JP2022047835A - Determination program, determination method and information processing device - Google Patents

Abstract

To improve analysis accuracy.SOLUTION: An information processing device calculates statistics of feature quantities of a plurality of sample data, about a plurality of datasets each including the plurality of sample data. The information processing device executes selection or determination of priorities of datasets as training data used for training of a determination model, among the plurality of datasets, based on the calculated statistics.SELECTED DRAWING: Figure 5

Description

The present invention relates to a determination program, a determination method and an information processing apparatus.

In recent years, call centers have recorded conversations between operators and customers and accumulated information on the recorded conversations. Call centers have a need to improve their services by utilizing the accumulated conversation information.

For example, there are the following technologies for utilizing the accumulated conversation information. There is a technology that determines the number of predetermined keywords included in a conversation between a customer and an operator, displays FAQ (Frequently Asked Questions) according to the number of determined keywords, and notifies the supervisor. In addition, there is a technology that converts the operator's voice into a character string and checks whether the character string contains the keyword to be transmitted to determine whether the operator properly conveys the content to be transmitted to the customer. ..

Japanese Unexamined Patent Publication No. 2015-99304

In a call center, there is a need to detect whether or not a specific conversation situation specified in advance exists in a conversation between an operator and a customer. Here, when detecting whether or not a specific conversation situation exists by using the above technology, keyword setting is comprehensively performed according to the conversation situation to be detected, and the set keyword is the conversation information. It is conceivable to perform a process of determining whether or not it is included in.

However, it is difficult to know in advance how many keywords should be covered in order to detect a specific conversation situation. In addition, it is difficult to comprehensively set keywords manually because there are various phrases even in conversations with the same meaning.

Therefore, it is conceivable to use a trained model trained by machine learning to analyze the conversation situation. In addition, the time required to generate the trained model and the performance of the trained model depend on the nature of the training data. However, it is difficult to specify the nature of the training data in advance, and even if it can be specified, it takes a lot of time.

In one aspect, it is an object of the present invention to provide a decision program, a decision method, and an information processing apparatus capable of improving analysis accuracy.

In the first proposal, the determination program causes a computer to perform a process of calculating a feature amount statistic of the plurality of sample data for a plurality of data sets each including a plurality of sample data. The determination program causes a computer to execute a process of selecting a data set as training data to be used for training a judgment model or determining a priority of the plurality of data sets based on the calculated statistic.

According to one embodiment, it is possible to improve the analysis accuracy.

FIG. 1 is a diagram illustrating the overall flow of quantization. FIG. 2 is a functional block diagram showing a functional configuration of the information processing apparatus according to the first embodiment. FIG. 3 is a diagram illustrating an example of an update conversation data set. FIG. 4 is a diagram illustrating an example of the certainty DB. FIG. 5 is a diagram illustrating a flow of calculating the certainty of the conversation data set for update. FIG. 6 is a diagram illustrating the setting of the coefficient that determines the update speed. FIG. 7 is a flowchart showing the flow of processing. FIG. 8 is a diagram showing an overall configuration example of the application system. FIG. 9 is a diagram for explaining an example of processing of the learning device and the determination device of the application system. FIG. 10 is a diagram for explaining a generation unit of the learning device. FIG. 11 is a diagram illustrating the update of the quantization table. FIG. 12 is a diagram illustrating the overall flow of adaptive control of the quantization table. FIG. 13 is a diagram illustrating an example of distribution of a conversation data set. FIG. 14 is a diagram illustrating an example in which the quantization table is updated with the conversation data set of line B in FIG. FIG. 15 is a diagram illustrating an example in which the quantization table is updated with the conversation data set of the R line in FIG. FIG. 16 is a flowchart showing the flow of processing according to the third embodiment. FIG. 17 is a diagram illustrating an update process according to the third embodiment. FIG. 18 is a diagram illustrating an application example. FIG. 19 is a diagram illustrating a hardware configuration example.

Hereinafter, examples of the determination program, determination method, and information processing apparatus disclosed in the present application will be described in detail with reference to the drawings. The present invention is not limited to this embodiment. In addition, each embodiment can be appropriately combined within a consistent range.

[Overall explanation]
In recent years, in order to support the operator of the contact center, the conversation situation including the detection of the occurrence of a response trouble has been analyzed. In contact centers, which are becoming popular as one of the customer contact points, waiting time is directly linked to customer dissatisfaction, so it is important to increase the operating rate of operators. However, when a trouble with a customer such as a complaint occurs, the response time becomes longer than usual, and as a result, the waiting time for the operator connection of another customer becomes long. In order to alleviate the situation as much as possible, it is necessary for the contact center manager to understand the response and provide support such as taking over at an early stage.

In order to realize support by the administrator, it is necessary to detect the occurrence of response troubles at an early stage. Therefore, when a response trouble occurs, the keyword (for example, "yes"), which is a language characteristic of the conversation between the customer and the operator, and the bias of the phoneme arrangement are detected by voice recognition, and the conversation situation is analyzed based on the result. Technology is used.

