Estimation device, estimation method, and estimation program

Abstract

A calculation unit (131): uses an encoder to calculate an observation variable with a high acquisition cost on the basis of an input observation variable with a low acquisition cost; uses the encoder and an input observation variable with a high acquisition cost to correct the observation variable with a high acquisition cost calculated using the encoder; and uses the encoder to calculate a latent variable on the basis of the corrected observation variable with a high acquisition cost. The calculation unit (131): uses a decoder to calculate an observation variable with a high acquisition cost on the basis of the latent variable calculated using the encoder; and uses the decoder to calculate an observation variable with a low acquisition cost on the basis of the observation variable with a high acquisition cost calculated using the decoder. An update unit (132) updates the parameters of the encoder and the decoder on the basis of the calculation result of the calculation unit (131).

Description

Estimation device, estimation method, and estimation program

[0001] The present invention relates to an estimation device, an estimation method, and an estimation program.

[0002] Conventionally, methods for analyzing human psychological states have been known. For example, WB (Well-being) may be measured based on a questionnaire. Another known method is to analyze posts on social networking services (SNS) using Russell's circular model (see, for example, Non-Patent Document 1).

[0003] Daniel Preotiuc-Pietro, H. Andrew Schwartz, Gregory Park, Johannes Eichstaedt, Margaret Kern, Lyle Ungar, Elisabeth Shulman, "Modeling Valence and Arousal in Facebook Posts," Proceedings of the 7th Workshop on Computational Approaches to Subjectivity, Sentiment and Social Media Analysis, 2016, [online], [Retrieved April 11, 2024], Internet (https: / / aclanthology.org / W16-0404 / )

[0004] However, conventional techniques often fail to efficiently measure WB. For example, when measuring WB based on a questionnaire, the questionnaire items for WB related to company meetings and WB at home are different, making it difficult to select effective items. Another problem is the high cost of conducting questionnaires. Furthermore, while arousal and valence in Russell's circumplex model can be used as measures of WB, determining arousal and valence requires specialized knowledge.

[0005] In order to solve the above-described problems and achieve the object, an estimation device includes: a first calculation unit that calculates a second observation variable using an encoder based on an input first observation variable, corrects the second observation variable calculated using the encoder using the input second observation variable, and calculates a latent variable using the encoder based on the corrected second observation variable; a second calculation unit that calculates the second observation variable using a decoder based on the latent variable calculated by the first calculation unit, and calculates the first observation variable using the decoder based on the second observation variable calculated using the decoder; and an update unit that updates parameters of the encoder and the decoder based on a calculation result of the second calculation unit.

[0006] According to the present invention, WB measurement can be carried out efficiently.

[0007] FIG. 1 is a diagram illustrating an example of the configuration of an estimation device according to a first embodiment. FIG. 2 is a diagram illustrating problem setting. FIG. 3 is a diagram illustrating an encoder. FIG. 4 is a diagram illustrating a decoder. FIG. 5 is a diagram illustrating Russell's circular ring model. FIG. 6 is a diagram illustrating transfer. FIG. 7 is a diagram illustrating a sample dataset. FIG. 8 is a diagram illustrating some samples. FIG. 9 is a flowchart illustrating the flow of a learning process. FIG. 10 is a flowchart illustrating the flow of an inference process using both low-cost observed variables and high-cost observed variables for inference. FIG. 11 is a flowchart illustrating the flow of an inference process using low-cost observed variables. FIG. 12 is a flowchart illustrating the flow of an inference process using high-cost observed variables. FIG. 13 is a diagram illustrating an example of a unique WB. FIG. 14 is a diagram illustrating an example of a computer that executes an estimation program.

[0008] DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0023] Exemplary embodiments of an estimation device, an estimation method, and an estimation program according to the present application will be described in detail below with reference to the accompanying drawings. However, the present invention is not limited to the following exemplary embodiments.

[0009] [Configuration of First Embodiment] First, the configuration of an estimation device according to a first embodiment will be described with reference to Fig. 1. Fig. 1 is a diagram showing an example of the configuration of an estimation device according to a first embodiment. For example, the estimation device 10 is a personal computer, a server device, a smartphone, a tablet terminal, or the like.

[0010] The estimation device 10 aims to measure true WB as accurately as possible from observation variables that have low acquisition costs. For example, text posted on a social networking site (SNS) can be an observation variable with low acquisition costs. On the other hand, arousal and valence determined based on text posted on a social networking site require specialized knowledge, and therefore can be considered observation variables with higher acquisition costs than the text itself.

