JP6783454B2 - Interesting quantitative evaluation device, interesting quantitative evaluation method, interesting adjustment method, and program - Google Patents

Description

Embodiments of the present invention relate to an interesting quantitative evaluation device, an interesting quantitative evaluation method, an interesting adjustment method, and a program.

In performances such as Manzai and Tale, exchanges aimed at making the audience laugh (interactions that produce laughter) are repeatedly performed as oral performances by the performers. Manzai usually consists of the interaction of two performers. Tale is organized as a skit. The forms and contents of the material elements that cause laughter in oral performances are various, and include not only those based on dialogue but also those based on gestures and movements (Non-Patent Document 1).

Yoshito Hirose, "Correlation between Manzai Components and" Laughter ": Blurring, Tsukkomi, Facial Expressions, Pretending, etc.", Nihon University National Literature Society, Language, No.146, pp.75-65, 2013

However, as mentioned above, there are various forms and contents of material elements in oral performances such as comics and comics. Therefore, it was difficult to quantitatively evaluate the fun of oral performance.

An object to be solved by the present invention is to more easily provide an fun quantitative evaluation device, a fun quantitative evaluation method, a fun adjustment method, and a program for quantitatively evaluating the fun of oral performance.

(1) In order to achieve the above object, the fun quantitative evaluation device according to one aspect of the present invention is the first analysis result showing the tendency of the performer's speaking style derived based on the voice emitted by the performer during the performance. A second analysis result showing the tendency of the content played by the performer derived based on the laughter of the viewer of the performance, and an evaluation unit for evaluating the fun of the performance performed by the performer based on the results. wherein the the first analysis result, Ru contain the results of the evaluation of the tempo of a plurality of human interaction is the performer.
(2) Further, the first analysis result in the above-mentioned fun quantitative evaluation device does not include the analysis result regarding the utterance content of the performer based on the voice uttered by the performer , and the second analysis result includes the analysis result. Includes the results of an evaluation of the proportion of non-verbal material elements in the performance .
(3) Further, the fun quantitative evaluation device according to one aspect of the present invention provides the first analysis result showing the tendency of the performer's speaking style derived based on the voice emitted by the performer during the performance, and the performance. The second analysis is provided with a second analysis result showing the tendency of the content performed by the performer derived based on the laughter of the viewer and an evaluation unit for evaluating the fun of the performance performed by the performer based on the second analysis result. The result includes information on the time interval in which the performer played the material element in the performance , or information on the time from when the performer played the material element in the latter half of the performance to the end of the performance.
( 4 ) Further, in the above-mentioned fun quantitative evaluation device, the material element played by the performer in the latter half of the performance is the material element played at the end of the performance.
( 5 ) Further, the fun quantitative evaluation device according to one aspect of the present invention provides the first analysis result showing the tendency of the performer's speaking style derived based on the voice emitted by the performer during the performance, and the performance. The first analysis is provided with a second analysis result showing a tendency of the content performed by the performer derived based on the laughter of the viewer, and an evaluation unit for evaluating the fun of the performance performed by the performer based on the result. The result includes information indicating the amount of information of the voice emitted by the performer.
( 6 ) Further, the amount of information of the voice in the above-mentioned fun quantitative evaluation device includes the time allotted to the dialogue, the number of mora assigned to the unit time, and the dialogue of the performer in the voice uttered by the performer. Among them, one or more of the number of mora of relatively long dialogue is included .
(7 ) Further, the evaluation unit in the above-mentioned fun quantitative evaluation device is the first in the performance performed by the performer in the first of the predetermined times of the contest in which the contest is divided into a plurality of times and selected stepwise. The first evaluation value of the fun of the performance performed by the performer based on the 1 analysis result and the second analysis result, and the performer in the second time different from the first time in the predetermined times. Based on the first analysis result and the second analysis result in the performance performed by the performer, the second evaluation value of the fun of the performance performed by the performer is derived, and the fun quantitative evaluation device is described. An estimation unit for estimating the fun of the performance performed by the performer after both the first and second evaluations based on the first evaluation value and the second evaluation value is further provided.
( 8 ) Further, the fun quantitative evaluation device according to one aspect of the present invention provides the first analysis result showing the tendency of the performer's speaking style derived based on the voice emitted by the performer during the performance, and the performance. The evaluation unit is provided with a second analysis result showing the tendency of the content performed by the performer derived based on the laughter of the viewer, and an evaluation unit for evaluating the fun of the performance performed by the performer based on the result. , The performer plays based on the first analysis result and the second analysis result in the performance performed by the performer in the first of the predetermined times of the contest which is divided into a plurality of times and selected stepwise. The first evaluation value of the fun of the performance, the first analysis result and the second analysis result in the performance performed by the performer in the second time different from the first time in the predetermined time. Based on the above, a second evaluation value of the fun of the performance performed by the performer is derived, and the fun quantitative evaluation device uses the first evaluation value, the second evaluation value, and the first evaluation. An estimation unit that estimates the fun of the performance performed by the performer after both the first and second times is further based on the value and the coefficient that weights the second evaluation value. Be prepared.
(9) Further, the above-mentioned fun quantitative evaluation device includes a model generation unit that generates a model for evaluating the fun of the play based on the feature amount of the play whose fun is evaluated and the evaluation result. In preparation, the feature quantity of the performance evaluated for the fun includes the voice emitted by the performer who performed the performance evaluated for the fun and the laughter of the person who watched the performance evaluated for the fun. included.
( 10 ) Further, the fun quantitative evaluation method according to one aspect of the present invention includes a first analysis result showing a tendency of the performer's speaking style derived by a computer based on the voice emitted by the performer during the performance. only contains a second analysis results showing the trend of what the performers, which is derived based on the people of laughter to watch the repertoire is played, the process of evaluating the interest of the performers play repertoire on the basis of, the The first analysis result includes the result of evaluation of the tempo of the interaction of a plurality of persons who are the performers, or the information indicating the amount of information of the voice emitted by the performer, or the second analysis result includes the result of evaluation. , Information on the time interval in which the performer played the material element in the performance, or information on the time from when the performer played the material element in the latter half of the performance to the end of the performance .
( 11 ) Further, the program according to one aspect of the present invention includes a first analysis result showing a tendency of the performer's speaking style derived based on the voice emitted by the performer during the performance, and a person who watches the performance. The computer of the quantification evaluation device is made to perform a step of evaluating the fun of the performance performed by the performer based on the second analysis result showing the tendency of the content played by the performer derived based on the laughter. The first analysis result includes information indicating the result of evaluation of the tempo of the interaction of a plurality of persons who are the performers, or the amount of information of the voice emitted by the performer, or The second analysis result includes information on the time interval in which the performer played the material element in the program, or the time from when the performer played the material element in the latter half of the program to the end of the program. It is a program that contains the information of .
(12) Further, the fun quantitative evaluation device according to one aspect of the present invention provides the first analysis result showing the tendency of the performer's speaking style derived based on the voice emitted by the performer during the performance, and the performance. To generate an evaluation value as a result of evaluating the fun of the performance performed by the performer based on the second analysis result showing the tendency of the content performed by the performer derived based on the laughter of the viewer. A formed evaluation unit is provided, and the composition of the script of the performance is adjusted so that the evaluation value becomes a desired value.
(13) Further, the fun quantitative evaluation device according to one aspect of the present invention provides the first analysis result showing the tendency of the performer's speaking style derived based on the voice emitted by the performer during the performance, and the performance. To generate an evaluation value as a result of evaluating the fun of the performance performed by the performer based on the second analysis result showing the tendency of the content performed by the performer derived based on the laughter of the viewer. The formed evaluation unit is provided, and the composition of the program is adjusted so that the evaluation value becomes a desired value, thereby supporting the creation of the script of the program.
(14) Further, the fun quantitative evaluation device according to one aspect of the present invention provides the first analysis result showing the tendency of the performer's speaking style derived based on the voice emitted by the performer during the performance, and the performance. To generate an evaluation value as a result of evaluating the fun of the performance performed by the performer based on the second analysis result showing the tendency of the content performed by the performer derived based on the laughter of the viewer. It is provided with a formed evaluation unit, and the script of the performance is quantitatively evaluated based on the evaluation value.
(15) Further, the fun adjustment method according to one aspect of the present invention is an fun adjustment method for the computer to support the adjustment of the fun of the performance, and the computer makes a sound produced by the performer during the performance. The first analysis result showing the tendency of the performer's speaking style derived based on the above, and the second analysis result showing the tendency of the content played by the performer derived based on the laughter of the viewer of the performance. The process of evaluating the fun of the first performance performed by the performer based on the above, and the process of the computer evaluating the fun of the second performance after the fun is adjusted based on the result of the evaluation. Including.
(16) Further, the fun adjustment method according to one aspect of the present invention includes a first analysis result showing a tendency of the performer's speaking style derived based on the voice emitted by the performer during the performance, and viewing the performance. A computer adjusted so that the fun of the performance performed by the performer can be evaluated based on the second analysis result showing the tendency of the content performed by the performer derived based on the laughter of the performer is the target performance. Includes a process of evaluating and supporting the creation of a script for the performance so that the desired performance can be obtained from the performance.

According to one aspect of the present invention, it is possible to more easily provide an fun quantitative evaluation device, a fun quantitative evaluation method, a fun adjustment method, and a program for quantitatively evaluating the fun of oral performance.

