Learning device, inference device, learning method, and program

Abstract

Provided is a learning device for training a prediction model for acquiring a content label for an utterance sentence of an interviewer in an interview in which the interviewer interviews a subject, the learning device comprising: a feature amount generation unit for generating a feature amount of the utterance sentence, which is extracted from dialogue data; and a training unit for training the prediction model by using, as inputs, the feature amount and teacher data pertaining to a content label for the utterance sentence.

Description

Learning device, inference device, learning method, and program

[0001] The present invention relates to a technique for analyzing dialogue in health guidance interviews.

[0002] A health guidance interview is an interview conducted by specialized staff (public health nurses, registered dietitians, etc.) with individuals at high risk of developing lifestyle-related diseases to support them in reviewing their lifestyle habits. Note that hereinafter, "health guidance" is used in a broad sense to include both specific health guidance and health guidance other than specific health guidance.

[0003] Generally, in order to analyze dialogue and provide feedback, the system must understand the meaning of the dialogue. Therefore, research into automatic topic labeling has been actively conducted. However, such technology has not been applied to health guidance consultations.

[0004] A concept similar to health guidance is motivational interviewing (MI) (e.g., Non-Patent Document 1), which is mainly used for smoking cessation support. However, health guidance differs from MI in that the interviewer talks to the subject about their health checkup results.

[0005] Predicting Counselor Behaviors in Motivational Interviewing Encounters, Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 1, Long Papers, pages 1128-1137, Valencia, Spain, April 3-7, 2017.

[0006] As described above, conventional technologies have not been able to automatically annotate content labels in real time during a health guidance interview between an interviewer and a subject. This issue can arise in any interview, not just health guidance interviews. Note that a content label is a label that indicates the content of the conversation.

[0007] The present invention has been made in consideration of the above points, and aims to provide a technology that enables an interviewer to automatically assign content labels to spoken sentences during an interview with a subject.

[0008] According to the disclosed technology, there is provided a learning device that learns a predictive model for obtaining content labels for sentences spoken by an interviewer in an interview with a target person, the learning device comprising: a feature generation unit that generates features for the sentences spoken extracted from dialogue data; and a learning unit that learns the predictive model using as input the features and training data for content labels for the sentences spoken.

[0009] The disclosed technology provides a technology that enables an interviewer to automatically assign content labels to spoken sentences during an interview with a subject.

[0010] FIG. 1 is a diagram showing an example of the configuration of a learning device 100. FIG. 2 is a diagram showing examples of hyperparameters stored in a prediction model storage unit 140. FIG. 3 is a diagram showing an example of the configuration of an inference device 200. FIG. 4 is a flowchart for explaining an example of the operation of the learning device 100. FIG. 5 is a flowchart for explaining an example of the operation of the inference device 200. FIG. 6 is a diagram showing an output image of the inference device 200. FIG. 7 is a diagram showing an example of the hardware configuration of the device.

[0011] Hereinafter, an embodiment of the present invention will be described with reference to the drawings. The embodiment described below is merely an example, and the embodiment to which the present invention is applied is not limited to the following embodiment.

[0012] Below, first, the conventional technology and its problems will be described in more detail, and then the technology according to this embodiment will be described.

[0013] (Regarding Prior Art) Non-Patent Document 1 discloses a technology for automatically annotating labels for counselor utterances in MI. The technology in Non-Patent Document 1 is called MITI (Motivational Interviewing Treatment Integrity). The labels in the technology related to Non-Patent Document 1 include, for example, the following: question, simple response, complex response, cooperation, emphasis on autonomy, self-respect, affirmation, presentation of information, advice with permission, advice without permission, and confrontation.

[0014] The features used in the automatic annotation are as follows:

[0015] - N-gram: frequency of word occurrence - Semantic features: frequency of occurrence of words that encourage self-reflection - Similarity: degree of agreement between client and counselor - Syntactic features: number of occurrences of specific grammatical forms obtained from grammatical analysis The technology in Non-Patent Document 1 classifies labels using an SVM (Support Vector Machine), and classification with a set of semantic and syntactic features is possible with an accuracy of up to 90%.

[0016] In addition to the technology disclosed in Non-Patent Document 1, there is a technology disclosed in Reference Document 1 (Observing Dialogue in Therapy: Categorizing and Forecasting Behavioral Codes, Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 5599-5611 Florence, Italy, July 28 - August 2, 2019.).

