JP2018004813A - Communication ability estimation device, method, and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a communication ability estimation device using a machine learning technique.SOLUTION: A communication ability estimation device according to the present invention comprises: means that extracts the amount of characteristics on speech from speech sound information of subjects participating in a group discussion; means that inputs the amount of characteristics on the speech of each subject and outputs a rating value of a communication ability of the subject for creation of teacher data; means that allows a learning machine to learn the created teacher data and creates an estimation model for estimating a communication ability; means that inputs the amount of characteristics on speech extracted from speech sound information of a subject to be rated by the means that extracts the amount of characteristics to the estimation model when the subject to be rated participates in the group discussion; and means that acquires an estimated value of a communication ability of the subject as an output of the estimation model.SELECTED DRAWING: Figure 1

Description

  The present invention relates to a technique for estimating communication ability.

  In recent years, in the field of artificial intelligence, social intelligence (Non-Patent Document 1), which is defined as "the ability of an action subject to build and use social relationships with others and solve problems through interaction", is modeled. Research is actively conducted (for example, Non-Patent Document 2).

  On the other hand, in the field of human resource development in recent years, it is an important issue to improve the communication ability of young people, and there is a strong demand for the establishment of an education and training base for that purpose.

Toyoaki Nishida, Yasuyuki Kado, Masahiro Matsumura: Social Intelligence Design, Knowledge Science, Ohmsha (2009)

Nishida, T .: Conversational informatics: an engineeringapproach, Vol. 9, Wiley.com (2008)

  The present invention has been made in view of the above, and an object of the present invention is to provide a communication ability estimation apparatus using a machine learning technique.

  As a result of intensive studies on the configuration of a communication ability estimation device using a machine learning technique, the present inventor has conceived the following configuration and has reached the present invention.

  That is, according to the present invention, an apparatus for estimating communication ability, a means for extracting a feature amount related to speech from speech speech information of each subject participating in a group discussion, and a feature amount related to the speech of each subject. Means for generating teacher data as an input and outputting the evaluation value of the subject's communication ability, means for generating an estimation model for estimating the communication ability by causing the learner to learn the generated teacher data, and an evaluation object When the subject who participates in the group discussion, the means for extracting the feature amount inputs the feature amount related to the utterance extracted from the utterance voice information of the subject to the estimation model, and the output of the estimation model Means for obtaining an estimated value of the communication ability of the subject. The means for extracting the feature value related to the utterance analyzes the utterance voice information to acquire language information, extracts a predetermined part of speech from the language information, and a first feature amount related to the number of appearances of the predetermined part of speech. A communication ability estimation device is provided that is characterized by extracting.

  As described above, according to the present invention, a communication ability estimation apparatus using a machine learning technique is provided.

The functional block diagram of the communication capability estimation apparatus of this embodiment. The flowchart of a teacher data generation process. The functional block diagram of the communication capability estimation apparatus of this embodiment. The flowchart of a communication ability estimation process. Bar graph showing the coefficient of determination R ² of the regression model generated in this embodiment. The bar graph which shows the correct answer rate of the classification model produced | generated by the present Example. The figure which shows the ROC curve of the classification model produced | generated in the present Example.

  Hereinafter, the present invention will be described with reference to embodiments shown in the drawings, but the present invention is not limited to the embodiments shown in the drawings. In the drawings referred to below, the same reference numerals are used for common elements, and the description thereof is omitted as appropriate.

  FIG. 1 shows a functional block diagram of a communication capability estimation apparatus 100 that is an embodiment of the present invention. Hereinafter, the functional configuration of the communication capability estimation apparatus 100 according to the present embodiment will be described with reference to FIG.

  As shown in FIG. 1, the communication ability estimation apparatus 100 of the present embodiment includes a corpus collection unit 101, a rating value input unit 102, a speech recognition unit 103, a morpheme analysis unit 104, an utterance speech analysis unit 105, The movement amount detection unit 106, the feature amount extraction unit 107, the teacher data generation unit 108, the learning execution unit 109, the learning device 110, the estimation execution unit 112, and the estimated value output unit 113 are configured. .

  The corpus collection unit 101 is means for collecting a corpus from subjects participating in a group discussion.