However, in order to detect language features, it is necessary to create a model of language-dependent information such as an acoustic model for speech recognition and a language model in advance. Here, these creation costs are required for each language, and in particular, the creation of an acoustic model is expensive, so the realization of conversational situation analysis that does not use language-dependent information is eagerly desired.

In recent years, a technique has been used to quantify the physical features of speech such as the power spectrum with reference to human auditory characteristics, and to detect the bias of vocalization used for analysis of conversational situations based on the results. There is.

FIG. 1 is a diagram illustrating the overall flow of quantization. As shown in FIG. 1, the quantization table is updated using an update conversation data set containing a plurality of speech data for update in order to adapt to the distribution of physical features of speech. .. Then, using the optimized quantization table, the quantization of the training data set containing a plurality of voice data for training is executed, and the training of the judgment model (machine learning) using the quantization result of each voice data. ) Is executed. After that, conversation data (evaluation conversation data) collected in real time is input to the trained judgment model, analysis results are acquired, and the occurrence of response trouble (abnormal call) is detected.

In general, the larger the difference in the distribution (mean, variance) of the quantization results between normal and abnormal conversations, the better the analysis accuracy of conversation data. For example, when updating the quantization table using normal conversation data, which is relatively easy to collect, the dispersion of the quantization result of the normal conversation data (normal conversation) becomes small at the time of evaluation, and at the time of abnormality. The dispersion of the quantization result of the conversation data (abnormal conversation) of is large.

Here, in the quantization table update process used for quantization, the speed of convergence at the time of updating the quantization table changes depending on the characteristics of the feature amount of the conversation data set for update, and the quantization result of the feature amount. Distribution changes. That is, when the convergence at the time of updating the quantization table is slow, the variance of the quantization result of the feature amount of the conversation data for evaluation becomes relatively large. For example, when the convergence is slow when updating the quantization table using the data set of the normal conversation, the dispersion of the quantization result of the normal conversation for evaluation is larger than when the update converges quickly, and the abnormality for evaluation is large. The difference between the quantized result of the feature quantity of the conversation is relatively small. That is, the analysis accuracy of normal conversation and abnormal conversation is lowered.

Therefore, the disclosed technique realizes selection of a conversation data set suitable for updating the quantization table by examining the characteristics of the features of the conversation data set for updating the quantization table.

[Functional configuration]
Next, the functional configuration of the information processing apparatus 50 that realizes the selection of the conversation data set suitable for updating the quantization table will be described. FIG. 2 is a functional block diagram showing a functional configuration of the information processing apparatus 50 according to the first embodiment.

As shown in FIG. 2, the information processing apparatus 50 includes a communication unit 51, a storage unit 52, and a control unit 60. The communication unit 51 is a processing unit that controls communication with other devices, and is realized by, for example, a communication interface. For example, the communication unit 51 receives various instructions and various data from the administrator terminal and the like, and transmits various results to the administrator terminal and the like.

The storage unit 52 is a processing unit that stores various data, programs executed by the control unit 60, and the like, and is realized by, for example, a memory or a hard disk. The storage unit 52 stores the conversation data set 53 for updating and the certainty DB 54.

The update conversation data set 53 has a plurality of data sets having a plurality of conversation data, and is a data set used for updating the above-mentioned quantization table. That is, the conversation data set 53 for update is a target for investigating the characteristics of the feature amount. For example, the update conversation data set 53 is a data set of normal conversation data collected at different contact centers.

FIG. 3 is a diagram illustrating an example of the conversation data set 53 for update. As shown in FIG. 3, the update conversation data set 53 stores the “data set name, data” in association with each other. The "data set name" is information that identifies the data set. "Data" is conversation data used to update the quantization table, and is time-series data. In the example of FIG. 3, it is shown that the conversation data set A is composed of a plurality of conversation data such as conversation data A1 and conversation data A2.

The certainty DB 54 stores the survey results for each conversation data set. That is, the certainty DB 54 stores the certainty used for determining the priority at the time of updating the quantization table for each conversation data set stored in the update conversation data set 53.

FIG. 4 is a diagram illustrating an example of the certainty DB 54. As shown in FIG. 4, the certainty DB 54 stores the “data set name, data, and certainty” in association with each other. The "data set name" is information that identifies the data set, and the "data" is conversation data used to update the quantization table. The “certainty” is a score calculated by the control unit 60, which will be described later, and is used for determining the priority at the time of updating the quantization table. In the example of FIG. 4, the conversation data set A is composed of a plurality of conversation data such as conversation data A1 and conversation data A2, and shows that "probability 1" is calculated as "probability".

The control unit 60 is a processing unit that controls the entire information processing device 50, and is realized by, for example, a processor. The control unit 60 has an acoustic analysis unit 61, a conversation statistics unit 62, an overall statistics unit 63, and a calculation unit 64. The acoustic analysis unit 61, the conversation statistics unit 62, the overall statistics unit 63, and the calculation unit 64 are realized by an electronic circuit possessed by the processor, a process executed by the processor, or the like.

The control unit 60 calculates the certainty of each conversation data set stored in the update conversation data set 53. FIG. 5 is a diagram illustrating a flow of calculating the certainty of the conversation data set for update. As shown in FIG. 5, the control unit 60 executes processing for each conversation data and processing for each conversation data set.