[0017] Reference 1 also discloses a technology for automatically annotating labels for counselor utterances in MI. The technology in Reference 1 is called MISC (Motivational Interviewing Skill Code). Examples of labels in the technology related to Reference 1 include the following: conversation promotion, information provision, simple response, complex response, closed question, open question, MI compliance (affirmation, advice with permission), and MI independence (advice without permission, confrontation).

[0018] The technology disclosed in Reference 1 uses an encoding method using a GRU (Gated Recurrent Unit), and has achieved an F-measure of around 60 to 70 (minimum 29.7, maximum 95.0).

[0019] (Regarding the issues) Conventional technologies such as MITI and MISC evaluate so-called "communication skills," and do not check how the dialogue unfolds during the interview.

[0020] In the technology according to the present embodiment, the content label to be assigned indicates the "content of the conversation." The technology according to the present embodiment differs from the conventional technology not only in the interview itself that is the target of the technology, but also in the label itself that is the target of annotation.

[0021] That is, in the conventional technology, it is not possible to automatically perform annotation of content labels in real time during an interview between an interviewer and a subject during health guidance.

[0022] (Overview of Technology According to the Embodiment) The technology according to the present embodiment for solving the above problems will be described in detail below. The "dialogue" used in the present embodiment is assumed to be a dialogue for determining a behavioral goal. The party who determines their own behavioral goal in the dialogue is defined as the subject, and the party who supports the determination of the behavioral goal is defined as the interviewer. In a health consultation, the interviewer is, for example, a parental care worker, and the subject is, for example, a person who has been diagnosed as being at risk of metabolic syndrome in a health checkup.

[0023] The users of the learning device 100 and the inference device 200 described below are interviewees. The learning device 100 learns a prediction model for automatically assigning content labels to sentences spoken by the interviewee from dialogue data. The inference device 200 uses the trained prediction model trained by the learning device 100 to automatically assign content labels to sentences spoken by the interviewee from the dialogue data.

[0024] The following describes in detail the device configurations and device operations of learning device 100 and inference device 200. Note that in this embodiment, an example will be described in which learning device 100 and inference device 200 are separate devices, but this is not limiting, and learning device 100 and inference device 200 may be the same device. For example, learning device 100 may be equipped with a prediction unit 230, allowing learning device 100 to perform both learning and inference.

[0025] 1 shows an example of the configuration of the learning device 100 according to this embodiment. As shown in FIG. 1, the learning device 100 includes a preprocessing unit 110, a feature generation unit 120, a learning unit 130, and a prediction model storage unit 140.

[0026] Transcribed dialogue data (text data in this embodiment) and content label data as training data are input to the learning device 100 as "data input" shown in Figure 1. The dialogue data may also be called a dialogue log. The functions of each unit are as follows:

[0027] The preprocessing unit 110 performs data cleansing on the input dialogue data. For example, the data cleansing may include removing filler expressions. The preprocessing unit 110 also assigns speaker labels to the dialogue data. The speaker labels may be assigned at the time of "data input."

[0028] Furthermore, the preprocessing unit 110 may be provided outside the learning device 100. When the preprocessing unit 110 is provided outside the learning device 100, preprocessed dialogue data is input to the learning device 100.

[0029] The feature generation unit 120 generates text features from dialogue data (text data). The features to be generated are not limited to specific ones, but in this embodiment, the feature generation unit 120 acquires vectors by, for example, text embedding processing. The feature generation unit 120 may also acquire importance within the text data as a feature by, for example, applying tf-idf.

[0030] The feature generating unit 120 may also assign a label indicating the position of the target utterance sentence in the interview to the dialogue data. Furthermore, the feature generating unit 120 may generate semantic information and syntactic information from the utterance sentence as features.

[0031] The learning unit 130 receives as input the features of the utterance sentence generated by the feature generating unit 120 and the correct content label (teaching data) corresponding to the utterance sentence, and learns a "prediction model that receives as input the features corresponding to the utterance sentence and outputs a prediction result of the content label corresponding to the utterance sentence." The prediction model is a machine learning model. The machine learning model may be any machine learning model, such as LightGBM or a neural network.

[0032] The prediction model storage unit 140 stores the parameters, hyperparameters, and prediction model of the prediction model learned by the learning unit 130. The parameters, hyperparameters, and prediction model are stored, for example, in a pickle-format file. Specifically, the parameters, hyperparameters, and prediction model are stored in binary format. If the prediction model is a neural network, the weight parameters are also stored in binary format.