  The rating value input unit 102 is means for receiving an input of a test subject's rating value from an expert who evaluates communication ability.

  The voice recognition unit 103 is means for analyzing utterance voice information of a subject and acquiring language information (text data).

  The morpheme analysis unit 104 is a unit that performs morpheme analysis on the linguistic information acquired by the speech recognition unit 103 and extracts predetermined parts of speech.

  The utterance voice analysis unit 105 is a means for extracting the utterance by analyzing the utterance voice information of the subject.

  The motion amount detection unit 106 is a means for detecting the motion amount from the motion information of the subject.

  The feature amount extraction unit 107 is a feature related to the subject's communication based on the predetermined part of speech extracted by the morphological analysis unit 104, the utterance extracted by the utterance voice analysis unit 105, and the operation amount detected by the operation amount detection unit 106. It is a means for extracting the quantity.

  The teacher data generation unit 108 is a means for generating teacher data for generating an estimation model of communication ability.

  The learning execution unit 109 is a unit that causes the learning device to learn the teacher data generated by the teacher data generation unit 108.

  The learning device 110 is a unit that learns teacher data and generates an estimation model of communication ability.

  The estimation execution unit 112 is means for inputting the feature amount of the subject to be evaluated into the estimation model.

  The estimated value output unit 113 is means for acquiring an estimated value of the subject's communication ability as an output of the estimated model.

  In the present embodiment, the communication capacity estimation apparatus 100 functions as the above-described units when a computer constituting the communication capacity estimation apparatus 100 executes a predetermined program.

  The functional configuration of the communication ability estimation device 100 has been described above. Next, processing executed by the communication ability estimation device 100 will be described. The communication ability estimation apparatus 100 performs “learning mode” in which machine learning is performed to generate an estimation model of communication ability, and “estimation mode” in which the communication ability of the subject is estimated using the generated estimation model (learning device). It has. Here, the “learning mode” will be described first with reference to FIG.

  In the “learning mode”, the communication ability estimation apparatus 100 first collects corpora. The procedure for collecting the corpus will be described below.

  Collect multiple subjects and divide them into groups. Give each group a common task and conduct a group discussion. At this time, it is preferable that the members of the group meet each other for the first time, and it is preferable to give a plurality of tasks having different tendencies to each group.

  During the group discussion, each subject is equipped with a directional voice microphone to individually acquire the speech voice information of each subject, and an arbitrary motion sensor (for example, a three-axis acceleration sensor) is attached to each subject. It is worn on the body (preferably the head), and motion information (for example, acceleration data) of each subject is acquired. In the present embodiment, “motion information” means RAW data from which the amount of movement of the subject can be extracted, and the sensor output of a non-contact sensor (for example, an infrared sensor) is used as a motion as long as the purpose is met. It may be acquired as information, or an image of a subject taken by a digital video camera may be acquired as motion information.

  The speech voice information and motion information acquired from each subject are linked to the subject's identification information (hereinafter referred to as subject ID) by the corpus collection unit 101 and stored in the storage area 120.

  After the corpus collection is completed, a video image showing the state of the group discussion is viewed by a plurality of pre-selected graders, and each grader is graded for the communication ability of each subject. In the present embodiment, it is preferable to select an expert who has an interview experience at a company or the like as an evaluator.

  On the other hand, the rating value input unit 102 provides the evaluator with a predetermined UI for inputting the rating value. In response to this, each evaluator inputs the evaluation value of each subject to the communication ability estimation apparatus 100 via the provided UI. In response to this, the rating value input unit 102 stores the input rating value of each subject in association with the subject ID in the storage area 120.

  In response to the corpus (speech voice information, motion information) and evaluation values of each subject being stored together in the storage area 120, the communication ability estimation device 100 executes “teacher data generation processing”. Hereinafter, the “teacher data generation process” executed by the communication ability estimation apparatus 100 will be described based on the flowchart shown in FIG.

  First, in step 101, the speech recognition unit 103, the morpheme analysis unit 104, the utterance speech analysis unit 105, and the feature amount extraction unit 107 cooperate to extract “features related to utterance” from the speech speech information of each subject. In the present embodiment, three types of feature amount groups, “utterance turn (S)”, “prosody (A)”, and “language (L)”, are extracted as “features related to speech”. Table 1 below collectively shows the feature values belonging to each feature value group.