It is a figure which shows the outline of the fun quantitative evaluation apparatus 1 of 1st Embodiment. It is a figure which shows the hardware structure of the fun quantitative evaluation apparatus 1 of this embodiment. It is a figure which shows the structure of the fun quantitative evaluation apparatus 1 of this embodiment. It is a figure which shows an example of the detailed structure of the fun quantitative evaluation apparatus 1 of this embodiment. It is a figure which shows an example of the interaction between performer A and performer B. It is a figure which shows the time from the start to the end of a comic book, and the position of the detected material element. This is the number of participating groups in the final and final finals of each Manzai contest. It is a figure which shows the number of the deleted part in the video recorded on the DVD of the comic dialogue contest. It is a block diagram of the neural network 241 which concerns on this embodiment. It is a figure which shows the learning data which concerns on this embodiment. It is a figure which shows the evaluation value which concerns on this embodiment. It is a flowchart which shows the verification process of the fun quantitative evaluation apparatus 1 which concerns on this embodiment. It is a figure which shows the prediction accuracy of the final game of MZ -2001-MZ -2010. It is a figure which shows the prediction accuracy of the final decisive battle of MZ -2001-MZ -2010. It is a figure which shows the degree of influence on the prediction accuracy of 6 indexes p1 to p6. It is a figure which shows the structure of the fun quantitative evaluation apparatus 1 of the 2nd Embodiment. It is a figure which shows the progress of M1-2015 which is an example of a comic dialogue contest. It is a figure which shows the outline of the ranking prediction processing of the final game of M1-2015 which is an example of a comic contest. It is a figure which shows the result of the prediction of the ranking of the final game which concerns on this embodiment. It is a block diagram of the fun quantitative evaluation apparatus 1 including the analysis apparatus of this embodiment. It is a figure which shows an example of the index obtained by the feature amount acquisition part 26 which concerns on this embodiment.

Hereinafter, with reference to the drawings, an interesting quantitative evaluation device, an interesting quantitative evaluation method, and an embodiment of the program will be described.

(First Embodiment)
[1. Configuration of quantitative evaluation device for fun]
FIG. 1 is a diagram showing an outline of the fun quantitative evaluation device 1 of the present embodiment.
The fun quantitative evaluation device 1 is derived from the first analysis result showing the tendency of the performer's speaking style derived based on the voice emitted by the performer during the performance, and the performer derived based on the laughter of the viewer of the performance. Based on the second analysis result showing the tendency of the content performed by the performer, the fun of the performance performed by the performer is evaluated.

The fun quantitative evaluation device 1 makes it possible to quantitatively evaluate the oral performance by using an index related to the content of the oral performance in a comic dialogue or a contest and an index related to the utterance. Further, the fun quantitative evaluation device 1 can derive a quantitative evaluation result for the oral performance, and the evaluation result can be compared with the evaluation result by other means. Therefore, the fun quantitative evaluation device 1 further enhances the accuracy of the evaluation by the fun quantitative evaluation device 1 by a machine learning method that utilizes the evaluation results by other means.

In the following description in the present embodiment, a manzai will be illustrated and described as an evaluation target. The fun quantitative evaluation device 1 quantitatively evaluates the fun based on the video data 2 of the comic book. For example, the fun quantitative evaluation device 1 uses the video of the comic contest as the video data 2 of the comic. The fun quantitative evaluation device 1 utilizes the evaluation result by the examiner of the above-mentioned comic dialogue contest as the evaluation result by other means for machine learning.

By the way, there are various types of manzai, "showing one story", "showing multiple developments in a specific situation", "starting from one topic and developing to similar topics", In addition to these, there are many other forms.

In evaluating the fun of such comics, the fun quantitative evaluation device 1 in the present embodiment evaluates the fun of comics by using an index that does not depend on a specific form of comics. For example, the fun quantitative evaluation device 1 of the present embodiment does not include an index showing the connection relationship before and after the material elements constituting the comic story such as the flow of the story and the continuity of the topic depending on the utterance content. May be good. The connection relationship before and after the material elements that make up the comic is related to the content of each material element, the meaning and degree of fun of each material element, and the semantic connection between the material elements. The fun quantitative evaluation device 1 exemplifies a case where the connection relationship before and after each individual material element as described above is not used as an evaluation index.

In addition, the fun quantitative evaluation device 1 is used to create a specific material element that brings laughter to the listener, a pattern of tsukkomi for blurring, and how to take the time between speeches. It is not necessary to include the factor of fun due to the relationship (relationship between the material elements) in the evaluation index. That is, the fun quantitative evaluation device 1 in the present embodiment does not use the information related to the time series of the oral presentation and the linguistic information as the index for evaluating the fun.

For example, the fun quantitative evaluation device 1 may predict the ranking of the comic contest based on the quantitative evaluation that does not depend on the above information. The fun quantitative evaluation device 1 regards comics as a chain of exchanges that cause the audience to laugh once (interactions that produce laughter). Here, the exchange that causes laughter once is defined as a "material element". Regardless of the form and content of the material element, it includes those based on dialogue, gestures and movements.

FIG. 2 is a diagram showing a hardware configuration of the fun quantitative evaluation device 1 of the present embodiment. The fun quantitative evaluation device 1 includes a CPU 1A, a ROM (Read Only Memory), an EEPROM (Electrically Erasable and Programmable Read Only Memory), a non-volatile storage device 1B of an HDD (Hard Disk Drive), and a RAM (Random Access Memory). A computer including a volatile storage device 1C such as a register, a portable recording medium drive device 1D, an input / output device 1E, a communication interface 1F, and the like, and a process for evaluating the fun of a comic speech by a program to be executed. To execute.

FIG. 3A is a diagram showing the configuration of the fun quantitative evaluation device 1 of the present embodiment. FIG. 3B is a diagram showing an example of a detailed configuration of the fun quantitative evaluation device 1 of the present embodiment. The fun quantitative evaluation device 1 includes a storage unit 10, an evaluation unit 21, an average score calculation unit 22, an order determination unit 23, a model generation unit 24, a feature quantity extraction unit 25, and a prediction accuracy calculation unit 26. To be equipped.

The storage unit 10 is realized by the non-volatile storage device 1B and the volatile storage device 1C. The storage unit 10 stores data such as learning data 11, model data 12, evaluation data 13, evaluation result data 14, examination result data 15, and the like, in addition to the program for operating the fun quantitative evaluation device 1. .. Details of the learning data 11, the model data 12, the evaluation data 13, the evaluation result data 14, and the examination result data 15 will be described later.

For example, each unit such as the evaluation unit 21, the average score calculation unit 22, the order determination unit 23, the model generation unit 24, the feature amount extraction unit 25, and the prediction accuracy calculation unit 26 stores a program stored in the storage unit 10. It functions when the CPU 1A executes it.

The evaluation unit 21 evaluates the fun of the performance performed by the performer of the group performing the performance (hereinafter, simply referred to as the performer) based on the evaluation data 13 read from the storage unit 10, and evaluates the evaluation value as the evaluation result. (Prediction evaluation point k (k = 1, ..., 10)) is output. The rating data 13 includes the first analysis result 131 and the second analysis result 132. The first analysis result 131 is a result of analyzing the tendency of the performer's speaking style based on the voice emitted by the performer during the performance. The second analysis result 132 is a result of analyzing the tendency of the content performed by the performer based on the laughter of the viewer of the program. For example, the evaluation unit 21 performs an evaluation process for evaluating the fun of the performance performed by the performer based on the evaluation model generated by the model generation unit 24 by the machine learning method, and performs the above evaluation value (predictive evaluation). The point k) is derived. The details of the process for evaluating the fun of the performance performed by the performer will be described later.

The average score calculation unit 22 calculates the average of the evaluation values (predicted evaluation points k) for each performer of each performance derived by the evaluation unit 21.

The order determination unit 23 determines the order of the fun of the oral performances of the target performers by the determination based on the average of the evaluation values for each group (performer) of each performance derived by the evaluation unit 21. Then, the evaluation result data 14 is generated. The fun quantitative evaluation device 1 may predict the fun of the oral performance to be performed in the future based on the order of the fun of the performers determined by the order determination unit 23.

The prediction accuracy calculation unit 26 compares the order of fun of the group (performer) determined by the order determination unit 23 (evaluation result data 14) with the order of fun of the performer indicated by the examination result data 15. , Calculate the prediction accuracy.

The model generation unit 24 generates an evaluation model to function as the evaluation unit 21 in advance, and generates model data 12. For example, the model generation unit 24 generates the above evaluation model based on a machine learning method. Hereinafter, as an example of the model generation unit 24, a case where a neural network is applied as an example of a machine learning method will be described. For example, the model generation unit 24 includes a neural network 241 and a learning processing unit 242. The learning processing unit 242 controls the preliminary learning of the neural network 241. The evaluation model is generated by the neural network 241 that has completed the preliminary learning. The model generation unit 24 stores the generated evaluation model as model data 12 in the storage unit 10. The above evaluation model includes numerical information including the weight of the connection between layers and the bias given to the node of each layer when the neural network is formulated. In this evaluation model, the above numerical information is determined by preliminary learning.