[0011] Compared to observable variables with low acquisition costs, it is easier to measure WB from observable variables with high acquisition costs. For example, arousal and valence can be indicators close to WB. On the other hand, it is difficult to grasp WB directly from text.

[0012] The problem handled by the estimation device 10 is shown in Figure 2. Figure 2 is a diagram for explaining the problem setting. X is an observation variable with a high acquisition cost. X' is an observation variable with a low acquisition cost. H is a latent variable representing the true WB.

[0013] For example, by actually surveying meeting participants, observational variables with high acquisition costs, such as whether the meeting is going well or not and whether communication is smooth or not, can be obtained. On the other hand, observational variables with low acquisition costs, such as the number of times meeting participants speak, the content of their comments, and whether they initiate conversations, can be obtained by simply observing the state of the meeting without conducting a survey. However, observational variables with low acquisition costs have more noise than observational variables with high acquisition costs.

[0014] The problem is formulated as in equation (1).

[0015]

[0016] Here, the estimation device 10 aims to estimate H from X and X'. By maximizing the joint distribution of X and X', loss functions are designed as shown in equations (2-1) to (2-8). D KL is the KL divergence.

[0017]

[0018] Here, q is an approximate distribution, and equation (2-8) is a variational lower bound of logP(X,X'). In other words, by maximizing equation (2-8), q(H|X,X') approaches the true posterior distribution p(H|X,X'). q(H|X,X') is the encoder, and p(X'|X)p(X|H) is the decoder.

[0019] The estimation device 10 trains a model including an encoder and a decoder. Using the trained model, the estimation device 10 can obtain a latent variable H representing the true WB from at least one of an observation variable X′ with a low acquisition cost and an observation variable X with a high acquisition cost.

[0020] FIG. 3 is a diagram illustrating an encoder. The encoder calculates X from X' (step S11). If X is input, the encoder corrects the calculated X using the input X (step S12). If X is not input, the encoder does not correct the calculated X. Then, the encoder calculates H from X (step S13). If X' is not input, the encoder calculates H from the input X. In this way, the encoder can calculate H whether both X and X' are input during inference, or whether only one of them is input.

[0021] 4 is a diagram illustrating a decoder. The decoder calculates X from H (step S21). Furthermore, the decoder calculates X' from X (step S22). When both X and X' are input in training, the decoder can improve the accuracy of the encoder.

[0022] Returning to FIG. 1 , the estimation device 10 includes a communication unit 11 , a storage unit 12 , and a control unit 13 .

[0023] The communication unit 11 performs data communication with other devices via a network, and is, for example, a network interface card (NIC).

[0024] The storage unit 12 is a storage device such as a hard disk drive (HDD), a solid state drive (SSD), an optical disk, etc. Note that the storage unit 12 may also be a data-rewritable semiconductor memory such as a random access memory (RAM), a flash memory, or a non-volatile static random access memory (NVSRAM). The storage unit 12 stores an operating system (OS) and various programs executed by the estimation device 10.

[0025] The storage unit 12 stores model information 121 , low-cost observation variables 122 , and high-cost observation variables 123 .

[0026] The model information 121 is information about a model. For example, the model information 121 is parameters for constructing an encoder and a decoder. For example, the parameters are weights and biases in a neural network that functions as an encoder and a decoder.

[0027] The low-cost observed variable 122 and the high-cost observed variable 123 are X′ and X for training, respectively.

[0028] The control unit 13 controls the entire estimation device 10. The control unit 13 is, for example, an electronic circuit such as a central processing unit (CPU), a micro processing unit (MPU), or a graphics processing unit (GPU), or an integrated circuit such as an application specific integrated circuit (ASIC) or a field programmable gate array (FPGA). The control unit 13 also has an internal memory for storing programs that define various processing procedures and control data, and executes each process using the internal memory.

[0029] The control unit 13 functions as various processing units by running various programs. For example, the control unit 13 includes a calculation unit 131 and an update unit 132.

[0030] The calculation unit 131 performs calculations using an encoder and a decoder constructed based on the model information 121. During inference, the calculation unit 131 calculates H using the encoder and outputs the calculated H. During learning, the calculation unit 131 calculates H using the encoder and calculates X and X' from H using the decoder.