  Hereinafter, the feature quantities belonging to each feature quantity group shown in Table 1 and the calculation method thereof will be described in order.

<Speaking turn (S)>
The utterance turn (S) is a group of feature amounts related to the amount of utterance (time length / number) of the subject, and “total utterance length”, “total utterance count”, “total utterance length (1 second or more)”, It consists of four feature quantities, “total number of utterances (1 second or more)”. In the present embodiment, the utterance voice analysis unit 105 and the feature amount extraction unit 107 cooperate to calculate each feature amount according to the following procedure.

(Total utterance length)
The utterance voice analysis unit 105 divides utterance voice data of each subject acquired during a group discussion period (hereinafter referred to as one session) for one subject into predetermined minute time units (for example, 0.01 second units). The sound pressure in each section is analyzed, and a section in which the sound pressure is equal to or greater than a predetermined threshold is detected as a “speech section”. In response to this, the feature amount extraction unit 107 sets a value (N × 0.01) obtained by multiplying the number N of detected utterance sections by the length of the utterance section (for example, 0.01 seconds) as “total utterance length”.

(Total number of utterances)
The feature amount extraction unit 107 extracts one or more continuous utterance sections as “utterance fragments”, and sets the total number of extracted utterance fragments as “total number of utterances”.

(Total utterance length (more than 1 second))
The feature quantity extraction unit 107 extracts utterance fragments having a time length of 1 second or more from the extracted utterance fragments, and calculates the total time length of the extracted one or more utterance fragments as “total utterance length (1 Sec.) ”.

(Total number of utterances (1 second or more))
The feature amount extraction unit 107 sets the total number of extracted utterance fragments having a time length of 1 second or longer as “total number of utterances (1 second or longer)”.

<Prosody (A)>
The prosody (A) is a group of features related to the prosody of the subject's utterance, and is “maximum pitch”, “minimum pitch”, “pitch average”, “maximum intensity”, “minimum intensity”, “sound pressure”. It consists of eight feature quantities: “width”, “intonation”, and “speech speed”. In the present embodiment, the utterance voice analysis unit 105 and the feature amount extraction unit 107 cooperate to calculate each feature amount according to the following procedure.

(Maximum pitch)
The utterance voice analysis unit 105 detects the highest pitch from each utterance fragment detected from the utterance voice data of each subject recorded in one session. In response to this, the feature quantity extraction unit 107 sets the average value of the highest pitches detected from each utterance fragment as the “maximum pitch”.

(Minimum pitch)
The speech voice analysis unit 105 detects the lowest pitch from each detected speech fragment. In response to this, the feature quantity extraction unit 107 sets the average value of the lowest pitches detected from each utterance fragment as the “minimum pitch”.

(Pitch average)
The utterance voice analysis unit 105 divides each detected utterance fragment into a predetermined minute time unit (for example, 0.1 second unit) and detects the pitch of each section. In response to this, the feature quantity extraction unit 107 calculates the average value of the pitches detected from each section. Then, the average value of the pitch average values calculated from each utterance fragment is referred to as “pitch average”.

(Maximum intensity)
The utterance voice analysis unit 105 detects the maximum sound pressure from each utterance fragment detected by the feature amount extraction unit 107. In response to this, the feature amount extraction unit 107 sets the average value of the maximum sound pressure detected from each utterance fragment as “maximum intensity”.

(Minimum intensity)
The utterance voice analysis unit 105 detects the minimum sound pressure from each utterance fragment detected by the feature amount extraction unit 107. In response to this, the feature quantity extraction unit 107 sets the average value of the minimum sound pressure detected from each utterance fragment as the “minimum intensity”.

(Width of sound pressure)
The feature quantity extraction unit 107 sets the difference between the maximum intensity and the minimum intensity as the “sound pressure width”.

(intonation)
The feature amount extraction unit 107 sets the difference between the maximum pitch and the minimum pitch as “intonation”.

(Speaking speed)
The utterance voice analysis unit 105 detects syllables (syllables) included in all utterance fragments detected by the feature amount extraction unit 107. In response to this, the feature amount extraction unit 107 sets a value obtained by dividing the total number of detected syllables by the “total utterance length” as “speech speed”.