The feature amount extraction unit 25 detects each of the voice emitted by the performer and the laughter of the viewer so as to be distinguishable from each other. For example, the feature amount extraction unit 25 detects the sound collected by the microphone arranged in the vicinity of the performer as the sound emitted by the performer. The feature amount extraction unit 25 identifies the number of characters constituting the voice from the voice emitted by the performer, and distinguishes between the vocalized period and the non-vocalized period.
The feature amount extraction unit 25 extracts the feature amount of the tendency of the performer's speech based on the number of characters identified from the voice emitted by the performer, the vocalization period, and the non-vocalization period, and the analysis result (first analysis) in the storage unit 10. Result) is written. A detailed explanation of the analysis of the features of the speaker's speaking tendency will be described later.

The feature amount extraction unit 25 detects the laughter of the viewer from the sound collected on the audience side. When the sound emitted by the performer and the laughter of the viewer are mixed, the sound emitted by the performer and the laughter of the viewer are separated from the mixed sound based on the characteristics of the sound. .. The feature amount extraction unit 25 analyzes the tendency of the content played by the performer based on the laughter of the separated viewer, and writes the analysis result (second analysis result) in the storage unit 10. A detailed explanation of the analysis of the tendency of the content performed by the performer will be described later.

The feature amount extraction unit 25 extracts the examination result by the examiner and writes the result in the storage unit 10.

[2. Method of quantitative evaluation of fun]
A method of quantitative evaluation of fun by the fun quantitative evaluation device 1 will be described. The fun quantitative evaluation device 1 quantifies the manzai oral performance in terms of two aspects: the performer's utterance and the content of the manzai. In addition, in this embodiment, an application example of quantitative evaluation is illustrated in which the value of one manzai oral performance is measured for each index, and the oral performance is not divided based on the time series as in the case of the front, middle, and rear stages. ..

[2.1 Evaluation index for comics]
The fun quantitative evaluation device 1 captures the manzai oral performance from two aspects: (A) utterance and (B) the content of the manzai.
The fun quantitative evaluation device 1 measures the following three characteristics for utterance. The three characteristics of utterance are (A-1) p1: 1 time that can be assigned to dialogue, (A-2) p2: 1 second (unit time), the number of mora that can be assigned to utterance, and (A). -3) p3: The number of mora in the longest line. The above is an example of the amount of voice information in the voice emitted by the performer.

On the other hand, the fun quantitative evaluation device 1 handles the material element, which is a factor that causes the audience to laugh, and measures the following three characteristics for the content of the comic dialogue. The three characteristics of the content of the comics are (B-1) p4: the average time interval of the material elements, (B-2) p5: the time from the last material element to the end of the comics, and (B-3). p6: The ratio of the frequency of appearance of the material elements that are only movements and have no dialogue to all the material elements.

[2.1.1 Utterance index]
Manzai is usually spoken by two performers. Movements such as gestures and gestures are also means of communicating the content of comics to the audience, but utterances are the main means. The fun quantitative evaluation device 1 quantifies the characteristics related to utterance. The utterance-related indexes p1 to p3 are calculated using the character data obtained by extracting or transcribing the dialogue from the video data 2. The feature quantity extraction unit 25 of the fun quantitative evaluation device 1 derives indexes p1 to p3 related to utterance based on the number of characters identified from the voice uttered by the performer, the utterance period, and the non-utterance period.

In this embodiment, one line is defined as one turn until the utterance in the conversation analysis moves to the next performer. FIG. 4 is a diagram showing an example of interaction between performer A and performer B. For example, the number of dialogues in the exchange in FIG. 4 is 5. In order to deal with the characteristics related to utterance, the number of mora that can obtain a value closer to the amount of utterance of the line is measured, not the number of characters of the transcribed line. The number of mora is basically the same as the number of hiragana, but youon is not counted. For example, "Gakufu", "Sea otter", and "Rae" are all 3 mora, but "Ryohi" is 2 mora. The fun quantitative evaluation device 1 converts all the character data extracted based on the performer's voice into hiragana for reading, and measures the number of mora for each line. The fun quantitative evaluation device 1 may convert all the character data extracted or transcribed based on the performer's voice into reading hiragana and measure it as the number of mora for each line.

By the way, the speed of the exchange is also related to the evaluation of the manzai, as there are expressions such as "slightly friendly exchange" and "slightly pleasant tempo". The fun quantitative evaluation device 1 includes p1 (time that can be allocated to one line) in the first analysis result 131. In addition, the fun quantitative evaluation device 1 includes p2 (the number of mora that can be assigned to one second of utterance) in the first analysis result 131 as an index related to the utterance speed. Further, the fun quantitative evaluation device 1 includes p3 (the number of mora of the longest dialogue) in the first analysis result 131 as another index regarding the interaction of the performers. This is because when the monopoly of utterance occurs, it affects the tempo of the exchange. There may be a manzai with a high proportion of long lines in all lines, but one or more relatively long lines or the longest line is extracted as a sample and its length is used as an index.

For example, the index p1 is calculated based on the equation (1).

For example, the index p2 is calculated based on the equation (2).

As shown in the above equations (1) and (2), the indexes p1 and p2 are different from the actual utterance speed values. In the formulas (1) and (2), the treatment of the oral presentation time is reversed between the numerator and the denominator. In the case of dialogue, the time per dialogue, and in the case of the number of mora related to speech speed, the number of mora per second can be grasped more intuitively, and it is applied to the practice of comic dialogue. By making it easier. The feature amount extraction unit 25 may derive the indexes p1 and p2 based on the above equations (1) and (2).

[2.1.2 Indicators related to the content of comics]
The content of the manzai has a great influence on the evaluation of the manzai oral performance. In this embodiment, the material element is treated as a factor that causes the audience to laugh. For example, it is defined that the larger the number of material elements, the higher the evaluation of the comic dialogue performance, and the greater the influence of the material elements on the evaluation. The fun quantitative evaluation device 1 measures p4 (average time interval of material elements).

On the other hand, it is said that it is good to bring excitement in the latter half of the oral performance, and the impression and evaluation of the oral performance depends on the matter immediately before the end. Therefore, it is assumed that the position of the last material element in the oral performance affects the evaluation of the comic dialogue. The fun quantitative evaluation device 1 measures p5 (the time from the last material element to the end of the comic dialogue).

As described above, p1 (time that can be assigned to one line) and p2 (the number of mora that can be assigned to one second of utterance) do not represent the actual utterance speed. The fun quantitative evaluation device 1 measures p6 (ratio of material elements without dialogue) in order to complement p1 and p2.

The case where the actual speech speed is used as the evaluation item is used as a comparative example, and is compared with the case of the present embodiment. In the case of the above comparative example, it is easy for the performer to change the utterance speed in the oral performance. The total time of the oral performance changes as the speech speed changes, and as a result, other indicators such as p1 (time that can be assigned to one dialogue), p4 (average time interval of the material element), and p5 (last material element). (Time from to the end of Manzai) is affected. P6 (ratio of material elements without dialogue), which is used as an index for evaluation of the present embodiment, has a specification independent of the index related to utterance, and has an advantage that it can be easily applied to an actual comic dialogue.

The feature amount extraction unit 25 measures the indexes p4 to p6 relating to the content of the comic dialogue by detecting the time point at which the laughter of the audience occurs in the video data 2 of the comic dialogue. The point to measure the laughter of the audience is only at the time of occurrence, and it is not necessary to measure the magnitude and duration of the laughter. For example, in the above-mentioned exchange shown in FIG. 4, laughter occurs at two places, and the feature amount extraction unit 25 determines that there are two material elements. If the audience laughs, the feature amount extraction unit 25 treats it as a material element regardless of the content. It is possible that the audience may laugh at a place not intended by the creator, or conversely, laughter may not occur at a place intended by the creator, but the author of the script cannot confirm which is the case. The fun quantitative evaluation device 1 handles the above two without distinguishing between them.

FIG. 5 is a diagram showing the time from the start to the end of the comic dialogue and the position of the detected material element. In the figure, t [i] and (i = 1 to N) indicate when the audience laughs. That is, t [i] is the generation time of the material element. N is the number of detected material elements. In the figure, t [0] is the start time of the comic book. Further, t [N + 1] is the end time of the comic dialogue, and its value differs for each group.

For example, the average time interval (p4) of the material element is calculated by the equation (3).

In the equation (3), N is the number of detected material elements.

In addition, the time (p5) from the last material element to the end of the comic dialogue is calculated by the equation (4).

[2.2 Target data]
The target data in this embodiment is the data in the comic contest.
As the target data in the present embodiment, a case where more specific data in the comic dialogue contest is applied will be illustrated, and the effect will be described. As an example of the Mansai Contest, the broadcast program "M-1 Grand Prix (registered trademark)" that is televised in Japan is known (for details, see http://www.m-1gp.com). The first M-1 Grand Prix was held in 2001, and since then, it has been held once a year until 2010, and the tenth held in 2010 was temporarily suspended.

The "M-1 Grand Prix" held 10 times is carried out according to the same rules, although the progress method differs depending on the year. For example, the "M-1 Grand Prix" will go through four stages (five stages only in the 10th round) qualifying rounds (1st round, 2nd round, 3rd round, quarterfinals (only 10th round), semifinals) and then to the finals. Nine groups to advance (10 groups only for the first time) are selected. From the second round onward, one group will be selected from the semi-final or quarter-final defeated groups among the groups that advance to the finals.