First, the control unit 60 reads out one conversation data set, executes acoustic analysis in a predetermined unit (predetermined frame unit) for each conversation data included in the conversation data set, and calculates the feature amount of the predetermined unit. Then, the control unit 60 calculates the statistic (average, variance) of the feature amount for each conversation data by using the feature amount of a predetermined unit. In this way, the control unit 60 calculates and stores the feature amount for each conversation data of each conversation data set. Next, the control unit 60 calculates the variance value in the conversation data set for each conversation data set by using the feature amount of each conversation data included in the conversation data set. Then, the control unit 60 calculates the certainty suitable for updating by using the distributed value of each conversation data set.

The acoustic analysis unit 61 is a processing unit that calculates a feature amount of conversation data. Specifically, the acoustic analysis unit 61 calculates MFCC (mel-frequency cepstrum coefficient), which is an example of the feature amount, for each conversation data stored in the update conversation data set 53 in units of analysis frames.

For example, the acoustic analysis unit 61 reads out the conversation data A1 and extracts each data of the analysis frame unit by the window hanging from the conversation data A1. Then, the acoustic analysis unit 61 calculates the feature amount MFCC (i, j, t) for each analysis frame and outputs it to the conversation statistics unit 62. In addition, i is a data number, t is an analysis frame number, and j is a dimension of MFCC.

The conversation statistics unit 62 is a processing unit that calculates statistics (average, variance) for each conversation data. Specifically, the conversation statistics unit 62 has a feature amount MFCC (i, j) for each analysis frame calculated by the acoustic analysis unit 61 for each conversation data of each conversation data set stored in the update conversation data set 53. , T) is used to calculate the statistics in the conversation data. For example, when the conversation statistics unit 62 calculates 30 feature quantities MFCC for the conversation data A1 in analysis frame units, the conversation statistics unit 62 calculates the statistics of the conversation data A using the 30 feature quantities MFCC.

More specifically, the conversation statistics unit 62 uses the feature amount MFCC (i, j, t) of each analysis frame unit in the conversation data for each conversation data, and averages the feature amount MFCC (i, j, t). m (i, j) is calculated by the formula (1), and the variance v (i, j) of the feature quantity MFCC (i, j, t) is calculated by the formula (2). Then, the conversation statistics unit 62 outputs the calculated average and variance to the overall statistics unit 63. Note that Nt in the equations (1) and (2) is the number of frames in the conversation data.

The overall statistics unit 63 is a processing unit that calculates statistics (mean, variance) for each conversation data set. Specifically, the overall statistics unit 63 uses the statistics of each conversation data calculated by the conversation statistics unit 62 for each conversation data set stored in the update conversation data set 53, and statistics in the conversation data set. Calculate the amount. For example, the overall statistics unit 63 uses the average and variance of the conversation data A1 in the conversation data set A, the average and the variance of the conversation data A2, and the like for the conversation data set A, and uses the statistics (average, average, etc.) of the conversation data set A. Dispersion) is calculated.

More specifically, the overall statistics unit 63 uses the mean m (i, j) for each conversation data and the variance v (i, j) for the mean m (i, i, for each conversation data included in the conversation data set. The average mm (j) of j) is calculated by the equation (3), and the average vm (j) of the variance v (i, j) is calculated by the equation (4). Further, the overall statistics unit 63 uses the average m (i, j) for each conversation data and the variance v (i, j) to determine the average m (i, j) for each conversation data included in the conversation data set. The variance mv (j) is calculated by the equation (5), and the variance vv (j) of the variance v (i, j) is calculated by the equation (6). Then, the overall statistics unit 63 outputs the calculated information to the calculation unit 64. In addition, Ni in equations (3) to (6) is the number of conversation data.

The calculation unit 64 is a processing unit that calculates the certainty suitable for updating for each conversation data set. Specifically, the calculation unit 64 calculates the certainty of each conversation data set using the statistics calculated by the overall statistics unit 63 for each conversation data set stored in the update conversation data set 53. .. For example, the calculation unit 64 calculates the certainty (score) of the conversation data set A using the statistics of the conversation data set A (mean and average variance, variance of variance and variance). In this way, the calculation unit 64 calculates the certainty of each conversation data set and stores it in the certainty DB 54.

More specifically, the calculation unit 64 uses the equation (7) to perform weighting addition and reciprocal calculation of the variances mv (i, j) and vv (i, j) in each data set to obtain a certainty (score). calculate. Nj in the equation (7) is an order of MFCC. That is, the calculation unit 64 calculates that the smaller the variance value of the conversation data set, the higher the certainty that it is suitable for updating the quantization table.

Here, the quantization table is updated so that the initial table is generated due to white noise or the like and the quantization error of the conversation data set for update is minimized. At this time, the calculation unit 64 can control the update speed so that the update speed becomes faster when the certainty suitable for the update is high.