<Language (L)>
The language (L) is a group of feature quantities related to the number of appearances of a predetermined part of speech (“noun”, “verb”, “impression verb”, “filler”) appearing in the speech of the subject. It consists of seven feature quantities: “number of verbs”, “number of impressions”, “number of fillers”, “number of new nouns”, “number of existing nouns”, and “new noun frequency”. In the present embodiment, the voice recognition unit 103, the morpheme analysis unit 104, and the feature amount extraction unit 107 cooperate to calculate each feature amount according to the following procedure.

(Nouns, verbs, impressions, fillers)
The voice recognition unit 103 analyzes the speech voice data of each subject recorded in one session, and acquires language information (text data representing the content of the speech). In response to this, the morphological analysis unit 104 performs morphological analysis on the acquired linguistic information and extracts four types of parts of speech (“noun”, “verb”, “impression verb”, “filler”). . Here, “impression verb” is a part of speech that is often observed when interacting with or responding to the other's remarks, such as “Oh”, “I see”, “Ye”, and “Filler” ”Or“ something ”, which is a part of speech often observed when a speaker tries to fill a gap between utterances. In response to the extraction of four types of part of speech (“noun”, “verb”, “impression verb”, “filler”), the feature quantity extraction unit 107 calculates the total number of “nouns” and the total of “verbs”. The number, the total number of “impression verbs”, and the total number of “fillers” are defined as “noun number”, “verb number”, “impression verb number”, and “filler number”, respectively.

(Number of new nouns)
In one session, the feature amount extraction unit 107 detects “nouns” spoken for the first time by the subject in question before other subjects as “new nouns” of the subject, and the total number of detected “new nouns” is “ The number of new nouns.

(Number of existing nouns)
The feature quantity extraction unit 107 sets the difference between the “number of nouns” and the “number of new nouns” as the “number of existing nouns”.

(New noun frequency)
The feature amount extraction unit 107 sets a value obtained by dividing the “number of new nouns” by the “total number of utterances” as “new noun frequency”.

  The feature amounts belonging to each feature amount group shown in Table 1 and the calculation method thereof have been described above. In this embodiment, the feature amount extraction unit 107 associates each subject's “utterance turn (S)”, “prosody (A)”, and “language (L)” with the subject ID and stores them in the storage area 130. To do.

  Returning again to FIG. 2, the description will be continued.

  In the subsequent step 102, the motion amount detection unit 106 and the feature amount extraction unit 107 cooperate to extract “feature amount related to motion” from the motion information of each subject. In the present embodiment, a feature value group “motion (M)” is extracted as “feature value related to motion”. Table 2 below collectively shows the feature quantities belonging to “motion (M)”.

  In the following, the feature quantities belonging to “motion (M)” shown in Table 2 and the calculation method thereof will be described in order.

<Operation (M)>
The motion (M) is composed of four feature amounts: “average of motion amount”, “standard deviation of motion amount”, “average of motion amount during speech”, and “standard deviation of motion amount during speech”. In this experiment, the motion amount detection unit 106 and the feature amount extraction unit 107 cooperate to calculate each feature amount according to the following procedure. In the following, a case where acceleration data of a triaxial acceleration sensor is acquired as motion information will be described as an example.

(Average of movement amount, standard deviation of movement amount)
The motion amount detection unit 106 analyzes the acceleration data of each subject recorded in one session, and calculates the norm | at | of the three-dimensional vector: at = {xt, yt, zt} of acceleration at time t. In response to this, the feature amount extraction unit 107 sets the calculated | at | average and standard deviation as “average of motion amount” and “standard deviation of motion amount”, respectively.

(Average of movement during utterance, standard deviation of movement during utterance)
The motion amount detection unit 106 analyzes the acceleration data of the subject recorded in the period corresponding to each utterance fragment detected from the speech data of each subject recorded in one session, and the motion amount ( Norm | at |) is calculated. In response to this, the feature amount extraction unit 107 calculates the average and standard deviation of the motion amount (norm | at |) calculated for each utterance fragment, and then calculates the calculated motion amount (norm | at |). The average average value is “average motion amount (utterance unit)”, and the average standard deviation of the calculated motion amount (norm | at |) is “standard deviation of motion amount (utterance unit)”.