Participating groups are groups within 10 years of formation, and the number reaches 1,000 groups each time. For example, 1,603 groups participate in the 1st M-1 Grand Prix 2001, and 4,835 groups participate in the 10th M-1 Grand Prix 2010. Among these, it can be judged that the ability of the group that can participate in the final match is high.

In the finals, seven judges will evaluate the performance of each group's oral performance and score each group's oral performance. Each group is scored by each examiner according to a scoring standard with a maximum of 100 points, and is evaluated by the total score. Only the groups with the highest total score (the top two in the first round) can get the right to advance to the final battle. In the final final, it is required to show a different manzai from the final. The points of the spectators are added only in the first final match.
The ranking of the final and final battles of the M-1 Grand Prix will be determined by the evaluation of the judges. The existence of evaluation items for examination and the judgment criteria for each judge are unknown.

The fun quantitative evaluation device 1 analyzes data on the final and final battles of a comic contest held a plurality of times, for example, such as the "M-1 Grand Prix". The feature amount extraction unit 25 is a manzai contest held multiple times, from the 1st manzai contest (hereinafter MZ-2001) to the 10th manzai contest (hereinafter MZ-2010). May be targeted. For example, the fun quantitative evaluation device 1 analyzes the video data 2 of the comic contests held a plurality of times, and extracts, for example, the values of the indexes p1 to p6. The fun quantitative evaluation device 1 predicts the ranking of each holding time by the leave-one-out method using the data of each holding time as one set, and measures the prediction accuracy.

The fun quantitative evaluation device 1 is used in the 1st Manzai Contest (MZ-2001) to the 10th Manzai Contest (MZ-2010), for example, the 1st M-1 Grand Prix 2001 (M1-2001). ) To the 10th M-1 Grand Prix 2010 (hereinafter M1-2010) may be applied. For example, commercially available video data is used for the data of M1-2001 to M1-2010. As commercially available video data, data from the DVD of M-1 Grand Prix the FINAL PREMIUM COLLECTION 2001-2010 (Yoshimoto R & C Co., Ltd.) may be used.

FIG. 6 is a diagram showing an example of the number of participating groups in the finals and finals of each of the Manzai contests. It should be noted that commercially available video data and the like include parts that have been deleted due to copyright, expression, and the like. FIG. 7 is a diagram showing an example of the number of deleted parts in the video recorded on the DVD of the Manzai contest. The lines of the deleted part are inevitably excluded from the analysis. Further, the video data to be analyzed is shorter than the actual time by the deletion time, but the time of the deleted video is unknown. Of the data in the group shown in the figure, the data used as the learning data includes a "learning data ID" (see FIG. 9 described later). The group name is a tentative name.

[2.3 Calculation of evaluation value of Manzai oral performance]
A model for calculating the evaluation value of Manzai's oral performance is constructed using a feedforward multi-layer neural network. The neural network 241 in the model generation unit 24 is an example of a feedforward type multi-layer neural network. For example, see Geoffrey E. Hinton, "Learning multiple layers of representation", Trends in Cognitive Sciences, Vol. 11, pp. 428-34, 2007. for more information.

FIG. 8 is a configuration diagram of the neural network 241 according to the present embodiment. As shown in the figure, the number of hidden layers of the neural network 241 is 3, and the number of nodes of each hidden layer is 10. The learning processing unit 242 trains the neural network 241 with a teacher, and reduces the evaluation error by using a stochastic gradient descent method using backpropagation (error backpropagation method) in the learning processing. Optimize the characteristics to improve the accuracy of the evaluation. For example, the neural network 241 may be configured using the H2O package, in which case the training parameters may use the default values of the package. For more information on the H2O package, see http://0xdata.com.

The number of nodes in the input layer of the neural network 241 is 6, and six indexes p1 to p6 extracted from the oral data of the comic dialogue are input to each of the nodes as parameters. The number of nodes in the output layer is 1, and the above-mentioned nodes output an evaluation value which is an evaluation result. As described above, the number of nodes in each hidden layer is larger than the number of nodes in the input layer and the number of nodes in the output layer.

FIG. 9 is a diagram showing learning data according to the present embodiment. The learning processing unit 242 learns and verifies the neural network 241 by the leave-one-out cross-validation method. In the leave-one-out cross-validation method, only one case is extracted from the sample group and used as test data (test case), and the rest is used as learning data (training case). The learning processing unit 242 converts all the cases into test data once, and repeatedly executes each verification. This is applied to the learning and verification of the neural network 241 of the present embodiment.

For the learning of the neural network 241, the data of typical cases of the finals of each held time is used. As shown in FIG. 9, the learning processing unit 242 extracts the top two groups, the bottom two groups, and the middle one group from the finals of each of the MZ-2001 to MZ-2010 times, and uses them as the learning data for each time. Therefore, the learning processing unit 242 uses about half of the data of the data set of the final match of each held time at the time of learning. Five out of ten MZ-2001, five out of eight MZ-2003 and MZ-2010, and five out of nine other than those are adopted as training data.

The final match data for each of the MZ-2001 to MZ-2010 will be set as one set, and the test data will be for all groups participating in the final and final finals of each event. For example, when MZ-2001 is the test data, the training data is the top 2 groups, the middle 1 group and the bottom 2 groups, 5 groups per set, and 45 groups in each final of MZ-2002 to MZ-2010. , 10 groups of the final match of MZ-2001 and 2 groups of the final match are test data. Nine groups have participated in the finals of MZ-2003 and MZ-2010, but due to copyright issues, groups not recorded on the DVD will not be analyzed, and both will be analyzed by eight groups.

As shown in FIG. 9 above, there are groups that are used multiple times for learning. For example, "PGB (tentative name)" is a group in which both X2006-05 and X2007-05 are used as learning data with lower ranks. On the other hand, there are also groups such as "KRN (tentative name)" that are used as all types of learning data, such as upper (X2005-02), middle (X2001-03), and lower (X2003-04). Even in the same group, the content of the oral presentation is different each time, and there is no duplication as learning data. Since the purpose of learning in the neural network is not group identification or feature extraction, the selected data is suitable as training data.

The learning processing unit 242 uses a standardized value for the evaluation value (examination result) included in the learning data 11 and used at the time of learning, not the total score of the scores voted by the judges in the actual final match. .. FIG. 10 is a diagram showing evaluation values according to the present embodiment. Reasons for not directly using the score by the examiner in this way include the fact that the examiner's examination items and the value of points added for each item are not specified, and that the objectivity of the evaluation criteria is unclear. In addition, it is difficult to directly compare the scores of different holding times because the values vary depending on the holding time. For example, the MZ-2002 finals 1st place group ("FTB-RAW- (tentative name)") scored 621 points, but the MZ-2005 finals 5th place group ("C-TRAR ("C-TRAR) " Tentative name) ”) is lower than the score (622 points).

For the above reasons, in this embodiment, the evaluation value of the oral performance of the group used for learning is not the total score of the scores cast by the judges, but the evaluation value (relative value) based on the ranking of each holding time shown in FIG. ) Is used. For example, the upper group is given 100 points, the middle group is given 50 points, and the lower group is given 0 points. As long as the evaluation value is based on the ranking, at least groups having the same ranking can be treated equally even if the groups are held at different times. On the other hand, even if the oral performance of the first group in one holding time is equivalent to the oral performance of the middle group in another holding time, different evaluation values may be given. Both the evaluation value based on the ranking and the evaluation value based on the examination points have learning defects, but the present embodiment is expected to have the effect of reducing the influence of the variation in each holding time. Then, the evaluation value based on the ranking is used.

The stochastic gradient descent method used for learning the neural network 241 has an initial value dependence. The model generation unit 24 of the present embodiment generates a plurality of models in order to eliminate the influence of the initial value dependence of the stochastic gradient descent method. For example, the neural network 241 of the model generation unit 24 performs learning based on the same data 10 times using random number seeds generated by the learning processing unit 242 and having different values, and 10 models having different characteristics are obtained. To construct.

These 10 models function as the evaluation unit 21. In each of the models (21-k) and (k = 1, ..., 10), the rating data 13 (test data) of the group is input for each group. The rating data 13 input to each of the models (21-k) is the same for each group. Each model (21-k) outputs the evaluation value Pg-k of the group g for each group. The evaluation value Pg-k indicates a predicted evaluation point k derived by the model (21-k) based on the evaluation data 13 of the group g.
The average point calculation unit 22 calculates the average value Pgy of 10 evaluation values (evaluation value Pg-k) y output from each model (21-k) for each group. The above relationship is shown in Equation (5). The configuration shown in the figure shows the processing of a specific holding time, and is prepared for each holding time corresponding to the evaluation data to be processed.

In the above formula (5), y refers to the test data of MZ-2001 to MZ-2010, 2001 means MZ-2001, 2002 means MZ-2002, and so on.
The order determination unit 23 determines the prediction order _Ry (y = 2001, ..., 2010) of each test data based on the average value Pgy. The order determination unit 23 assigns the prediction order R _y in descending order of the average value Pgy of the evaluation values output by the “average point calculation unit of the group g”.