FIG. 6 is a diagram illustrating the setting of the coefficient that determines the update speed. As shown in FIG. 6, the coefficient α is set to a value as the minimum value is 0 and the maximum value is 1, and the greater the certainty, the larger the value. The update formula of the quantization table using such a coefficient α is "quantization point after update n = ((1-α) × quantization point n before update) + (α × quantization point before update). It can be expressed by "the average of the feature quantities quantized by n)". As shown in FIG. 6, the coefficient α that determines the update speed is the fastest in the range of 0 ≦ α ≦ 1 when α = 1, and the update is not executed when α = 0.

[Processing flow]
FIG. 7 is a flowchart showing the flow of processing. As shown in FIG. 7, when the processing start is instructed (S101: Yes), the acoustic analysis unit 61 executes an acoustic analysis for calculating the feature amount of the conversation data in the analysis frame unit (S102).

Subsequently, the conversation statistics unit 62 calculates the average and variance for each conversation data using the feature amount of each analysis frame unit (S103). Then, the overall statistics unit 63 calculates the average and variance in each update conversation data set by using the average and variance for each conversation data (S104).

After that, the calculation unit 64 calculates the certainty of each conversation data set by using the average and the variance in each conversation data set (S105), and uses the certainty of each conversation data set to update the speed. A coefficient for determining the value is set (S106).

[effect]
As described above, the information processing apparatus 50 can calculate the certainty indicating the degree of suitability for updating each conversation data set for updating before updating the quantization table. In this way, by examining the characteristics of the features of the conversation data set for updating the quantization table, it is possible to realize the selection of the conversation data set suitable for updating the quantization table, and as a result, a specific conversation. It is possible to improve the analysis accuracy of analyzing whether or not the situation is included in the voice data.

Next, an application system that updates the quantization table using the update conversation data set for which the "probability" described in the first embodiment is calculated, and executes conversation analysis using the updated quantization table. Will be explained.

[Applicable system]
FIG. 8 is a diagram showing an overall configuration example of the application system. As shown in FIG. 8, this system includes a customer terminal 10, an operator terminal 15, a call recording device 30, an administrator terminal 40, a learning device 100, and a determination device 200. The information processing device 50 described in the first embodiment may be incorporated in the learning device 100, or may be realized by separate devices.

The customer terminal 10 and the operator terminal 15 are connected to each other via a network 1 such as an IP (Internet Protocol) network. Further, the operator terminal 15, the call recording device 30, the administrator terminal 40, the learning device 100, and the determination device 200 are also connected to each other by a predetermined network. In addition, various communication networks such as the Internet and a dedicated line can be adopted for each network regardless of whether they are wired or wireless.

The customer terminal 10 is a terminal device used by a customer to have a conversation (call) with an operator. The operator terminal 15 is a terminal device used by an operator to have a conversation with a customer.

The call recording device 30 is a device that records the voice of a conversation transmitted and received between the customer terminal 10 and the operator terminal 15. At the time of learning, the voice data recorded by the call recording device 30 is notified to the learning device 100 and used as voice data for learning. When the abnormal conversation is detected, the voice data recorded by the call recording device 30 is notified to the determination device 200, and it is determined whether or not the voice data includes an abnormal conversation situation.

The administrator terminal 40 is a terminal device used by an administrator who manages an operator who talks with a customer by using the operator terminal 15. For example, when the determination device 200 determines that the conversation between the customer and the operator includes an abnormal conversation situation, the determination device notifies the administrator terminal 40 of information that the abnormal conversation situation has been detected. Will be done.

The learning device 100 is a device that learns (trains) the LSTM (Long Short Term Memory) model 110c and the DNN (Deep Neural Network) model 110d of FIG. 9 using the learning (training) voice data and the correct answer information. Is. The learning device 100 notifies the determination device 200 of the information of the learned LSTM model 110c and the DNN model 110d.

The determination device 200 uses the LSTM model 110c and the DNN model 110d notified from the learning device 100 to determine whether or not the conversation between the customer terminal 10 and the operator terminal 15 includes an abnormal conversation situation. Is. When the determination device 200 determines that the conversation between the customer and the operator includes an abnormal conversation situation, the determination device 200 notifies the administrator terminal 40 of information that the abnormal conversation situation has been detected.

[Description of processing]
FIG. 9 is a diagram for explaining an example of processing of the learning device and the determination device of the application system. As shown in FIG. 9, the learning device 100 includes a learning voice database 110a, a generation unit 120, a first calculation unit 130, a third calculation unit 140, a second calculation unit 150, and a learning unit 160. Have.

The learning voice database 110a stores a plurality of learning voice data generated by dividing each conversation data, and each learning voice data is associated with the correct answer information 110b. The correct answer information 110b is information given to each voice data indicating whether or not a specific conversation situation is included. In the first embodiment, as an example, a specific conversation situation is referred to as an "abnormal conversation situation". Anomalous conversational situations include "unusual situations" such as customer dissatisfaction, anger, or intimidation.

The generation unit 120 acquires learning voice data from the learning voice database 110a. In the following description of the learning device 100, the learning voice data acquired from the learning voice database may be simply referred to as "voice data". The generation unit 120 performs, for example, vector quantization on the voice data, and generates information (quantization series) of the quantization result. For example, the quantization series is an example of an index showing the bias of vocalization. The generation unit 120 converts each quantization result into a One Hot vector, and outputs the One Hot vector of each quantization result to the first calculation unit 130.