[0033] When LightGBM is used as the prediction model, an example of hyperparameters stored in the prediction model storage unit 140 is shown in Fig. 2. In the following description, hyperparameters are also defined as part of the parameters.

[0034] The parameters of the trained prediction model and the prediction model stored in the prediction model storage unit 140 are read out and stored in the prediction model storage unit 240 of the inference device 200. Alternatively, the prediction model storage unit 140 and the prediction model storage unit 240 may be the same storage unit.

[0035] (Configuration Example of Inference Apparatus 200) Fig. 3 shows a configuration example of the inference apparatus 200 according to this embodiment. As shown in Fig. 3, the inference apparatus 200 includes a preprocessing unit 210, a feature generating unit 220, a prediction unit 230, and a prediction model storage unit 240.

[0036] Dialogue data (text data) for which a content label is to be predicted is input to the inference device 200 as "data input" shown in FIG.

[0037] The preprocessing unit 210 and the feature generation unit 220 have the same functions as the preprocessing unit 110 and the feature generation unit 120 described above. Note that the preprocessing unit 210 may be provided outside the inference device 200. When the preprocessing unit 210 is provided outside the inference device 200, preprocessed dialogue data is input to the inference device 200.

[0038] The prediction model storage unit 240 stores the prediction model and parameters of the prediction model learned by the learning device 100. When LightGBM is used as the prediction model, examples of parameters stored in the prediction model storage unit 240 are as shown in FIG.

[0039] The prediction unit 230 reads out and stores a prediction model and its parameters from the prediction model storage unit 240. The prediction unit 230 receives as input the features of the utterance sentences in the dialogue data generated by the feature generation unit 220. The prediction unit 230 uses a prediction model (trained prediction model) in which the parameters read out from the prediction model storage unit 240 are set. The prediction unit 230 inputs the features of the utterance sentences into the trained prediction model, and obtains a content label of the utterance sentence as an output from the trained prediction model. The prediction unit 230 outputs the content label as a prediction result.

[0040] The prediction unit 230 may output multiple content labels corresponding to multiple spoken sentences at once.

[0041] (Operation of the Learning Device 100) An example of the operation of the learning device 100 will be described with reference to the flowchart in Fig. 4. Specific examples of dialogue data etc. will be described in the explanation of the operation of the inference device 200.

[0042] <S101 (Step 101)> In S101, dialogue data (a collection of utterances) and the correct content label for each utterance by the interviewee are input to the learning device 100. The correct content label may be called the objective variable, and the utterance corresponding to the content label may be called the explanatory variable.

[0043] In this embodiment, the content labels are expressed by applying one-hot encoding to six content labels. A content label that is assigned to the corresponding utterance sentence is represented by 1, and a content label that is not assigned is represented by 0. Note that the six content labels shown below are merely examples.

[0044] Examples of six content labels are as follows:

[0045] Building trust (Bt): utterances related to greetings, self-introductions, and empathy or interest in the subject's story; Information gathering (Ig): utterances asking about the subject's hobbies, interests, and lifestyle habits; Feedback on health check-up results (Fh): utterances explaining health check-up results; Motivation (Mo): utterances explaining the benefits of healthy behavior; Action planning (Ap): utterances proposing an action plan; Interview closing (Ic): utterances confirming whether the subject has any questions at the end of the interview. <S102> In S102, the preprocessing unit 110 extracts only the interviewee's spoken sentences from the dialogue data and performs data cleansing on the spoken sentences. Specifically, the preprocessing unit 110 removes speaker labels, fillers, and corrects typos and omissions. The preprocessing unit 110 sends the data-cleansed spoken sentences of the interviewee to the feature generation unit 120.

[0046] <S103> In S103, the feature generation unit 120 generates a vector as a feature for each sentence uttered by the interviewee. The feature generation unit 120 may also generate other features (e.g., semantic information, syntactic information) in addition to the vector. The feature generation unit 120 also assigns a label to the feature (vector) indicating the utterance position within the interview of the utterance sentence corresponding to the feature. The feature with the utterance position label assigned is sent to the learning unit 130. Note that the utterance position label may not be assigned.

[0047] <S104> In S104, the learning unit 130 inputs the spoken sentence (more specifically, its feature quantities) of the interviewee and learns the prediction model so as to output the correct content label.

[0048] <S105> In S105, the learning unit 130 stores the learned prediction model parameters and the prediction model in the prediction model storage unit 140.