The prediction accuracy calculation unit 26 compares the predicted rank R _y with the actual rank RR _y (y = 2001, ..., 2010) and calculates the rank correlation coefficient. The prediction accuracy calculation unit 26 sets this rank correlation coefficient as the prediction accuracy Sy (y = 2001, ..., 2010) of the test data.

The prediction accuracy S for all the test data of MZ-2001 to MZ-2010 is shown in Equation (6). The order determination unit 23 averages the rank prediction accuracy Sy of the model for each test data, and sets the prediction accuracy S for all the test data of MZ-2001 to MZ-2010.

The fun quantitative evaluation device 1 prepares two types of evaluation data (test data) used for the verification for each of 10 sets (Y sets) of data sets corresponding to each time of MZ-2001 to MZ-2010. One is the data of the final match, and the fun quantitative evaluation device 1 evaluates the prediction accuracy for the data by Spearman's rank correlation coefficient. The other is the data of the final battle, and the fun quantitative evaluation device 1 evaluates the prediction accuracy for the data based on whether or not the first-ranked group (the winning group of the holding times) can be predicted.

FIG. 11 is a flowchart showing a verification process of the fun quantitative evaluation device 1. For example, a case where the test data is MZ-2001 will be illustrated and described in detail.

First, the learning processing unit 242 executes learning 10 times (M times) using a total of 45 comic dialogue data of MZ-2002 to MZ-2010 shown in FIG. 9 as learning data 11 (S10), 10 Generate (M) models (S15). As a result, 10 (M) independent models are generated.

Next, the data of the groups of 10 groups (Group G) who participated in the final match of MZ-2001 (for example, the final match of MZ-2001) is input to each model of the evaluation unit 21, and the evaluation unit 21 is used for each group. The evaluation value is derived for each model (S20).
After the fun quantitative evaluation device 1 finishes processing all the models in the same group, the average score calculation unit 23 derives the average value of the evaluation values derived by each model in the same group (S25).
The fun quantitative evaluation device 1 determines whether or not the processing of all the groups is completed (S30), and repeats the processing from S20 until the processing of all the models of the group is completed.
After finishing the processing of all the groups, the order determination unit 23 derives the rank Ryk based on the average score of the evaluation values of each group (S35), and derives the rank of 10 groups (G group).

Next, the prediction accuracy calculation unit 26 derives the rank correlation coefficient (Spearman) between the rank derived in S35 and the actual rank (S40). The prediction accuracy calculation unit 26 uses the derived rank correlation coefficient as the prediction accuracy of the MZ-2001 final match.

Next, the data of the two groups (Group D, D <G) who participated in the final battle of MZ-2001 (MZ-2001 final battle) was used to predict the final battle of 10 (M) models (M). Input to the evaluation unit 21), the evaluation unit 21 derives 10 (M) evaluation values for each group (S45), and the mean score calculation unit 23 derives the evaluation values derived from each model in the same group. The average value is derived (S50). Next, the order determination unit 23 selects 10 (Y) groups having the highest prediction based on the average value of the evaluation values for each group (S55).

The fun quantitative evaluation device 1 repeats the processing from S45 until the processing of all the models is completed (S60). After finishing the processing of all the models, the prediction accuracy calculation unit 26 derives the ratio of the group of the predicted 1st place selected in S50 to the group that actually won the 1st place (S65), and determines the result as the final battle. The prediction accuracy of. The fun quantitative evaluation device 1 executes the above processing for each holding time targeted for evaluation.

In addition, the judging method of the comic contest to be analyzed may differ depending on the number of times it is held, and there are times when the actual ranking is unknown. In such a case, the examination method may be adjusted within the range that can be unified. For example, in M1-2001 shown as an example, a total of 300 spectators' points by 100 spectators at the Yoshimoto Kogyo theaters in Sapporo, Osaka, and Fukuoka will be added to the judges' points only in the final match, and the ranking will be decided. Has been done. Since M1-2002 and later, the score was changed only by the judges, so in order to unify the evaluation criteria for ranking, the ranking based only on the judges' scores will be used in the analysis of the final match of M1-2001. Therefore, in the final match of M1-2001, the ranking of the 5 groups is different from the actual ranking. On the other hand, the ranking of the final battle depends only on the judges, so the actual ranking was used. If the ranking determined only by the judges' scores is used, the data of the target group may not exist. If the data of the target group does not exist, the data of the group having a rank close to the rank of the target group may be used. For example, when adjusting the ranking as described above, even if the group that did not participate in the final battle is used, the group that does not have the oral data of the final battle is excluded from the analysis.

Further, when the evaluation points by the judges are the same, the rankings of these two groups may be used for the analysis based on the number of the highest points for each judge.

Also, in the final battle, the actual ranking may be unclear. Nine groups (10 groups only for M1-2001) will participate in the final, and the judges will give each group a maximum of 100 points. The total score of all judges will be the score of the group, and the top three groups will advance to the final battle. In the final battle, each judge will vote in one group, and the group with the most votes will win. Therefore, in the final decisive battle, the number of votes of the two groups other than the winning group may be the same including zero, and the second and third places may not be decided. For example, in the final decisive battle of M1-2009, all judges voted for the same group, so the 1st place is clear, but the 2nd and 3rd place groups are unknown. Therefore, for the final battle, the accuracy of the proposed method will be measured based on whether the first group can be predicted.

[3. result]
[3.1 Prediction accuracy of Manzai contest ranking]
FIG. 12 is a diagram showing the prediction accuracy of the final match of MZ-2001 to MZ-2010. The figure shows an example of the rank correlation coefficient of the predicted ranking of the test data MZ-2001 to MZ-2010. This rank correlation coefficient is derived from each of the 10 models in which the same test data is input. The prediction accuracy can be determined based on the rank correlation coefficient. The average value at the bottom is the average value of all test data values. The specific numerical values are illustrated when the test data M1-2001 to M1-2010 are used as the test data MZ-2001 to MZ-2010. Refer to FIG. 6 for the number of participating groups in that case.

The average prediction accuracy based on all the test data of M1-2001 to M1-2010 was 0.58 in the ranking correlation coefficient of the final match. The value (0.58) of the ranking correlation coefficient of the final match indicates that the predicted ranking of the final match is correlated with the actual ranking. The accuracy of the final match is Spearman's rank correlation coefficient, and correlated results were obtained for seven test data.
As shown in the figure, test data with 6 strong correlations (MZ-2001, MZ-2002, MZ-2006, MZ-2007, MZ-2008, MZ-2009) and test data with 1 weak correlation (MZ-2001, MZ-2002, MZ-2006, MZ-2007, MZ-2008, MZ-2009). It is MZ-2005). The remaining three test data (MZ-2003, MZ-2004, MZ-2010) are uncorrelated, and the prediction accuracy (0.17) of MZ-2004 is the lowest among them.

FIG. 13 is a diagram showing the prediction accuracy of the final battle of MZ -2001 to MZ -2010. The figure shows an example of the prediction accuracy of the first place in the final battle of test data MZ -2001 to MZ -2010. This prediction accuracy is the average value of the prediction accuracy derived from each of the 10 models in which the same test data is input. In the "High accuracy" column, as a comparative example, the number of times that a result with higher accuracy than the case where the first place is randomly predicted is obtained, and when it is judged to be high accuracy, "GOOD" is described. There is. The average at the bottom is the average of all test data values. The specific numerical values are illustrated when the test data M1-2001 to M1-2010 are used as the test data MZ-2001 to MZ-2010. Refer to FIG. 6 for the number of participating groups in that case. See FIG. 6 for the number of participating groups.

When the test data MZ-2001 to MZ-2010 were used for all the test data of M1-2001 to M1-2010, the average prediction accuracy was 0.68 for the first place prediction accuracy in the final battle. Since the average prediction accuracy (0.68) of the final battle is about twice the prediction accuracy (0.35) when the first place is randomly predicted as a comparative example, the first place group can be predicted. Is shown. The prediction accuracy when the first place of the comparative example is randomly predicted is 0.5 for M1-2001 and 0.33 for the others, and the average value of M1-2001 to M1-2010 is 0.35 ( = (0.5 + 9x0.33) / 10).

Of the 10 test data of M1-2001 to M1-2010 corresponding to MZ-2001 to MZ-2010, 8 test data are higher than the random prediction accuracy value illustrated as a comparative example. It has become. Furthermore, in the four test data (M1-2004, M1-2005, M1-2006, M1-2010), the first place is predicted with a 100% probability. The accuracy of the two test data (M1-2003 and M1-2007) is slightly lower. On the other hand, the prediction of the first place failed in the two test data (M1-2008 and M1-2009), and the prediction accuracy in the case of random prediction illustrated as a comparative example is 0.33 (= 1/3). ) Is lower than.

The effectiveness of the fun quantitative evaluation device 1 was verified using the data of the M-1 Grand Prix. The criteria for manzai evaluation by the judges of the M-1 Grand Prix are unknown. The evaluation criteria of the fun quantitative evaluation device 1 are not consistent with the above-mentioned criteria for manzai evaluation by the judges.

The fun quantitative evaluation device 1 evaluates 10 test data of M1-2001 to M1-2010 corresponding to MZ-2001 to MZ-2010 according to the evaluation criteria set independently, and as a result, the judge evaluates the comics. We derived an evaluation result that correlates with the result of.

