performs predetermined preprocessing on the target input data input in step S201 (step S202). That is, the preprocessing unit 202 performs data cleansing on the dialogue data, which is the target input data, as preprocessing, such as correcting typos in the text represented by the dialogue data, deleting fillers, deleting speaker labels, etc.

[0057] The feature generation unit 203 generates features of the target input data after the preprocessing in step S202 (step S203). That is, the feature generation unit 203 divides the text represented by the dialogue data, which is the target input data after the preprocessing in step S202, into tokens, and generates an embedding vector representing the embedded expression for each token as a feature. As a result, a series of embedding vectors is obtained as the features of the target input data after the preprocessing.

[0058] In addition, when the target input data contains data representing various information related to the dialogue (e.g., personal information of the interviewer or the subject, the interview format, etc.), for example, the information can be quantified or vectorized, or a series of such information, and a series of embedded vectors can be obtained as features of the target input data after preprocessing.

[0059] The prediction unit 204 predicts the plan continuity prediction value using the trained prediction model stored in the prediction model storage unit 301 and the feature values ​​generated in step S203 (step S204). That is, the prediction unit 204 predicts the plan continuity prediction value by inputting the feature values ​​generated in step S203 into the trained prediction model.

[0060] The output unit 207 outputs the plan continuity prediction value predicted in step S204 to a predetermined output destination (step S205).

[0061] Through the above prediction process, when a dialogue is held between two or more users, including a target person, with the aim of determining an action plan for the target person, a plan continuity prediction value can be obtained that represents the degree to which the target person will continue to act in accordance with the action plan.

[0062] <Output Example of Plan Continuity Prediction Value> An output example of the plan continuity prediction value will be described below.

[0063] Output example 1: The plan continuity is information that indicates whether or not a planned action will be carried out after a predetermined number of days, i.e., X days, has passed, using 0 or 1. In this case, the plan continuity prediction value is output in a format such as "Plan continued after X days" or "Plan not continued after X days."

[0064] Output Example 2: Plan continuity is information that expresses the possibility of implementing a planned action after a predetermined number of days, i.e., X days, has passed, as a percentage. In this case, when the possibility of implementation is a%, the predicted value of plan continuity is output in a format such as "Plan continuity possibility after X days has passed: a%."

[0065] Output Example 3: The plan continuity is information that represents, as a percentage, the feasibility of implementing a planned action for each number of days that has elapsed until a predetermined number of days, i.e., 100 days, has elapsed. In this case, the plan continuity prediction value is output, for example, as a graph 2000 shown in Fig. 8. The graph 2000 shown in Fig. 8 represents the plan continuity prediction value for each number of days that has elapsed, with the vertical axis representing the plan continuity value and the horizontal axis representing the number of days that have elapsed.

[0066] [Modifications] Modifications of the above embodiment will be described below. Note that the following modifications can be combined as appropriate as long as they do not contradict each other.

[0067] Variation 1: Even after the learning phase has been executed once, the learning phase may be executed again using the learned parameters as the initial values ​​of the learnable parameters. Furthermore, in this case, learning data may be used in which the target input data of the inference phase executed before the learning phase is used as the "input data," and the actual value of the plan continuity for the action plan determined in the dialogue represented by the dialogue data included in the target input data is used as the "teacher data." This allows the learnable parameters of the predictive model to be re-learned, which is expected to result in a more accurate predictive model.

[0068] Variation 2 In the above embodiment, the embedded representation of the text represented by the dialogue data is used as the feature of the dialogue data. However, the feature of the dialogue data may be generated by, for example, TF-IDF (Term Frequency-Inverse Document Frequency).

[0069] Variation 3 The features of the input data may include, for example, the emotion estimation result using the text represented by the dialogue data, or may include, as a feature, a change in the frequency of fillers deleted by the preprocessing unit 202. The change in the frequency of fillers refers to, for example, the change over time in the number of fillers that occur in the text represented by the dialogue data (the change over time). This is because fillers are likely to increase when expressing one's own thoughts, and this change over time is used as a feature.