[0031] 3, the calculation unit 131 calculates an observation variable X with a high acquisition cost using an encoder based on an input observation variable X' with a low acquisition cost, corrects the observation variable X with a high acquisition cost calculated using the encoder using the input observation variable X with a high acquisition cost, and calculates a latent variable H using the encoder based on the corrected observation variable X with a high acquisition cost. During inference, the calculation unit 131 outputs the latent variable H.

[0032] Furthermore, the calculation unit 131 calculates an observation variable X with a high acquisition cost using a decoder based on the latent variable H calculated using the encoder, and calculates an observation variable X' with a low acquisition cost using a decoder based on the observation variable X with a high acquisition cost calculated using the decoder.

[0033] The update unit 132 updates the parameters of the encoder and decoder, that is, the model information 121, so that the loss function shown in equation (2-8) becomes larger. For example, the update unit 132 updates the parameters by backpropagation.

[0034] This allows the estimation device 10 to measure true WB from multiple surveys. Furthermore, the estimation device 10 can predict survey results (observation variables with high acquisition costs) from utterances (observation variables with low acquisition costs). As a result, even when observation variables with particularly high acquisition costs cannot be acquired, the estimation device 10 can efficiently measure WB.

[0035] For example, an observation variable with a low acquisition cost is a text entered by a user, and an observation variable with a high acquisition cost is a result of a multi-level evaluation that the user fills out in a questionnaire related to the text.

[0036] Russell's circular model is shown in Figure 5. Figure 5 is a diagram explaining Russell's circular model. Arousal and valence correspond to the vertical and horizontal axes, respectively. Emotions such as "HAPPY," "ANGRY," "SAD," and "RELAXED" are assigned to each quadrant. In this example, the greater the arousal and valence, the more improved the WB.

[0037] The estimation device 10 can perform model transfer. FIG. 6 is a diagram illustrating transfer. For example, the estimation device 10 performs learning by defining text as X' and arousal and valence in Russell's ring model as X. Then, the estimation device 10 can perform inference by defining text as X' and a unique WB that is different from the arousal and valence in Russell's ring model as X.

[0038] (Experimental Results) The results of an experiment using this embodiment will be described. In the experiment, approximately 3,000 pieces of text posted on a social networking site were used as X'. Two experts scored each piece of text on two WB scales, Arousal and Valence, from 1 to 9. This resulted in the data set shown in FIG. 7. FIG. 7 is a diagram showing a sample data set. In the calculation, A, which is the average of Arousal1 and Arousal2, and V, which is the average of Valence1 and Valence2, were used as X. H was two-dimensional data.

[0039] First, as a preliminary experiment, we embedded text (Text in Figure 7) using gpt3.5 and visualized it using t-sne. The results showed that there was little relationship between the proximity of words in the text and the proximity of A and V.

[0040] Experiments were then conducted on the cases where H was calculated from only X', only X, and both X and X'. In particular, the accuracy was significantly improved when H was calculated from both X and X'.

[0041] Figure 8 shows some samples. Figure 8 shows the results of taking four samples from the larger values ​​of one dimension of H (top four rows of Figure 8) and four samples from the smaller values ​​(bottom four rows of Figure 8) when calculating H from only X. From Figure 8, it can be said that the influence of WB is reflected in the latent variable H.

[0042] [Processing of First Embodiment] Fig. 9 is a flowchart showing the flow of the learning process. As shown in Fig. 9, the estimation device 10 first reads low-cost observation variables for learning and high-cost observation variables for learning (step S101). The estimation device 10 also reads parameters of a model including an encoder and a decoder (step S102).

[0043] Next, the estimation device 10 encodes the low-cost observation variables for training to calculate high-cost observation variables (step S103), and then corrects the high-cost observation variables for training using the high-cost observation variables obtained by encoding (step S104).

[0044] The estimation device 10 then encodes the corrected high-cost observed variables to calculate latent variables (step S105), decodes the calculated latent variables to calculate high-cost observed variables (step S106), and decodes the high-cost observed variables obtained by decoding to calculate low-cost observed variables (step S107).

[0045] The estimation device 10 calculates a loss function based on each variable (step S108), and updates the model parameters so as to reduce the loss function (step S109).

[0046] If the learning has not finished (step S110, No), the estimation device 10 returns to step S102 and repeats the process. If the learning has finished (step S110, Yes), the estimation device 10 ends the process. For example, the estimation device 10 determines that the learning has finished when the parameter updates have been performed a certain number of times, when the parameter update amount has converged, etc.