[3.2 Indicators p1 to p6]
The proposed method uses six indexes, p1 to p6, for the evaluation of manzai oral performance. The contribution rate of each index to the prediction accuracy was investigated. Since a neural network is used for ranking prediction, it is difficult to accurately grasp the contribution rate of each index, but by examining the change in accuracy when forecasting excluding individual indexes, each index Estimate the degree of impact.

FIG. 14 is a diagram showing the degree of influence of the six indexes p1 to p6 on the prediction accuracy. The figure shows the average ranking prediction accuracy of the finals of M1-2001 to M1-2010 when one of the six indexes p1 to p6 is removed. The figure shows the average of the rank correlation coefficients 10 times each time by the leave-one-out method as in the case where all the indexes of p1 to p6 are used. The "difference from the case of using p1 to p6" is the difference from the value (0.58) of the average prediction accuracy of the final match when all the indexes of p1 to p6 shown in FIG. 12 are used. ..

As shown in the figure, even if any of the six indexes p1 to p6 is removed, the ranking prediction accuracy is lower than when all the indexes are used. From the size of the value whose prediction accuracy has decreased, it can be seen that the index p4 (average time interval of the material elements) has the greatest influence among the indexes. Below, index p1 (time that can be assigned to one line), index p2 (number of mora that can be assigned to one second utterance), index p6 (ratio of material elements without dialogue), and index p5 (from the last material element). Time until the end of Manzai) and so on. The influence of the index p3 (the number of mora of the longest line) is the lowest. Since the index p4 is obtained by measuring the density of the material elements per hour, it can be said that the number of material elements has a great influence on the evaluation of the oral performance. The index p1 is an index related to the tempo of the performer's interaction, and it can be seen that the "feeling of interaction" and the interval between the material elements have the same importance. Other indicators are important, but have a lower impact. From the figure, it can be seen that the speed of utterance (index p2) is less important than the speed of interaction between performers (index p1). In addition, the ratio of the material element without dialogue (index p6) has the same effect as the speed of utterance (index p2), and further has a greater effect than the index p5 (time from the last material element to the end of comic dialogue). Having is mentioned as a feature.

As described above, it is estimated that the contribution rate of the time interval (p4) of the material element is the highest among the six indexes. Here, it is assumed that the number of laughter is equivalent to the number of material elements contained in the script. In other words, the fun of comics is related to the number of laughter, and it is hypothesized that the greater the number of laughter that occurs in the audience, the more interesting comics are. In order to verify this, the relationship between the rank based on the value of the number of laughter per hour (the reciprocal of p4) and the actual rank is shown based on the rank correlation coefficient.

Figure 12 above shows the ranking correlation coefficient between the ranking based on the number of laughter per hour (laughing density: 1 = p4) in the finals of M1-2001 to M1-2010 and the actual ranking, and the proposed method. The rank correlation coefficient between the rank and the actual rank is shown. As shown in the figure, the average value of the rank correlation coefficient of the final match in M1-2001 to M1-2010 is 0.34, which is an uncorrelated value, and the average of the rank correlation coefficient predicted by the proposed method. Lower than the value (0.58).

Furthermore, in the case of M1-2010, it was (-0.24), which was a negative value and was uncorrelated. These are contrary to the hypothesis that the more laughter you have, the more interesting it is. From these facts, it can be seen that the number and density of laughter cannot be used as indicators by themselves in evaluating the fun of comics. Furthermore, it can be seen from the above results that it is not enough to simply add a lot of material elements when creating a script for a comic book.

According to the above embodiment, the first analysis result showing the tendency of the performer's speaking style derived based on the voice emitted by the performer during the performance and the performer derived based on the laughter of the viewer of the performance are derived. By evaluating the fun of the performance performed by the performer based on the second analysis result showing the tendency of the performance, the fun of the oral performance can be quantitatively evaluated.

(Second Embodiment)
A second embodiment will be described. In the first embodiment described above, the evaluation result of the fun quantitative evaluation device 1 based on the data (history data) of each holding time of M1-2001 to M1-2010 is compared with the examination result of each holding time. The effectiveness was confirmed. Instead, in the second embodiment, an example of predicting the judging result of the comic dialogue contest will be described. For example, as for the M-1 Grand Prix, the 11th M-1 Grand Prix 2015 (referred to as M1-2015) was held after a four-year non-holding period after the above-mentioned 10 times were held. An example of applying the method of this embodiment to M1-2015 and predicting the ranking of the final and the first group of the final final before the final is carried out will be described.

FIG. 15 is a diagram showing the configuration of the fun quantitative evaluation device 1 of the present embodiment. The differences from FIG. 3 will be mainly described.
The storage unit 10 in the fun quantitative evaluation device 1 stores data (history data) for 10 times from M1-2001 to M1-2010 as learning data 11, and serves as evaluation data 13 for the preliminary meeting in M1-2015. And store the quarterfinal data.
The evaluation unit 21 evaluates the fun of the oral performance of the target group based on the data of the qualifying session and the quarterfinals in M1-2015, which is the evaluation data 13, and the prediction accuracy calculation unit 26 determines the ranking after the final match. Predict.

FIG. 16 is a diagram showing the progress of M1-2015, which is an example of a comic contest. In M1-2015, after three qualifying rounds of the first, second and third rounds were held, the quarterfinals and semifinals were held, and eight groups were selected as finalists in the semifinals. In M1-2015, performers who made interesting oral presentations will be selected in stages by dividing them into multiple times. Participants (performers) of M1-2015 will perform the performance at each predetermined time and will be selected based on the degree of fun of the performance. The results of the selection of the eight groups were released on November 19, 2015.

In addition to the eight groups selected above, one group will be selected and added to the M1-2015 finalists from the groups that lost by the semi-finals. The results will be released on the day of the final (December 6, 2015).

The final of M1-2015 will be contested by the above 9 groups (final contestants), and the 3 groups that have won will be selected as the groups that will participate in the final final. In the final battle, the first group (winner) will be selected from the three selected groups. The final and final battle will be held on the same day.

The fun quantitative evaluation device 1 of the present embodiment predicts the ranking in the finals of the eight groups selected in the semifinals and the winner of the final final at the stage when the semifinals of M1-2015 are completed. However, the results of this forecast do not include one group that will be added on the day of the final.

FIG. 17 is a diagram showing an outline of the ranking prediction process of the final match of M1-2015, which is an example of the comic contest. The ranking prediction process of the final match of M1-2015 is carried out by the evaluation unit 21, the order determination unit 23, and the prediction accuracy calculation unit 26. The order determination unit 23 and the prediction accuracy calculation unit 26 are examples of the estimation unit. For example, the order determination unit 23 and the prediction accuracy calculation unit 26 estimate the ranking of the group based on the evaluation result by the evaluation unit 21.

The evaluation unit 21 uses the above-mentioned 10 models in the process of predicting the ranking in the final. These 10 models are learned and generated by the procedure shown in the first embodiment. The evaluation unit 21 is based on the analysis results of the images of the third round and the quarterfinals of M1-2015.
Evaluate the fun of the oral performance of each group.

For example, the evaluation unit 21 derives the evaluation value for each group of eight groups selected in the semi-finals based on the analysis result of the video of the third round and the analysis result of the video of the quarter-finals. .. The evaluation unit 21 derives the average (average evaluation value) of the evaluation values of each time for each group. That is, the evaluation unit 21 is interested in the performance performed by the group based on the first analysis result and the second analysis result in the performance performed by the group in the third round (first round) of the predetermined rounds. The first evaluation value of is derived. In addition, the evaluation unit 21 groups based on the first analysis result and the second analysis result in the performance performed by the group in the quarterfinals (the second time different from the first time) among the predetermined times. Derivation of the second evaluation value of the fun of the performance played by. For example, the evaluation unit 21 derives the average (average evaluation value) of the first evaluation value and the second evaluation value. For example, the evaluation unit 21 derives a weighted average (average evaluation value) of the first evaluation value and the second evaluation value based on a predetermined coefficient value.

The order determination unit 23 derives an evaluation point of the fun of the oral performance of the target group based on the average evaluation value for each group derived by the evaluation unit 21 by the weighted average that emphasizes the quarterfinals.
The prediction accuracy calculation unit 26 predicts the ranking of each group based on the average score of the fun of the oral performance derived by the order determination unit 23.

FIG. 18 is a diagram showing the result of prediction of the ranking of the final match according to the present embodiment. As shown in the figure, the order determination unit 23 predicted the top three groups from the above results. In addition, the figure also shows the actual examination results. As shown in the figure, the results of the predictions for the top three groups were the same as the actual results. In addition, the value of the ranking correlation coefficient of the final match was higher than the average value indicating the result of the verification of the first embodiment. The value for all 8 groups was 0.69.

The condition from which the above values are derived is different from the case where the result of the first embodiment is derived in the following points. The data used for the prediction of M1-2015 is based on the oral performances of the same group in the qualifying rounds held prior to the finals. The content of the oral performance in the preliminary round is different from the oral performance in the final battle. As described above, according to the above-described embodiment, the skill of the group can be evaluated without being limited to evaluating the oral performance by a specific line or group.

According to the above-described embodiment, in addition to achieving the same effect as that of the first embodiment, the judging result of the comic dialogue contest can be predicted by quantitatively evaluating the fun of the oral performance.