  Heretofore, the feature quantities belonging to “motion (M)” shown in Table 2 and the calculation method thereof have been described. In the present embodiment, the feature amount extraction unit 107 stores the extracted “motion (M)” of each subject in association with the subject ID in the storage area 130.

  Returning again to FIG. 2, the description will be continued.

  In the subsequent step 103, the teacher data generation unit 108 generates teacher data in which the characteristic amount of each subject is input (X) and the evaluation value of the communication ability of the participant is output (Y). Specifically, first, from the four types of feature groups (utterance turn (S), prosody (A), language (L), action (M)) associated with the subject ID of each subject, “language (L ) ”Including a plurality of feature groups (hereinafter referred to as feature amount sets) is read from the storage area 130. Here, the feature amount set is preferably composed of three or more feature groups including “language (L)”. Thereafter, a feature vector having each feature amount (real number) constituting the subject feature amount set as an input (X) is used, and a rating value (real number) associated with the subject ID of the subject is read from the storage area 120 and output. Based on (Y), a set of input (X) and output (Y) is generated as teacher data.

  Finally, in step 104, the learning execution unit 109 causes the learning device 110 to learn the teacher data generated in step 103. Thereby, the learning device 110 generates an estimation model for estimating the communication ability. As the learning device 110, a support vector machine (Support Vector Machine: SVM) or a support vector regression (Support Vector Regression: SVR) can be exemplified.

  The processing executed by the communication ability estimation device 100 in the “learning mode” has been described above. Next, the processing executed by the communication ability estimation device 100 in the “estimation mode” will be described. In the following description, the functional block diagram shown in FIG. 3 is referred to as appropriate.

  In the “estimation mode”, the subject to be evaluated is allowed to participate in the group discussion, and the corpus (utterance voice information, motion information) of the subject at that time is acquired. Specifically, in the same procedure as the above-described “learning mode”, the communication ability estimation apparatus 100 is configured to give a predetermined task to a group including a subject whose communication ability is to be evaluated and perform a group discussion. Utterance voice information and motion information are acquired from a voice microphone and a motion sensor attached to the subject. In response to this, the corpus collection unit 101 stores the utterance voice information and motion information acquired from the subject in the storage area 120 in association with the subject ID of the subject.

  In response to storing the utterance voice information and motion information of the subject in the storage area 120, the communication ability estimation device 100 executes “communication ability estimation processing”. Hereinafter, “communication ability estimation processing” executed by the communication ability estimation apparatus 100 will be described based on the flowchart shown in FIG.

  First, in step 201, the speech recognition unit 103, the morpheme analysis unit 104, the utterance speech analysis unit 105, and the feature amount extraction unit 107 cooperate with each other from the utterance speech information of the subject in the same procedure as the “learning mode” described above. “Speech turn (S)”, “prosody (A)”, and “language (L)” are extracted and associated with the subject ID of the subject and stored in the storage area 130.

  In the following step 202, the motion amount detection unit 106 and the feature amount extraction unit 107 cooperate to extract “motion (M)” from the motion information of the subject in the same procedure as the “learning mode” described above, and the subject The subject ID is stored in the storage area 130.

  Finally, in step 203, the estimation execution part 112 is a part of four types of feature groups (speech turn (S), prosody (A), language (L), action (M)) associated with the subject ID of the subject. Alternatively, a feature quantity set consisting of all and having the same combination as the feature quantity set used when generating the estimation model (learning device 110) is read from the storage area 130 and input to the estimation model. In response to this, the estimation model outputs an estimated value of the communication ability of the subject, and the estimated value output unit 113 acquires the estimated value. Thereafter, the estimated value output unit 113 stores the acquired estimated value in association with the subject ID of the subject in the storage area 140 and outputs it to a predetermined output destination in response to a request from the user or application.

  As described above, according to the present embodiment, a novel communication ability estimation device using a machine learning technique is provided. In addition, the communication ability estimation apparatus of this embodiment can be thought of as an apparatus that generates an estimation model of communication ability, and a device that generates teacher data for generating the estimation model. I can think of it. Moreover, according to this embodiment, the method of producing | generating the estimation model of communication ability is provided. The method for generating a communication ability estimation model according to the present embodiment can be considered as a method for generating teacher data for generating a communication ability estimation model.