[0047] 10 is a flowchart showing the flow of inference processing using both low-cost observed variables and high-cost observed variables for inference. As shown in FIG. 10, the estimation device 10 first reads low-cost observed variables for inference and high-cost observed variables for inference (step S201). The estimation device 10 also reads parameters of a model including an encoder (step S202).

[0048] The estimation device 10 encodes the low-cost observed variables for inference to calculate high-cost observed variables (step S203), and then corrects the high-cost observed variables for inference using the high-cost observed variables obtained by encoding (step S204).

[0049] Furthermore, the estimation device 10 encodes the corrected high-cost observed variables to calculate latent variables (step S205), and outputs the calculated latent variables (step S206).

[0050] 11 is a flowchart showing the flow of inference processing using low-cost observed variables. As shown in FIG. 11, the estimation device 10 first reads low-cost observed variables for inference (step S211). The estimation device 10 also reads parameters of a model including an encoder (step S212).

[0051] The estimation device 10 encodes the low-cost observed variables for inference to calculate high-cost observed variables (step S213), then encodes the high-cost observed variables obtained by encoding to calculate latent variables (step S214), and outputs the calculated latent variables (step S215).

[0052] 12 is a flowchart showing the flow of inference processing using high-cost observed variables. As shown in FIG. 12, the estimation device 10 first reads high-cost observed variables for inference (step S221). The estimation device 10 also reads parameters of a model including an encoder (step S222).

[0053] The estimation device 10 encodes the high-cost observed variables for inference to calculate latent variables (step S223), and outputs the calculated latent variables (step S224).

[0054] (Transfer Experiment Results) Fig. 13 is a diagram showing an example of a unique WB used in a transfer experiment. Fig. 13 shows an example of a unique WB in the transfer described in Fig. 6.

[0055] In the experiment, six players played a cooperative game, and then responded to a questionnaire with a five-point rating and free text about their impressions. The five-point rating represents a unique WB that replaces Arousal and Valence, and corresponds to the observation variable X, which has a high acquisition cost. The free text corresponds to the observation variable X', which has a low acquisition cost. In this experiment, too, improvements were observed in the WB measurement results by performing inference using both X and X'.

[0056] [Example] The estimation device 10 according to the embodiment can estimate the WB of an employee based on text acquired from, for example, an in-house chat system. For example, when a new system is introduced as an in-house measure, the estimation device 10 can continuously estimate the WB of an employee without conducting a survey by collecting text related to the system circulated in the in-house chat as X'.

[0057] Specifically, in the learning phase, the estimation device 10 collects and learns from text of comments made in the chat system and survey results. Then, in the operation phase, the estimation device 10 estimates WB from the text of comments made in the chat system. If the estimation device 10 can obtain survey results in the operation phase, it can estimate WB with even greater accuracy.

[0058] [System Configuration, etc.] The components of each device shown in the figure are conceptual functional units and do not necessarily have to be physically configured as shown. In other words, the specific form of distribution and integration of each device is not limited to that shown in the figure, and all or part of the devices can be functionally or physically distributed or integrated in any unit depending on various loads, usage conditions, etc. Furthermore, all or any part of the processing functions performed by each device can be realized by a CPU (Central Processing Unit) and a program analyzed and executed by the CPU, or can be realized as hardware using wired logic. The program may be executed not only by the CPU but also by other processors such as a GPU.

[0059] Furthermore, among the processes described in this embodiment, all or part of the processes described as being performed automatically can be performed manually, or all or part of the processes described as being performed manually can be performed automatically using a known method.In addition, the information including the processing procedures, control procedures, specific names, various data and parameters shown in the above documents and drawings can be changed as desired unless otherwise specified.

[0060] [Program] In one embodiment, the estimation device 10 can be implemented by installing an estimation program that executes the above-described estimation process as package software or online software on a desired computer. For example, by causing an information processing device to execute the above-described estimation program, the information processing device can function as the estimation device 10. The information processing device referred to here includes desktop and notebook personal computers. In addition, information processing devices also include mobile communication terminals such as smartphones, mobile phones, and PHS (Personal Handyphone Systems), as well as slate terminals such as PDAs (Personal Digital Assistants).