[0049] (Operation of the inference device 200) Next, an example of the operation of the inference device 200 will be described with reference to the flowchart in Figure 5. Specific data examples will also be described here as appropriate. It is assumed that the prediction model storage unit 240 stores a trained prediction model and its parameters, and in S200, the prediction unit 230 acquires the trained prediction model and its parameters from the prediction model storage unit 240 and sets the parameters in the prediction model.

[0050] <S201> In S201, dialogue data (one or more spoken sentences) is input to the inference device 200. An example of the dialogue data to be input is as follows: "Subject: Good morning. Interviewer: My name is **, and I'm a public health nurse. Please excuse me for remaining seated. Subject: Nice to meet you. Interviewer: I will. Well, first of all, I have received my medical examination results. Subject: Hello."

[0051] <S202> In S202, the preprocessing unit 210 extracts only the spoken sentences of the interviewee from the input dialogue data and performs data cleansing on the spoken sentences. Specifically, the preprocessing unit 210 may delete speaker labels, delete fillers, correct typos, etc. The preprocessing unit 210 sends the spoken sentences of the interviewee that have been subjected to data cleansing to the feature generation unit 220.

[0052] For the specific example shown in S201, an example of dialogue data processed by the pre-processing unit 210 is as follows: "Good morning. My name is **, and I am a public health nurse. Please excuse me for remaining seated. First, I have received my medical examination results."

[0053] <S203> In S203, the feature generation unit 220 generates a vector as a feature for each sentence uttered by the interviewee. The feature generation unit 220 also assigns a label to the generated vector indicating the utterance position in the interview of the utterance sentence corresponding to the vector. The feature generation unit 220 may also generate other features (e.g., semantic information, syntactic information) in addition to the vector. The generated feature (vector) and utterance position label for each utterance sentence are input to the prediction unit 230. Note that an utterance position label may not be assigned.

[0054] The utterance position label may be any information that indicates a position within an interview, but for example, a decile within the interview may be used as the utterance position label. For example, suppose there are three utterance sentences (utterance sentence A, utterance sentence B, and utterance sentence C), and utterance sentence A is spoken 1 / 10 of the way through the interview, utterance sentence B is spoken 5 / 10 of the way through the interview, and utterance sentence C is spoken 7 / 10 of the way through the interview.

[0055] Furthermore, if the three vectors serving as features of (utterance sentence A, utterance sentence B, utterance sentence C) are (vector A, vector B, vector C) and the utterance position labels of (utterance sentence A, utterance sentence B, utterance sentence C) are represented as 0.1, 0.5, and 0.7, the three vectors assigned the utterance position labels are ((0.1, vector A), (0.5, vector B), and (0.7, vector C)).

[0056] <S204> In S204, the prediction unit 230 inputs the feature quantity for each utterance sentence into the trained prediction model and obtains a prediction result of the content label for each utterance sentence output from the prediction model. The prediction unit 230 outputs the prediction result. The output prediction result includes, for example, the utterance sentence of the interviewee and the content label of the utterance sentence.

[0057] An output image is shown in Fig. 6. In the example of Fig. 6, for example, it is shown that the content label assigned to the utterance sentence number 0 is "item_id_30".

[0058] (Hardware Configuration Example) Any of the devices described in this embodiment (learning device 100, inference device 200) can be realized, for example, by causing a computer to execute a program. This computer may be a physical computer or a virtual machine on the cloud.

[0059] That is, the device can be realized by executing a program corresponding to the processing performed by the device using hardware resources such as a CPU and memory built into a computer. The program can be recorded on a computer-readable recording medium (such as a portable memory) and stored or distributed. The program can also be provided via a network such as the Internet or email.

[0060] Fig. 7 is a diagram showing an example of the hardware configuration of the computer. The computer in Fig. 7 includes a drive device 1000, an auxiliary storage device 1002, a memory device 1003, a CPU 1004, an interface device 1005, a display device 1006, an input device 1007, an output device 1008, and the like, all of which are interconnected via a bus B. The computer may further include a GPU.

[0061] The program that realizes the processing on the computer is provided by a recording medium 1001, such as a CD-ROM or a memory card. When the recording medium 1001 storing the program is set in the drive device 1000, the program is installed from the recording medium 1001 to the auxiliary storage device 1002 via the drive device 1000. However, the program does not necessarily have to be installed from the recording medium 1001, but may be downloaded from another computer via a network. The auxiliary storage device 1002 stores the installed program as well as necessary files, data, etc.