The first calculation unit 130 is a processing unit that calculates an internal vector by sequentially inputting a One Hot vector of a quantization result into a first network having a recursive path and performing a calculation based on the parameters of the first network. .. For example, the first network corresponds to LSTM. The first calculation unit 130 inputs the One Hot vector of the quantization result generated from the voice data to the first network, and outputs each internal vector obtained by inputting to the third calculation unit 140.

The third calculation unit 140 is a processing unit that averages a plurality of internal vectors output from the first calculation unit 130. The third calculation unit 140 outputs the averaged internal vector to the second calculation unit 150. In the following description, the averaged internal vector is referred to as an "average vector".

The second calculation unit 150 is a processing unit that calculates an output value (neuron value) by inputting an average vector into the second network having no recursive path and performing a calculation based on the parameters of the second network. The second calculation unit 150 outputs the output value to the learning unit 160.

When the learning unit 160 inputs the voice data to the first calculation unit 130, the first calculation unit 130 so that the output value output from the second calculation unit 150 approaches the correct answer information 110b corresponding to the voice data. The parameter of the second calculation unit 150 and the parameter of the second calculation unit 150 are learned (learning by the error back propagation method).

The learning unit 160 repeatedly executes error back propagation learning until the learning stop condition is satisfied, and generates an LSTM model 110c and a DNN model 110d. The LSTM model 110c is information corresponding to the trained parameters of the first network. The DNN model 110d is information corresponding to the learned parameters of the second network. The learning device 100 notifies the determination device 200 of the information of the LSTM model 110c and the information of the DNN model 110d. The learning unit 160 may notify the determination device 200 of the information of the LSTM model 110c and the information of the DNN model 110d via the network, or the learning device 100 and the determination device 200 are directly connected to each other. Then, the information of the LSTM model 110c and the information of the DNN model 110d may be notified to the determination device 200.

The determination device 200 includes a generation unit 220, a first calculation unit 230, a third calculation unit 240, a second calculation unit 250, and a determination unit 260.

The generation unit 220 accepts the input of voice data to be detected as to whether or not the conversation status is abnormal. In the following description of the determination device 200, the voice data to be detected whether or not the conversation status is abnormal may be simply referred to as voice data. The generation unit 220 performs, for example, vector quantization on the voice data, and generates information on the quantization result. The generation unit 220 converts each quantization result into a One Hot vector, and outputs the One Hot vector of each quantization result to the first calculation unit 230.

The first calculation unit 230 is a processing unit that calculates an internal vector by sequentially inputting the One Hot vector of each quantization result into the first network having a recursive path and performing a calculation based on the parameters of the first network. be. The first calculation unit 230 uses the parameters of the LSTM model 110c as the parameters to be set in the first network. The first calculation unit 230 inputs the One Hot vector of the quantization result generated from the voice data to the first network, and outputs each internal vector obtained by inputting to the third calculation unit 240.

The third calculation unit 240 is a processing unit that averages a plurality of internal vectors output from the first calculation unit 130. The third calculation unit 140 outputs the averaged internal vector (average vector) to the second calculation unit 250.

The second calculation unit 250 is a processing unit that calculates an output value (neuron value) by inputting an average vector into the second network having no recursive path and performing a calculation based on the parameters of the second network. The second calculation unit 250 uses the parameters of the DNN model 110d as the parameters to be set in the second network. The second calculation unit 250 outputs the output value to the determination unit 260.

The determination unit 260 is a processing unit that compares the output value output from the second calculation unit 250 with the threshold value and determines whether or not the voice data includes an abnormal conversation situation. For example, the determination unit 260 determines that the voice data includes an abnormal conversation situation when the output value is equal to or greater than the threshold value.

As described above, the learning device 100 executes machine learning of the LSTM model 110c and the DNN model 110d by using the set of the quantization result extracted from the learning voice data and the correct answer information. Therefore, machine learning of the LSTM model 110c and the DNN model 110d can be executed without using trial and error, skillful knowledge, and know-how for setting a keyword for detecting a specific conversation situation. Further, the determination device 200 appropriately determines whether or not a specific conversation situation is included in the voice data by processing the voice data using the learned LSTM model 110c and the DNN model 110d. Can be done.

[Explanation of generator]
FIG. 10 is a diagram for explaining a generation unit of the learning device. As shown in FIG. 10, the generation unit 120 includes an acoustic processing unit 121, a quantization table 122, a vector quantization unit 123, and a vectorization unit 124.

The sound processing unit 121 is a processing unit that extracts information used for voice recognition from voice data. The information extracted from the voice data is called a feature quantity. The sound processing unit 121 sets a short section called a frame of about 32 ms in the voice data, and extracts a feature amount while shifting by about 10 ms. For example, the sound processing unit 121 extracts a feature amount from voice data based on the MFCC. The sound processing unit 121 outputs the feature amount to the vector quantization unit 123.

The quantization table 122 is a vector table used for quantization of voice data. This quantization table 122 is a plurality of representative points of feature quantities of speech data, and is pre-optimized by the quantization table generation unit 125 using each voice data for adaptive processing of the quantization table 122. ..