[0061] The estimation device 10 can also be implemented as an estimation server device that provides a service related to the above-mentioned estimation process to a client terminal device used by a user. For example, the estimation server device is implemented as a server device that receives text as input and outputs an estimated WB. In this case, the estimation server device may be implemented as a web server or as a cloud that provides a service related to the above-mentioned estimation process through outsourcing.

[0062] 14 is a diagram showing an example of a computer that executes an estimation program. The computer 1000 includes, for example, a memory 1010 and a CPU 1020. The computer 1000 also includes a hard disk drive interface 1030, a disk drive interface 1040, a serial port interface 1050, a video adapter 1060, and a network interface 1070. These components are connected by a bus 1080.

[0063] The memory 1010 includes a read-only memory (ROM) 1011 and a random access memory (RAM) 1012. The ROM 1011 stores a boot program such as a basic input / output system (BIOS). The hard disk drive interface 1030 is connected to a hard disk drive 1090. The disk drive interface 1040 is connected to a disk drive 1100. A removable storage medium such as a magnetic disk or optical disk is inserted into the disk drive 1100. The serial port interface 1050 is connected to a mouse 1110 and a keyboard 1120, for example. The video adapter 1060 is connected to a display 1130, for example.

[0064] The hard disk drive 1090 stores, for example, an OS 1091, an application program 1092, a program module 1093, and program data 1094. That is, a program that defines each process of the estimation device 10 is implemented as a program module 1093 in which computer-executable code is written. The program module 1093 is stored, for example, in the hard disk drive 1090. For example, the program module 1093 for executing processes similar to those of the functional configuration of the estimation device 10 is stored in the hard disk drive 1090. Note that the hard disk drive 1090 may be replaced with an SSD (Solid State Drive).

[0065] Furthermore, setting data used in the processing of the above-described embodiment is stored as program data 1094, for example, in the memory 1010 or the hard disk drive 1090. Then, the CPU 1020 reads the program module 1093 or the program data 1094 stored in the memory 1010 or the hard disk drive 1090 into the RAM 1012 as necessary, and executes the processing of the above-described embodiment.

[0066] The program module 1093 and program data 1094 may not necessarily be stored in the hard disk drive 1090, but may also be stored in a removable storage medium and read by the CPU 1020 via the disk drive 1100 or the like. Alternatively, the program module 1093 and program data 1094 may be stored in another computer connected via a network (such as a local area network (LAN) or a wide area network (WAN)). The program module 1093 and program data 1094 may then be read by the CPU 1020 from the other computer via the network interface 1070.

[0067] REFERENCE SIGNS LIST 10 Estimation device 11 Communication unit 12 Storage unit 13 Control unit 121 Model information 122 Low-cost observation variable 123 High-cost observation variable 131 Calculation unit 132 Update unit

Claims

1. An estimation device comprising: a first calculation unit that calculates a second observation variable using an encoder based on an input first observation variable, corrects the second observation variable calculated using the encoder using the input second observation variable, and calculates a latent variable using the encoder based on the corrected second observation variable; a second calculation unit that calculates the second observation variable using a decoder based on the latent variable calculated by the first calculation unit, and calculates the first observation variable using the decoder based on the second observation variable calculated using the decoder; and an update unit that updates parameters of the encoder and the decoder based on the calculation result of the second calculation unit.

2. The estimation device described in claim 1, characterized in that the first calculation unit accepts input of text entered by a user as the first observation variable, and accepts input of the results of a multi-level evaluation filled out by the user in a questionnaire related to the text as the second observation variable.

3. An estimation method comprising: a first calculation step of calculating a second observation variable using an encoder based on an input first observation variable, correcting the second observation variable calculated using the encoder using the input second observation variable, and calculating a latent variable using the encoder based on the corrected second observation variable; a second calculation step of calculating the second observation variable using a decoder based on the latent variable calculated in the first calculation step, and calculating the first observation variable using the decoder based on the second observation variable calculated using the decoder; and an update step of updating parameters of the encoder and the decoder based on the calculation result of the second calculation step.

4. An estimation program that causes a computer to execute the following steps: a first calculation step of calculating a second observation variable using an encoder based on an input first observation variable, correcting the second observation variable calculated using the encoder using the input second observation variable, and calculating a latent variable using the encoder based on the corrected second observation variable; a second calculation step of calculating the second observation variable using a decoder based on the latent variable calculated in the first calculation step, and calculating the first observation variable using the decoder based on the second observation variable calculated using the decoder; and an update step of updating parameters of the encoder and the decoder based on the calculation result of the second calculation step.

Images