  Note that each function of the above-described embodiment can be realized by a device-executable program described in C, C ++, C #, Java (registered trademark), and the like. It can be stored and distributed on a device-readable recording medium such as a CD-ROM, MO, DVD, flexible disk, EEPROM, EPROM, etc., and can be transmitted via a network in a format that other devices can.

  As described above, the present invention has been described with the embodiment. However, the present invention is not limited to the above-described embodiment, and the functions and effects of the present invention are within the scope of other embodiments that can be considered by those skilled in the art. As long as it plays, it is included in the scope of the present invention.

  Hereinafter, the method for generating an estimation model of communication ability according to the present invention will be described more specifically with reference to examples. However, the present invention is not limited to the examples described later.

(1) Collection of corpus 40 subjects (university students) are divided into 10 groups of 4 people in a group so that the members of the group face each other for the first time, and a group discussion is given to each group with common issues. I got it. In this experiment, in order to avoid over- / under-evaluation of each subject's communication ability due to the pros and cons of the task, the following three tasks (a) to (c) were set and decided: We had a discussion within the time given.
(A) Tasks to rank celebrities to call for school festivals (15 minutes)
(I) Challenges for making a store opening plan at a school festival (20 minutes)
(C) Making a plan to entertain foreign friends in Japan (20 minutes)

  In this experiment, each subject had a discussion with a headset microphone (audio-technica: HYP-190H) and an acceleration sensor (ATR-Promotions: WAA-010) attached to the back of the head. Taken with a video camera. During that time, utterance voice data was recorded via the headset microphone of each subject, and acceleration data in three axial directions (x, y, z) was acquired at 30 fps via the acceleration sensor of each subject.

(2) Assessment of subject's communication skills Twenty-one former HR recruiters with experience of recruitment interviews in companies of different industries (retail, personnel dispatch, IT-related, etc.) and scale (from SMEs to large companies) The subjects were evaluated for their communication skills. During the experiment, each evaluator was asked to watch a video containing the discussion of each group, and each subject was evaluated for the following six items.

  According to the Japan Vocational Ability Development Association (JAVADA), the communication skills that young people should acquire are “communication” that facilitates two-way communication and information transmission appropriately and accurately, and “collaboration” that adapts to the organization. It is said that it consists of three items, “self-expression” regarding presentation ability. Among them, the ability of “communication”, where individual differences are most noticeable in group discussions, is “attentive listening”, “two-way smoothness” It is said to consist of five element items: “communication”, “opinion gathering ability”, “information transmission ability”, and “logical and clear assertion”. In response to this, in this experiment, each grader was asked to evaluate the above-mentioned five element items in five stages, and an item of “comprehensive communication ability”, which is a comprehensive evaluation in consideration of the five element items. We had you evaluate in ten steps. Table 3 below summarizes the six evaluation items and their contents.

  When the evaluation values of 21 evaluators were obtained for a total of 120 subjects (40 people x 3 tasks), and the clone reliability α reliability coefficient of each evaluation item was calculated, the value of the evaluation item (0) Only (α = 0.66) was below the reliability standard of 0.8, but the values of the other evaluation items (1) to (5) were all above 0.8. Accordingly, in this experiment, only the results of the evaluation items (1) to (5) are adopted, and the average value of the evaluation values answered by 21 evaluators for each evaluation item is determined as the evaluation value related to the evaluation item. did.

(3) Extraction of feature amount The speech data of each subject recorded in one session is analyzed using speech analysis software Praat, and “utterance turn (S)”, “prosody (A)” described above, “Language (L)” was extracted. Similarly, the above-mentioned “motion (M)” was extracted from the acceleration data of each subject recorded in one session.

(4) Generation of Communication Capability Estimation Model 15 types of feature amount sets were prepared by combining four feature amount groups (S, A, M, L) extracted from the corpus of each subject. Table 4 below summarizes the contents of the prepared feature quantity sets (1) to (15).

  In this experiment, machine learning (regression learning / classification learning) was performed by the following procedure using each feature amount set shown in Table 4 below, and an estimation model for estimating communication ability was generated.