[0070] Variation 4: The dialogue data included in the target input data does not need to be dialogue data representing a dialogue from the start to the end of a dialogue aimed at determining the target's action plan, but may be dialogue data representing a dialogue from the start to the present time of a dialogue aimed at determining the target's action plan. This makes it possible to obtain a predicted value of the target's plan continuity even in the middle of a dialogue aimed at determining the target's action plan.

[0071] Furthermore, for example, target input data including dialogue data representing a dialogue over a certain time span may be used to sequentially predict a plan continuity prediction value for each time span. This makes it possible to sequentially obtain a plan continuity prediction value for the target person for each time span, and for example, by ensuring a desired accuracy and making the time span as short as possible, it becomes possible to obtain a plan continuity prediction value with real-time properties.

[0072] [Summary] As described above, the prediction device 10 according to the present embodiment can predict the continuity of behavior according to the action plan using dialogue data from a dialogue aimed at formulating an action plan, using a trained prediction model. This allows, for example, the interviewer of the interview in which the dialogue took place to take measures such as providing early follow-up to a subject who has low continuity of behavior according to the action plan. Furthermore, for example, the interviewer can use the predicted results of the continuity of behavior according to the action plan to review the interview and aim to achieve a more effective dialogue. Furthermore, since the prediction results can be obtained immediately after the interview, for example, the interviewer can immediately determine whether the action plan was successful in motivating the subject and can correct the course of the dialogue as necessary.

[0073] The present invention is not limited to the above-described specifically disclosed embodiments, and various modifications, changes, and combinations with known technologies are possible without departing from the scope of the claims.

[0074] REFERENCE SIGNS LIST 10 Prediction device 101 Input device 102 Display device 103 External I / F 103a Recording medium 104 Communication I / F 105 RAM 106 ROM 107 Auxiliary storage device 108 Processor 109 Bus 201 Input unit 202 Preprocessing unit 203 Feature generation unit 204 Prediction unit 205 Update unit 206 End determination unit 207 Output unit 301 Prediction model storage unit 302 Learning data storage unit

Claims

1. A learning device having: an input unit that inputs input data including dialogue data of a dialogue intended to determine a behavior plan for a subject, and a correct answer to a behavior continuity index that indicates the degree to which the subject continuously performs behavior in accordance with the behavior plan; a prediction unit that predicts a predicted value of the behavior continuity index based on features of the input data and a prediction model that inputs the features and outputs a predicted value of the behavior continuity index; and an update unit that updates learnable parameters of the prediction model based on the correct answer to the behavior continuity index and the predicted value of the behavior continuity index.

2. A prediction device having: an input unit that inputs input data including dialogue data of a dialogue intended to determine a behavior plan for a subject; a prediction unit that predicts a predicted value of behavior continuity based on features of the input data and a learned prediction model that uses the features as input and outputs a predicted value of behavior continuity that represents the degree to which the subject will continuously perform behavior in accordance with the behavior plan; and an output unit that outputs the predicted value of behavior continuity to a specified output destination.

3. A method in which a computer executes the following steps: an input step of inputting input data including dialogue data of a dialogue intended to determine a behavior plan for a subject, and a correct answer to a behavior continuity index that indicates the degree to which the subject continuously performs behavior in accordance with the behavior plan; a prediction step of predicting a predicted value of the behavior continuity index based on features of the input data and a prediction model that inputs the features and outputs a predicted value of the behavior continuity index; and an update step of updating learnable parameters of the prediction model based on the correct answer to the behavior continuity index and the predicted value of the behavior continuity index.

4. A program that causes a computer to execute the following steps: an input step of inputting input data including dialogue data of a dialogue intended to determine a behavior plan for a subject, and a correct answer to a behavior continuity index that indicates the degree to which the subject continuously performs behavior in accordance with the behavior plan; a prediction step of predicting a predicted value of the behavior continuity index based on features of the input data and a prediction model that inputs the features and outputs a predicted value of the behavior continuity index; and an update step of updating learnable parameters of the prediction model based on the correct answer to the behavior continuity index and the predicted value of the behavior continuity index.

Images