The vector quantization unit 123 collates the feature quantity with the quantization table 122, and corresponds to the feature quantity based on each quantization result (for example, corresponding to the quantization number) corresponding to each quantization point. It is a processing unit that executes a process of outputting a quantization result every time a feature amount is received from the acoustic processing unit 121. The vector quantization unit 123 outputs the information of the quantization series in which the quantization results corresponding to each feature amount are arranged in time series to the vectorization unit 124.

The vectorization unit 124 is a processing unit that converts each quantization result included in the quantization series into a vector. The vectorization unit outputs each vector corresponding to each quantization result of the quantization series to the first calculation unit 130. For example, the vectorization unit 124 represents each quantization result as a 40-dimensional One Hot vector. In the input One Hot vector of the quantization result, "1" is set in the dimension of the input quantization result, and "0" is set in the other dimensions.

The quantization table generation unit 125 is a processing unit that generates the quantization table 122. Specifically, the quantization table generation unit 125 generates the quantization table 122 using the update conversation data set having the certainty calculated by the first embodiment. For example, the quantization table generation unit 125 updates the quantization table 122 by using the update conversation data set having the highest probability or the update conversation data set having the probability equal to or higher than the threshold value.

FIG. 11 is a diagram illustrating the update of the quantization table 122. As shown in FIG. 11, the quantization table generation unit 125 of the generation unit 120 of the learning device 100 performs each voice for adaptive processing which is an example of the conversation data set for update when the quantization table 122 is optimized. Acoustic analysis is performed on the data, and the acoustic processing unit 121 generates a feature amount. Then, the quantization table generation unit 125 collates each feature amount generated from each voice data with the quantization table 122, accumulates the quantization result, and minimizes the quantization error. Quantization table 122 is optimized by repeating the update of.

FIG. 12 is a diagram illustrating an overall flow of adaptive control of the quantization table 122. Here, it is a two-dimensional feature quantity. As shown in FIG. 12, the quantization table generation unit 125 generates the initial value of the quantization table 122 from white noise and the like, quantizes the voice data for adaptive processing, and converts each vector of the quantization table 122. The vector is updated with the average of the selected feature quantities, and this update is repeated to reduce the quantization error. That is, the quantization table generation unit 125 repeatedly updates the quantization table 122 so as to minimize the quantization error according to the distribution of the physical features of the voice data.

At this time, the quantization table generation unit 125 updates the quantization table 122 using the conversation data set having the highest probability or the conversation data set having the probability equal to or higher than the threshold value. Further, the quantization table generation unit 125 controls the update speed by using the coefficient α calculated in the first embodiment.

[effect]
As described above, since the learning device 100 can execute the optimization of the quantized table using the conversation data set with high certainty, it is possible to suppress the convergence delay of the optimization of the quantized table, and appropriate data. The quantization table can be optimized with the optimum convergence rate using.

Here, a specific example of the conversation data set used for optimization (update) will be described. FIG. 13 is a diagram illustrating an example of distribution of a conversation data set. FIG. 13 illustrates using two normal conversation data sets collected at different contact centers. Each normal conversation data set is composed of 30 conversation data, calculates the average and variance of the feature amount (MFCC) for each conversation data, and further calculates the variance in the entire data set.

FIG. 13A shows an example of calculating the variance in a conversation data set using the average for each conversation data. That is, it is the average variance mv (j) of the conversation data in the conversation data set. FIG. 13B shows an example of calculating the variance in a conversation data set using the variance for each conversation data. That is, the variance vv (j) of the variance of the conversation data in the conversation dataset. As shown in FIGS. 13 (a) and 13 (b), the B-line conversation data set has a smaller variance than the R-line conversation data set, and is suitable for updating the quantization table.

The result of updating the quantization table will be described using the conversation data set corresponding to each line shown in FIG. FIG. 14 is a diagram illustrating an example in which the quantization table is updated with the B-line conversation data set in FIG. 13, and FIG. 15 is a diagram showing the update of the quantization table with the R-line conversation data set in FIG. It is a figure explaining the execution example.

In FIG. 14, the conversation data set corresponding to line B in FIG. 13 is used, and the quantization table updated with the features of the normal conversation is used, and about 200 conversations (voice section) each of the normal conversation and the abnormal conversation for evaluation are used. ) Is quantized and the mean and variance are shown. In FIG. 15, the conversation data set corresponding to the R line in FIG. 13 is used, and the quantization table updated with the features of the normal conversation is used. ) Is quantized and the mean and variance are shown.

As shown in FIG. 14 (a), the average of the quantization results obtained by quantizing the normal conversation and the anomalous conversation were quantized using the quantization table optimized using the conversation data set corresponding to the B line. The correlation with the average of the quantization results was "0.68". Further, as shown in FIG. 14 (b), the dispersion of the quantization result and the anomalous conversation obtained by quantizing the normal conversation are quantized by using the quantization table optimized by using the conversation data set corresponding to the B line. The correlation with the dispersion of the quantized result was "0.15".