(Regression learning)
A feature vector having each feature quantity (real number) constituting each feature quantity set shown in Table 4 as an element is input (X), and a rating value (for each rating item) of the subject is output (Y). As described above, teacher data composed of a pair of input / output (X) and output (Y) is generated, and regression learning is performed using linear support vector regression (SVR) based on the generated teacher data. Note that ε tolerance is used for the error function in SVR, and the search range of [0, 0.01, 0.1, 1] is used for ε for adjusting error tolerance, and the trade-off between loss function and margin size For parameter C that adjusts, a search range of [0.01, 0.1, 1, 5, 10] was provided, and the optimal parameter was determined by grid search through cross-validation.

(Classification learning)
For the teacher data described above, the average value m and standard deviation σ of the rating values (output (Y)) are calculated, the evaluation values of m + 0.1σ or higher are classified into high groups, and the evaluation values of m−0.1σ or lower are calculated. After classifying into the low group, the evaluation value (output (Y)) is converted into a high and low binary label, and the evaluation value near the average value that is not classified into either the high group or the low group is output (Y). Group to be excluded from the teacher data. Based on the teacher data with the above modifications, classification learning was performed using a linear support vector machine (SVM). A parameter C that adjusts the trade-off between loss and margin size in SVM was searched in the range of [0.01, 0.1, 1, 5, 10] and used for testing.

(4) Evaluation of regression / classification model In this experiment, the estimated performance of each model was evaluated by cross-validation. Specifically, out of 10 groups that participated in the experiment, after generating a model by machine learning using teacher data of each subject belonging to 9 groups, the teacher data of each subject belonging to the remaining 1 group is The estimation performance of the generated model was evaluated using test data.

(Estimated performance evaluation of regression model)
In the evaluation of estimation performance of the regression model was the decision for the test data coefficient R ² as an index. Figure 5 shows a bar graph of the coefficient of determination R ² for each 15 kinds of characteristic quantity set according to each of the evaluation items shown in Table 3 (1) to (5). In the bar graph shown in FIG. 5, the number assigned to each bar means the number of the feature amount set shown in Table 4 (hereinafter the same as in FIG. 6).

  As shown in FIG. 5, in each of the evaluation items (1) to (4), the regression model generated using the feature amount set of “utterance turn (S) + motion (M)” is the maximum value ( 0.47, 0.54, 0.56, 0.52). On the other hand, in “evaluation item (5): comprehensive communication ability”, the regression model generated using the feature amount set of “utterance turn (S) + language (L) + motion (M)” is the maximum value. (0.62).

(Estimated performance evaluation of classification model)
In evaluating the estimation performance of the classification model, the accuracy rate (Accuracy) was used as an index. FIG. 6 shows a bar graph of the correct answer rate for each of 15 types of feature quantity sets according to each of the evaluation items (1) to (5) shown in Table 3.

  As shown in FIG. 6, in both “evaluation item (1): two-way smooth communication” and “evaluation item (2): opinion gathering power”, both “prosody (A) + language (L) + operation” The classification model generated using the feature quantity set of (M) ”showed the maximum value (0.85, 0.83). Further, in “evaluation item (3): information transmission ability” and “evaluation item (4): logical and clear assertion”, both “prosody (A) + utterance turn (S) + language (L)” The classification model generated using the feature quantity set showed the maximum value (0.87, 0.85). On the other hand, in “evaluation item (5): comprehensive communication ability”, the classification model generated using the feature quantity set of “prosody (A) + language (L) + motion (M)” is the maximum value ( 0.93).

(Robustness evaluation of classification model)
FIG. 7 is a graph showing an estimation result of “Evaluation item (5): Comprehensive communication ability” based on a classification model generated using different feature quantity sets (A, S, L, M, A + L + M). The figure which plotted the curve is shown. In FIG. 7, the vertical axis indicates the true positive rate (True Positive (TP): the rate of correctly classifying low group data), and the horizontal axis indicates the false positive rate (FalsePositive (FP): low group data incorrectly high). The ratio of classification into groups). As shown in FIG. 7, the AUC (Area Under the Curve) of the classification model using the feature value set of “prosody (A) + language (L) + motion (M)” is 0.97, and A, S, L, M The AUC of the model using each was improved, and the TP was 0.91 or more at FP = 0.05. As a result, the classification model generated using the feature quantity set of “prosody (A) + language (L) + motion (M)” can maintain the low classification accuracy of the high group data while maintaining the low group accuracy. It shows that the data of can be correctly classified at a rate of 91% or more.