(Third Embodiment)
A third embodiment will be described. In the first embodiment and the second embodiment described above, a manzai has been illustrated and described. Instead of this, in the third embodiment, a control will be illustrated and described.

In the case of the above-mentioned manzai, the oral performance was evaluated based on the six indexes of p1 to p6. In the case of Tale, the oral performance is evaluated based on the following eight indicators excluding p1 to p9 and p4. I will mainly explain the differences from Manzai. Even in the case of the control, among the above eight indexes, the five indexes of the indexes p1 to p3 and the indexes p5 and p6 are the same as in the case of the comic book.

p7 is the time per pronunciation. p8 is the time from the penultimate laughter to the end of the oral performance. p9 is the time from the start of the oral performance to the first laughter. The above index p7 (time per pronunciation) may be defined as, for example, the equation (7).

With the change in the number of the above indexes, the number of input nodes of the model generated by the neural network 241 and the model generation unit 24 in the model generation unit 24 is set to eight, and the index targeted by the evaluation unit 21 is set to the above eight. Make it an individual.
For example, the process related to model generation and the process related to evaluation are the same as in the first embodiment.

According to the above embodiment, in addition to producing the same effect as that of the first embodiment, it is possible to control the target performance.

(Fourth Embodiment)
A fourth embodiment will be described. In the fourth embodiment, a method for improving the accuracy of detecting the laughter of the audience and detecting the time of occurrence (time when the laughter occurs) will be described.

Accurate measurement of the loudness and duration of laughter can make it difficult to identify the loudness of a voice, such as when obtaining the results of an automatic identification process based on a voice waveform. In addition, when a person makes a judgment, there are individual differences, and when the judgment is different, it becomes difficult to make a judgment based on the same judgment criteria. For example, when the determination becomes difficult, the following cases can be mentioned. In the continuous material element, the next laugh occurs before the laugh of the previous audience converges. In such a case, the laughter of the audience does not have to be interrupted, and even if the next laughter occurs before the laughter of the previous audience converges, it is measured as a new material element.

In the present embodiment, in such a case, three judges independently detect the laughter of the spectator and the time when the laughter occurs (the time when the laughter occurs), and the spectators detected by two or more of them. Laughter is used for analysis as a material element.

FIG. 19 is a configuration diagram of an interesting quantitative evaluation device 1 including the analysis device of the present embodiment.
The fun quantitative evaluation device 1 includes a storage unit 10, an evaluation unit 21, an order determination unit 23, a prediction accuracy calculation unit 26, a model generation unit 24, and a feature quantity acquisition unit 26.

The feature amount acquisition unit 26 includes an operation detection unit 261, an operation determination unit 262, and a timekeeping unit 263.

For example, the input / output device 1E (FIG. 2) is provided with a dialogue material element detection key operated by a judge who determines laughter in oral performance, and an operation material element detection key. The dialogue material element detection key is a key that is operated when there is a material element containing a dialogue. The operation material element detection key is a key that is operated when there is a material element that does not include a dialogue. For example, the operation material element detection key is operated in the case of an operation-only material element.

For example, a dialogue material element detection key and an operation material element detection key are paired, and the pair is provided in association with each judge.

The operation detection unit 261 detects the operation of the dialogue material element detection key and the operation material element detection key. For example, the operation detection unit 261 detects operations on the three sets of dialogue material element detection keys and operation material element detection keys, which are operated by the three judges, respectively.

The operation determination unit 262 detects whether or not the same type of key is operated among the above three sets of keys, and detects the occurrence of a material element corresponding to the key according to the type and number of the detected keys. To do. The operation determination unit 262 determines that the state in which the operation of 0 or 1 key is detected among the three keys of the same type is the "non-detection state", and 2 or 3 from the "non-detection state". When the key operation changes to the detected state, it is determined as the "detected state". That is, the operation determination unit 262 determines the state of returning from the "detection state" to 0 or 1 as the above-mentioned "non-detection state".
The above detection method is common to the dialogue material element detection key and the operation material element detection key.

The timekeeping unit 263 measures and records the elapsed time from the start time of the comic dialogue in synchronization with the detection of the occurrence of the material element.

The feature amount acquisition unit 26 detects the time of occurrence of the material element based on the result of the determination by the operation determination unit 262, and stores the result in the storage unit 10. For example, the judge presses a predetermined key at the moment when it is determined that the laughter of the audience is occurring while watching the video. At that time, the judge operates the dialogue material element detection key in the case of the material element including the dialogue, and operates the operation material element detection key in the case of the material element containing only the operation. The feature amount acquisition unit 26 detects an operation of the determiner and measures the timing at which the operation is detected to automatically collect data for calculating the index p6. When the feature amount acquisition unit 26 detects the operation of each key, the feature amount acquisition unit 26 records the elapsed time from the start time of the comic book. The feature amount acquisition unit 26 calculates the average value of the time data of two or three judges who have determined that the audience is laughing, and the calculated time is used as the generation time of the material element by the evaluation unit 21. Used for evaluation etc.

When comparing comic dialogue performances with different stage settings, number of spectators, and situations, the feature amount acquisition unit 26 may detect the magnitude of laughter or may normalize the duration of laughter. ..

FIG. 20 is a diagram showing an example of an index obtained by the feature amount acquisition unit 26 according to the present embodiment. The value of the index shown in the figure is used as training data of the neural network 241.

According to the above embodiment, in addition to achieving the same effect as that of the first embodiment, it is possible to improve the accuracy of detecting the laughter of the audience and the time when the laughter occurs (time when the laughter occurs). ..

(Fifth Embodiment)
A fifth embodiment will be described. In the fifth embodiment, a method for supporting the creation of a comic script will be described.

The fun quantitative evaluation device 1 supports the creation of a highly evaluated comic script by being used for the creation of a comic script.

For example, a user who creates a script creates a script that describes the lines while measuring the number of lines, the number of mora of the longest line, and the total number of mora. At this time, if the oral performance time is not specified, the oral performance time is determined.

It is assumed that the fun quantitative evaluation device 1 is in a state where it is possible to evaluate each oral performance after completing the preliminary learning.

The fun quantitative evaluation device 1 derives the time that can be allocated to one line and the number of mora that can be allocated to one second of utterance based on the number of lines, the total number of mora, and the oral performance time. The fun quantitative evaluation device 1 measures the interval of the material element (the part where laughter is expected to occur) that causes the audience to laugh by the number of mora. For example, the fun quantitative evaluation device 1 calculates the approximate value of the time between the material elements from the number of mora that can be assigned to one second, or measures the number of mora based on the voice read aloud by the user.

In addition, the fun quantitative evaluation device 1 adjusts the composition of the script so that the last material element is arranged at the end as much as possible. The evaluation unit 21 of the fun quantitative evaluation device 1 carries out the evaluation based on the result of the above adjustment. The fun quantitative evaluation device 1 uses the result of the evaluation by the evaluation unit 21 as the evaluation value of the script.

In addition, when a rehearsal exercise in which the listener is present on behalf of the audience is carried out, the fun quantitative evaluation device 1 records the time when the listener laughs in the rehearsal exercise, and the position and time interval of the material element. To recalculate. If the time that can be allocated to the dialogue and the number of mora that can be allocated to 1 second are not included in the recommended value range, a good evaluation value cannot be obtained. The fun quantitative evaluation device 1 may add or delete lines, mora, movements and gestures until the evaluation value becomes a good value, and repeatedly adjust the composition of the script. As a result, when the value of the evaluation item such as the time per line of the script approaches the recommended value, the fun quantitative evaluation device 1 outputs an evaluation value of a good value.

In this way, the fun quantitative evaluation device 1 supports the creation of the comic script by quantitatively evaluating the comic script. The fun quantitative evaluation device 1 uses an index that can be evaluated quantitatively to evaluate the manzai performance, verify how much the script writing ability has improved, detect points to be improved, and perform the manzai. It enables objective comparison of. The fun quantitative evaluation device 1 objectively evaluates the oral performance without depending on the linguistic information and the characteristics related to the time series by grasping the global characteristics of the comic dialogue performance with a quantitative index.

According to the above-described embodiment, in addition to achieving the same effect as that of the first embodiment, it is possible to support the script creation of a comic book.

Although the preferred embodiments of the present invention have been described in detail above, the present invention is not limited to the specific embodiments, and varies within the scope of the gist of the present invention described in the claims. Can be transformed / changed.

In the above embodiment, a case is illustrated in which only the comic dialogue groups that participated in the specific comic dialogue contest (for example, M-1 Grand Prix) that are held repeatedly are compared within a limited range. Each of the specific Manzai contests that are held repeatedly is held without any major changes in the number of judges, the size of the audience, etc. If it is held with the same number of judges and the size of the audience, it is not easily affected by the difference. In order to improve the prediction performance, the fun quantitative evaluation device 1 may evaluate a comic dialogue performed on a stage under the same or similar conditions as the training data, such as the M-1 Grand Prix. In addition to the M-1 Grand Prix, a broadcast program "THE MANZAI" and the like are known. The present embodiment may also be applied to "THE MANZAI" and the like.