Similarly, as shown in FIG. 15 (a), the average of the quantization results obtained by quantizing the normal conversation and the abnormal conversation are obtained using the quantization table optimized by using the conversation data set corresponding to the R line. The correlation with the average of the quantized quantized results was "0.95". Further, as shown in FIG. 15 (b), the dispersion of the quantization result obtained by quantizing the normal conversation and the anomalous conversation are quantized by using the quantization table optimized by using the conversation data set corresponding to the R line. The correlation with the dispersion of the quantized result was "0.79".

Therefore, the average correlation is "0.68" for the B line and "0.95" for the R line, and the difference in the distribution of the quantization results is larger in the quantization table updated using the B line. big. The variance correlation is "0.15" for the B line and "0.79" for the R line, and the difference in the distribution of the quantization results is larger in the quantization table updated using the B line. big. Therefore, it is better to perform conversation analysis using a quantization table updated with a conversation data set with high certainty, so that normal conversation and abnormal conversation can be accurately classified.

As a result of evaluating the analysis accuracy of the conversation situation, it was confirmed that the analysis accuracy was higher when the conversation data set (line B) having a small variance in the data set was used for updating the quantization table. For example, when updating with a conversation data set (B line) with a small variance in the data set, the analysis accuracy was about 90%, but when updating with a conversation data set (R line) with a large variance in the data set, the analysis accuracy was about 90%. The analysis accuracy was about 86%.

By the way, in the second embodiment, an example of selecting the conversation data set having a high “certainty” as described in the first embodiment for updating (applying) the quantization table has been described, but the present invention is not limited to this. It is also possible to determine the priority of application based on "certainty".

FIG. 16 is a flowchart showing the flow of processing according to the third embodiment. As shown in FIG. 13, when the processing start is instructed (S201: Yes), the acoustic analysis unit 61 executes an acoustic analysis for calculating the feature amount of the conversation data in the analysis frame unit (S202).

Subsequently, the conversation statistics unit 62 calculates the average and variance for each conversation data using the feature amount of each analysis frame unit (S203). Then, the overall statistics unit 63 calculates the average and variance in each update conversation data set by using the average and variance for each conversation data (S204).

After that, the calculation unit 64 calculates the certainty of each conversation data set using the average and variance in each update conversation data set (S205), and uses the certainty of each conversation data set to use the conversation for updating. The order of the data sets is determined (S206).

FIG. 17 is a diagram illustrating an update process according to the third embodiment. As shown in FIG. 17, it is assumed that the certainty 1 to m is calculated for the conversation data set 1 to m for update. However, it is assumed that the certainty 1 <... <the certainty m. In this case, the quantization table is updated by reading in order from the conversation data set 1 for update in ascending order of certainty. As a result, the data having a small variance can be updated later, and the convergence speed can be increased.

By the way, although the examples of the present invention have been described so far, the present invention may be carried out in various different forms other than the above-mentioned examples.

[Application example]
An application example of an analysis system using a determination model trained (machine learning) using the quantization table described in Example 1-3 will be described with reference to FIG. FIG. 18 is a diagram illustrating an application example. As shown in FIG. 18, the analysis system includes a GW (Gatway) 2, a SW (Switch) 3, an IP contact center system 4, an operator terminal 5, and an administrator terminal 6. The operator terminal 5 is equipped with a trouble detection function, and the administrator terminal 6 is equipped with a function of linking with a monitoring application.

The IP contact center system 4 receives a telephone call from a client via GW2 or SW3 and connects to the operator terminal 5. The operator terminal 5 analyzes the conversation with the client by using the determination model (LSTM model 110c and DNN model 110d) learned by using the quantization table.

Here, the operator terminal 5 analyzes the conversation online, and when a response trouble (abnormal call) is detected, outputs a detection message to the administrator terminal 6. The administrator terminal 6 can immediately support the operator's response by displaying the situation in real time by call monitoring or the like so that the administrator can confirm it.

[Numerical values, etc.]
The numerical examples, threshold values, and the like used in the above examples are merely examples and can be arbitrarily changed. It is also possible to use only one of the variance and the mean. Further, in the present embodiment, the specific conversation situation has been described as an "abnormal conversation situation", but the specific conversation situation is not limited to the abnormal conversation situation. For example, the specific conversation situation may be a conversation situation in which the meeting is stagnant, a conversation situation in which a trouble occurs, a conversation situation favorable to the customer, and the like. Further, the quantization point is not limited to the two-dimensional vector, and a multidimensional vector can be used.

[system]
Information including processing procedures, control procedures, specific names, various data and parameters shown in the above documents and drawings can be arbitrarily changed unless otherwise specified. The conversation statistics unit 62 and the overall statistics unit 63 correspond to the calculation unit, and the calculation unit 64 corresponds to the execution unit. The LSTM model 110c and the DNN model 110d are examples of the determination unit. The conversation data is an example of sample data, and the conversation data set is an example of a data set.

Further, each component of each of the illustrated devices is a functional concept, and does not necessarily have to be physically configured as shown in the figure. That is, the specific form of distribution or integration of each device is not limited to the one shown in the figure. That is, all or a part thereof can be functionally or physically distributed / integrated in any unit according to various loads, usage conditions, and the like. For example, the control unit 105 of the learning device 100 has the same function as the control unit 205 of the determination device 200, learns the LSTM model 110c and the DNN model 110d, and whether or not the voice data includes a specific conversation situation. May be determined.