DESCRIPTION OF SYMBOLS 100 ... Communication ability estimation apparatus 101 ... Corpus collection part 102 ... Rating value input part 103 ... Speech recognition part 104 ... Morphological analysis part 105 ... Speech speech analysis part 106 ... Motion amount detection part 107 ... Feature quantity extraction part 108 ... Teacher data generation Unit 109 ... Learning execution unit 110 ... Learner 112 ... Estimation execution unit 113 ... Estimated value output unit 120, 130, 140 ... Storage area

Claims (10)

A method for generating teacher data for generating an estimation model of communication ability,
Extracting a feature amount related to the utterance of the subject from the utterance voice information of each subject participating in the group discussion;
Generating teacher data having as input a feature amount related to the utterance of each subject and outputting a rating value of the subject's communication ability;
Including
The step of extracting a feature amount related to the utterance includes:
Analyzing the speech information, obtaining linguistic information, extracting a predetermined part of speech from the linguistic information, and extracting a first feature quantity relating to the number of appearances of the predetermined part of speech;
Method. The step of extracting a feature amount related to the utterance includes:
Analyzing the utterance voice information, extracting a subject's utterance, and extracting a second feature amount relating to the amount of the utterance;
The method of claim 1. The step of extracting a feature amount related to the utterance includes:
Analyzing the utterance voice information, extracting a subject's utterance, and extracting a third feature amount related to the prosody of the utterance;
The method according to claim 1 or 2. Furthermore, obtaining the motion information of each subject,
Extracting a feature amount related to a motion amount from the motion information of each subject,
The step of generating the teacher data includes:
Generating teacher data with the feature amount related to the utterance and the feature amount related to the motion amount of each subject as inputs, and output as a rating value of the subject's communication ability,
The method as described in any one of Claims 1-3. A method for generating an estimated model of communication ability,
Generating teacher data by the method according to any one of claims 1 to 4,
Learning the generated teacher data with a learning device to generate the estimation model;
Including methods. The estimation model is a regression model or a classification model;
The method of claim 5.   The program for making a computer perform each step of the method as described in any one of Claims 1-6. An apparatus for generating teacher data for generating an estimation model of communication ability,
Means for extracting feature values related to utterance from utterance voice information of each subject participating in the group discussion;
Means for generating teacher data having as input a feature amount related to the utterance of each subject and outputting a rating value of the subject's communication ability;
Including
The means for extracting the feature amount related to the utterance is:
Analyzing the speech information, obtaining linguistic information, extracting a predetermined part of speech from the linguistic information, and extracting a first feature quantity relating to the number of appearances of the predetermined part of speech;
apparatus. An apparatus for generating an estimation model of communication ability,
Means for extracting feature values related to utterance from utterance voice information of each subject participating in the group discussion;
Means for generating teacher data having as input a feature amount related to the utterance of each subject and outputting a rating value of the subject's communication ability;
Means for causing the learning device to learn the teacher data and generating the estimation model;
Including
The means for extracting the feature amount related to the utterance is:
Analyzing the speech information, obtaining linguistic information, extracting a predetermined part of speech from the linguistic information, and extracting a first feature quantity relating to the number of appearances of the predetermined part of speech;
apparatus. A device for estimating communication ability,
Means for extracting feature values related to utterance from utterance voice information of each subject participating in the group discussion;
Means for generating teacher data having as input a feature amount related to the utterance of each subject and outputting a rating value of the subject's communication ability;
Means for generating an estimation model for causing the learning device to learn the generated teacher data and estimating communication ability;
Means for inputting, to the estimation model, a feature amount related to an utterance extracted by the means for extracting the feature amount from the utterance voice information of the subject when the subject to be evaluated participates in the group discussion;
Means for obtaining an estimate of the communication ability of the subject as an output of the estimation model;
Including
The means for extracting the feature amount related to the utterance is:
Analyzing the speech information, obtaining linguistic information, extracting a predetermined part of speech from the linguistic information, and extracting a first feature quantity relating to the number of appearances of the predetermined part of speech;
Communication ability estimation device.