On the other hand, even in the case of analyzing including comics performed on the stage under different conditions, the fun quantitative evaluation device 1 can evaluate based on the generality of the fun of oral performance. It is natural to think that the manzai that the audience finds interesting changes over time. The fact that we were able to predict the results of the 2015 contest using the 2001-2010 Manzai is the core of the fun of the Manzai, which is independent of the long-term changes in taste and fashion / obsolescence of 10 years. It means that we were able to catch the above and that we were able to obtain an evaluation method of fun that would be valid five years later. Furthermore, since there is no overlap between the judges of the 2001-2010 and 2015 M-1 Grand Prix, it can be seen that it does not depend on the combination of judges.

In the above embodiment, it has been shown that the fun of comics can be quantitatively evaluated by using only static and non-verbal information. However, the fun quantitative evaluation device 1 is dynamic or linguistic. Information may be evaluated in combination with static and non-verbal information.

The fun quantitative evaluation device 1 may be applied not only to the applications shown in the embodiment but also to the following applications.
For example, since the fun quantitative evaluation device 1 evaluates using a quantitative index, it makes a judgment for changing a prepared script or a manzai oral performance, clarifies the points to be changed, teaches a manzai oral performance, etc. It can be applied to the purpose of. Furthermore, when making a script for a manzai, it is considered to be useful for providing an interesting script by using it for evaluation before speaking or for evaluation of an automatically generated script.

In addition, we are analyzing the highly skilled manzai groups that have passed the preliminary rounds among the participating groups. However, among them, there is a difference between the upper and lower ranks, and by clarifying the cause of the difference, it is possible to set specific goals for demonstrating a more highly evaluated manzai. For example, it can be used as an examiner robot that judges qualifying for a comic contest.

In addition, the fun quantitative evaluation device 1 can extract "interesting" comics. This feature may be used for the production and planning of comics, or for organizing programs such as selecting a group to broadcast and determining the order of groups.

1 ... Interesting quantitative evaluation device, 10 ... Storage unit, 11 ... Learning data, 12 ... Model data, 13 ... Evaluation data, 14 ... Evaluation result data, 15 ... Examination result data, 21 ... Evaluation unit, 22 ... Average Point calculation unit, 23 ... order determination unit, 24 ... model generation unit, 25 ... feature quantity extraction unit, 26 ... prediction accuracy calculation unit, 131 ... first analysis result, 132 ... second analysis result.

Claims (16)

The first analysis result showing the tendency of the performer's speaking style derived based on the voice emitted by the performer during the performance, and the tendency of the content performed by the performer derived based on the laughter of the viewer of the performance. The evaluation unit that evaluates the fun of the performance performed by the performer based on the second analysis result showing
Equipped with a,
The first analysis result includes the result of evaluation of the tempo of the interaction of the plurality of performers.
Interesting quantitative evaluation device. The first analysis result does not include the analysis result regarding the utterance content of the performer based on the voice uttered by the performer .
The second analysis result includes the result of evaluation of the ratio of the material element without dialogue in the performance.
The fun quantitative evaluation device according to claim 1. The first analysis result showing the tendency of the performer's speaking style derived based on the voice emitted by the performer during the performance, and the tendency of the content performed by the performer derived based on the laughter of the viewer of the performance. Evaluation unit that evaluates the fun of the performance performed by the performer based on the second analysis result showing
With
The second analysis result includes information on the time interval in which the performer played the material element in the performance , or information on the time interval.
The latter half of the performance includes information on the time from when the performer plays the material element to the end of the performance.
Interesting quantitative evaluation device. The material element played by the performer in the latter half of the performance is the material element played at the end of the performance.
The fun quantitative evaluation device according to claim 3 . The first analysis result showing the tendency of the performer's speaking style derived based on the voice emitted by the performer during the performance, and the tendency of the content performed by the performer derived based on the laughter of the viewer of the performance. Evaluation unit that evaluates the fun of the performance performed by the performer based on the second analysis result showing
With
The first analysis result includes information indicating the amount of information of the voice emitted by the performer.
Surface white of the quantitative evaluation device. The amount of information in the voice includes the time allotted to the dialogue, the number of mora assigned to the unit time in the voice uttered by the performer, one of the number of mora of the relatively long dialogue of the performer, or Includes multiple
The fun quantitative evaluation device according to claim 5 . The evaluation unit is based on the first analysis result and the second analysis result of the performance performed by the performer in the first of the predetermined times of the contest which is divided into a plurality of times and selected stepwise. The first evaluation value of the fun of the performance performed by the performer and
The fun of the performance performed by the performer based on the first analysis result and the second analysis result in the performance performed by the performer in the second time different from the first time among the predetermined times. Derived from the second evaluation value of
An estimation unit for estimating the fun of the performance performed by the performer after both the first and second evaluations based on the first evaluation value and the second evaluation value is further provided.
The fun quantitative evaluation device according to any one of claims 1 to 6 . The first analysis result showing the tendency of the performer's speaking style derived based on the voice emitted by the performer during the performance, and the tendency of the content performed by the performer derived based on the laughter of the viewer of the performance. Evaluation unit that evaluates the fun of the performance performed by the performer based on the second analysis result showing
With
The evaluation unit
The performer performed based on the first analysis result and the second analysis result of the performance performed by the performer in the first of the predetermined times of the contest which is divided into a plurality of times and selected stepwise. The first evaluation value of the fun of the performance and
The fun of the performance performed by the performer based on the first analysis result and the second analysis result in the performance performed by the performer in the second time different from the first time among the predetermined times. Derived from the second evaluation value of
Both the first and second evaluations are based on the first evaluation value, the second evaluation value, and the coefficients that weight the first evaluation value and the second evaluation value. It further includes an estimation unit that estimates the fun of the performance performed by the performer after the times .
Interesting quantitative evaluation device. A model generation unit that generates a model for evaluating the fun of the performance based on the feature amount of the performance whose fun is evaluated and the evaluation result.
With
The feature quantity of the performance evaluated for the fun includes the voice emitted by the performer who performed the performance evaluated for the fun and the laughter of the person who watched the performance evaluated for the fun. ,
The fun quantitative evaluation device according to any one of claims 1 to 8. The computer
The first analysis result showing the tendency of the performer's speaking style derived based on the voice emitted by the performer during the performance, and the tendency of the content performed by the performer derived based on the laughter of the viewer of the performance. The process of evaluating the fun of the performance performed by the performer based on the second analysis result showing
Only including,
The first analysis result includes the result of evaluating the tempo of the interaction of a plurality of people who are the performers, or the information indicating the amount of information of the voice emitted by the performer, or
The second analysis result includes information on the time interval in which the performer played the material element in the performance, or the time from when the performer played the material element in the latter half of the performance to the end of the performance. Contains information,
Interesting quantitative evaluation method. The first analysis result showing the tendency of the performer's speaking style derived based on the voice emitted by the performer during the program, and the tendency of the content performed by the performer derived based on the laughter of the viewer of the program. and a second analysis result showing the step of evaluating the fun of the performers plays repertoire based on a program for causing a computer to execute the interesting quantitative evaluation device,
The first analysis result includes the result of evaluating the tempo of the interaction of a plurality of people who are the performers, or the information indicating the amount of information of the voice emitted by the performer, or
The second analysis result includes information on the time interval in which the performer played the material element in the performance, or the time from when the performer played the material element in the latter half of the performance to the end of the performance. Contains information,
Program . The first analysis result showing the tendency of the performer's speaking style derived based on the voice emitted by the performer during the performance, and the tendency of the content performed by the performer derived based on the laughter of the viewer of the performance. An evaluation unit formed so as to generate an evaluation value of the result of evaluating the fun of the performance performed by the performer based on the second analysis result indicating
With
Adjust the composition of the script of the performance so that the evaluation value becomes a desired value.
Interesting quantitative evaluation device. The first analysis result showing the tendency of the performer's speaking style derived based on the voice emitted by the performer during the performance, and the tendency of the content performed by the performer derived based on the laughter of the viewer of the performance. An evaluation unit formed so as to generate an evaluation value of the result of evaluating the fun of the performance performed by the performer based on the second analysis result indicating
With
By adjusting the composition of the program so that the evaluation value becomes a desired value, the creation of the script of the program is supported.
Interesting quantitative evaluation device. The first analysis result showing the tendency of the performer's speaking style derived based on the voice emitted by the performer during the performance, and the tendency of the content performed by the performer derived based on the laughter of the viewer of the performance. It is provided with an evaluation unit formed so as to generate an evaluation value of the result of evaluating the fun of the performance performed by the performer based on the second analysis result showing.
Quantitatively evaluate the script of the performance based on the evaluation value.
Interesting quantitative evaluation device. It is a fun adjustment method for the computer to help adjust the fun of the performance.
The computer played the first analysis result showing the tendency of the performer's speaking style derived based on the voice emitted by the performer during the performance, and the performer derived based on the laughter of the viewer of the performance. The process of evaluating the fun of the first performance performed by the performer based on the second analysis result showing the tendency of the content, and
The process in which the computer evaluates the fun of the second performance after the fun is adjusted based on the result of the evaluation .
Interesting adjustment method including . The first analysis result showing the tendency of the performer's speaking style derived based on the voice emitted by the performer during the performance, and the tendency of the content performed by the performer derived based on the laughter of the viewer of the performance. A computer adjusted so that the fun of the performance performed by the performer can be evaluated based on the second analysis result showing
The process of evaluating the target program and supporting the creation of the script of the target program so that the desired fun can be obtained from the target program .
Interesting adjustment method including .

Images