Further, each processing function performed by each device may be realized by a CPU and a program analyzed and executed by the CPU, or may be realized as hardware by wired logic.

[hardware]
Next, a hardware configuration example of each device described in the above embodiment will be described. Since each device has the same hardware configuration, the information processing device 500 will be described here. FIG. 19 is a diagram illustrating a hardware configuration example. As shown in FIG. 19, the information processing device 500 includes a communication device 500a, an HDD (Hard Disk Drive) 500b, a memory 500c, and a processor 500d. Further, the parts shown in FIG. 19 are connected to each other by a bus or the like.

The communication device 500a is a network interface card or the like, and communicates with other servers. The HDD 500b stores a program or DB that operates the functions shown in FIG. 2 and the like.

The processor 500d reads a program that executes the same processing as each processing unit shown in FIG. 2 or the like from the HDD 500b or the like and expands the program into the memory 500c to operate a process that executes each function described in FIG. 2 or the like. .. For example, taking the information processing apparatus 50 as an example, this process executes the same functions as each processing unit of the information processing apparatus 50. Specifically, the processor 500d reads a program having the same functions as the acoustic analysis unit 61, the conversation statistics unit 62, the overall statistics unit 63, the calculation unit 64, etc. from the HDD 500b or the like. Then, the processor 500d executes a process of executing the same processing as the acoustic analysis unit 61, the conversation statistics unit 62, the overall statistics unit 63, the calculation unit 64, and the like.

In this way, the information processing apparatus 500 operates as an information processing apparatus that executes various processing methods by reading and executing the program. Further, the information processing apparatus 500 can realize the same function as that of the above-described embodiment by reading the program from the recording medium by the medium reader and executing the read program. The program referred to in the other embodiment is not limited to being executed by the information processing apparatus 500. For example, the present invention can be similarly applied when other computers or servers execute programs, or when they execute programs in cooperation with each other.

This program can be distributed over networks such as the Internet. In addition, this program is recorded on a computer-readable recording medium such as a hard disk, flexible disk (FD), CD-ROM, MO (Magneto-Optical disk), or DVD (Digital Versatile Disc), and is recorded from the recording medium by the computer. It can be executed by being read.

50 Information processing device 51 Communication unit 52 Storage unit 53 Conversation data set for update 54 Accuracy DB
60 Control unit 61 Acoustic analysis unit 62 Conversation statistics unit 63 Overall statistics unit 64 Calculation unit

Claims (7)

On the computer
For a plurality of data sets, each of which contains a plurality of sample data, the statistic of the feature amount of the plurality of sample data is calculated.
Based on the calculated statistic, the data set is selected or the priority is determined as training data to be used for training the judgment model for the plurality of data sets.
A decision program characterized by executing a process. In the calculation process, the variance of the feature amount is calculated using the feature amount for each predetermined unit for each of the plurality of sample data.
In the process to be executed, a data set having sample data whose variance is less than the threshold calculated for each of the plurality of sample data is selected as the training data, or the priority is determined in ascending order of the variance. The decision program according to claim 1, wherein the determination program is performed. In the calculation process, the variance calculated for each of the plurality of sample data possessed by each of the plurality of data sets is used to calculate the variance of the feature amount for each of the plurality of data sets.
The decision program according to claim 2, wherein the process to be executed selects a data set having a variance less than a threshold value as the training data, or executes the determination of the priority in ascending order of the variance. .. For a quantization table containing a plurality of quantization points, the selectivity of each of the plurality of quantization points was calculated based on the quantization data obtained by quantizing the feature quantities of the plurality of speech data included in the training data set. By updating a plurality of quantization points based on the selectivity, the computer is made to perform a process of updating the quantization table.
The determination according to claim 3, wherein the updating process uses a data set having a variance less than a threshold value as the training data set to calculate the selectivity of each of the plurality of quantization points. program. For a quantization table containing a plurality of quantization points, the selectivity of each of the plurality of quantization points was calculated based on the quantization data obtained by quantizing the feature quantities of the plurality of speech data included in the training data set. By updating a plurality of quantization points based on the selectivity, the computer is made to perform a process of updating the quantization table.
The third aspect of the present invention is characterized in that the updating process selects the plurality of data sets as the training data sets in ascending order of priority and calculates the selectivity of each of the plurality of quantization points. Described decision program. The computer
For a plurality of data sets, each of which contains a plurality of sample data, the statistic of the feature amount of the plurality of sample data is calculated.
Based on the calculated statistic, the data set is selected or the priority is determined as training data to be used for training the judgment model for the plurality of data sets.
A decision method characterized by performing an operation. For a plurality of data sets each containing a plurality of sample data, a calculation unit for calculating a statistic of the feature amount of the plurality of sample data, and a calculation unit.
An execution unit that executes selection or priority determination of a data set as training data used for training a judgment model for the plurality of data sets based on the calculated statistic.
An information processing device characterized by having.

Images