[0062] The memory device 1003 reads and stores a program from the auxiliary storage device 1002 when an instruction to start the program is received. The CPU 1004 realizes functions related to the device in accordance with the program stored in the memory device 1003. The interface device 1005 is used as an interface for connecting to a network, etc. The display device 1006 displays a GUI (Graphical User Interface) or the like according to the program. The input device 1007 is composed of a keyboard, mouse, buttons, a touch panel, etc., and is used to input various operation instructions. The output device 1008 outputs the results of calculations.

[0063] (Summary, Effects, etc. of the Embodiments) As described above, the technology described in the present embodiment makes it possible to automatically assign content labels to sentences spoken by interviewees during interviews such as health guidance interviews. Furthermore, accuracy can be improved by adding the utterance position of the sentence to the features input to the prediction model.

[0064] The technology according to the present embodiment can realize a tool that supports the interviewer in reviewing how to proceed with the interview and how to convey information based on the content of the conversation, for example. Furthermore, by using the technology according to the present embodiment in real time, the interviewer can become aware of how to proceed with the conversation during the interview.

[0065] The following additional notes are provided regarding the above-described embodiments.

[0066] <Additional Notes> (Additional Note 1) A learning device that trains a prediction model for obtaining content labels for sentences uttered by an interviewer in an interview with a target person, comprising: a memory; and at least one processor connected to the memory, wherein the processor generates features of the utterance sentences extracted from dialogue data, and learns the prediction model using the features and training data of content labels for the utterance sentences as inputs. (Additional Note 2) An inference device that obtains content labels for sentences uttered by an interviewer in an interview with a target person, comprising: a memory; and at least one processor connected to the memory, wherein the processor generates features of the utterance sentences extracted from dialogue data, and inputs the features to a trained prediction model trained to output content labels for the utterance sentences, thereby obtaining content labels for the utterance sentences corresponding to the features as output from the prediction model. (Supplementary Item 3) A learning method executed by a learning device that learns a predictive model for obtaining content labels for sentences spoken by an interviewer in an interview conducted by the interviewer with a target person, the learning method comprising: a step of generating features of the utterance sentences extracted from dialogue data, and a step of learning the predictive model using the features and teacher data of content labels for the utterance sentences as inputs. (Supplementary Item 4) A non-transitory storage medium that stores a program to cause a computer used as a learning device that learns a predictive model for obtaining content labels for sentences spoken by an interviewer in an interview conducted by the interviewer with a target person to function as: a feature generation unit that generates features of the utterance sentences extracted from dialogue data, and a learning unit that learns the predictive model using the features and teacher data of content labels for the utterance sentences as inputs.

[0067] Although the present embodiment has been described above, the present invention is not limited to such a specific embodiment, and various modifications and changes are possible within the scope of the gist of the present invention described in the claims.

[0068] REFERENCE SIGNS LIST 100 Learning device 110 Preprocessing unit 120 Feature generation unit 130 Learning unit 140 Prediction model storage unit 200 Inference device 210 Preprocessing unit 220 Feature generation unit 230 Prediction unit 240 Prediction model storage unit 1000 Drive device 1001 Recording medium 1002 Auxiliary storage device 1003 Memory device 1004 CPU 1005 Interface device 1006 Display device 1007 Input device 1008 Output device

Claims

1. A learning device that trains a predictive model for obtaining content labels for sentences spoken by an interviewer in an interview with a target person, the learning device comprising: a feature generation unit that generates features for the sentences spoken extracted from dialogue data; and a learning unit that trains the predictive model using as input the features and training data for content labels for the sentences spoken.

2. An inference device that obtains content labels for sentences spoken by an interviewer in an interview with a subject, comprising: a feature generation unit that generates features for the sentences spoken extracted from dialogue data; and a prediction unit that inputs the features into a trained prediction model that has been trained to output content labels for sentences spoken, and thereby obtains content labels for the sentences spoken corresponding to the features as output from the prediction model.

3. A learning method executed by a learning device that learns a predictive model for obtaining content labels for sentences spoken by an interviewer in an interview with a target person, the learning method comprising: a step of generating features of the sentences spoken extracted from dialogue data; and a step of learning the predictive model using the features and training data of content labels for the sentences spoken as input.

4. A program for causing a computer used as a learning device to learn a predictive model for obtaining content labels for sentences spoken by an interviewer in an interview with a subject, to function as: a feature generation unit that generates features for the sentences spoken extracted from dialogue data; and a learning unit that learns the predictive model using as input the features and training data for content labels for the sentences spoken.

Images