JP7216627B2 - INPUT SUPPORT METHOD, INPUT SUPPORT SYSTEM, AND PROGRAM - Google Patents

Description

An embodiment of the present invention relates to an input support system, an input support method, and a program.

2. Description of the Related Art Conventionally, in a call center, an employee (hereinafter referred to as an operator) responds to an inquiry, inquiry, or request from a customer. An operator who has finished responding to a customer by telephone inputs the progress of the customer response into a system that manages the case so that the information can be shared with other operators.

Patent Literature 1 discloses a system for displaying, in response to an inquiry from a customer, the type of incident associated with the inquiry and the content of actions that can be taken in association with the type of incident in a call center. Further, Patent Literature 1 describes that an operator inputs into the system how he/she responded to an inquiry from a customer by selecting from a pull-down menu. By inputting the progress of the processing of the matter into the system by the operator, even if another operator takes over the matter, the other operator can carry out the rest of the processing.

JP 2019-8722 A

Actions that can be taken differ depending on the type of matter handled by the call center, and actions that can be taken at the stage of dealing with the matter also differ. Therefore, there are cases in which there are dozens or more, or hundreds or more, of fixed phrases of the progress status to be input to an application for managing matters (hereinafter referred to as matter management application). When an operator selects a target fixed phrase using a pull-down method, it becomes difficult to find the target fixed phrase from hundreds or more of items. In addition, it takes a long time to input to the matter management application when inputting text instead of searching for the target fixed form using a pull-down method. Further, if the operator's skill level is low, more time is required.

In view of the above problems, an object of one embodiment of the present invention is to provide an input support method when an operator inputs fixed form sentences for registration in a case management application.

An input support method according to an embodiment of the present invention extracts words and characters included in speech text data, and extracts one word, a plurality of consecutive words, and one character and a plurality of consecutive characters. A feature vector weighted by appearance frequency and rarity is generated, and when the feature vector is input, a machine learning model generates candidates corresponding to the contents of the call text data to be registered in the application that manages the matter. Output fixed phrases.

In the above method, the machine learning model includes logistic regression, neural network.

In the above method, the machine learning model is updated based on the feature vectors and the fixed phrases selected for the speech text data.

In the above method, the feature vector is obtained by extracting text data uttered by the operator from the call text data, extracting words and characters contained in the text data uttered by the operator, and extracting one word and a plurality of consecutive words, Also, one character and a plurality of consecutive characters are weighted by appearance frequency and rarity.

In the above method, the call text data includes a tag indicating the start of the call and a tag indicating the end of the call.

In the above method, the call text data includes a tag indicating the start of the call and a delimiter tag indicating the middle of the call. Each time the call text data is acquired, the call text data is converted into a feature vector, and the feature vector is input By the calculation of the machine learning model as described above, the fixed sentences of the candidates corresponding to the speech text data output to the application are updated.

A program for causing a computer to execute an input support method according to an embodiment of the present invention.

An input support system according to an embodiment of the present invention includes a text data conversion unit that converts call voice data into text, a text data acquisition unit that acquires call text data, and extracts words and characters included in the call text data. a preprocessing unit that generates feature vectors weighted by appearance frequency and rarity for each of one word, a plurality of consecutive words, and a single character and a plurality of consecutive characters; and a classifier that, when received, outputs candidate boilerplate sentences corresponding to the content of the call text data for registration in an application that manages cases by means of a machine learning model.

In the above system, the text data conversion unit is included in the file server, and the text data acquisition unit, preprocessing unit, and classifier are included in the analysis server.

According to one embodiment of the present invention, it is possible to reduce the search time for a target fixed phrase when an operator inputs a fixed phrase to be registered in a case management application. As a result, it is possible to significantly reduce the time required to search for the target fixed phrase, thereby improving work efficiency.

1 is a block diagram of an input support system according to one embodiment of the present invention; FIG. It is an example of call text data. It is a figure which shows the hardware constitutions of an analysis server. 3 is a block diagram of an analysis server; FIG. 4 is a flow chart illustrating the learning stage; 4 is a flow chart illustrating a method of generating word vectors; 4 is a flow chart illustrating a method of generating a character vector; 3 is a block diagram of an analysis server; FIG. 4 is a flow chart describing the estimation stage; It is an example of fixed phrases displayed on the screen of the terminal device. It is an example of call text data. 4 is a flow chart describing the estimation stage; 1 is a block diagram of an input support system according to one embodiment of the present invention; FIG.

An embodiment of the present invention will be described below with reference to the drawings. The embodiments shown below are examples of embodiments of the present invention, and the present invention is not limited to these embodiments. In the drawings referred to in this embodiment, the same parts or parts having similar functions are denoted by the same reference numerals or similar reference numerals (reference numerals followed by A, B, etc.). may be omitted.

(First embodiment)
In this embodiment, an input support system 1 according to one embodiment of the present invention will be described with reference to FIGS. 1 to 10. FIG.

[Overview of input support system]
First, an overview of an input support system 1 according to an embodiment of the present invention will be described. FIG. 1 is a block diagram of an input support system 1 according to one embodiment of the invention. The input support system 1 infers a fixed phrase indicating appropriate progress of a project from the content of conversation between a customer and an operator in call center operations. In this way, input support for a fixed phrase to be registered in the matter management application is provided.

In the call center business, when a call is received from the customer call terminal 41 to the call center, it is connected to the operator call terminal 30. - 特許庁The operator call terminal 30 records the conversation between the customer and the operator as call voice data. Call voice data is recorded in the file server 10 . When the connection between the customer call terminal 41 and the operator call terminal 30 is disconnected, the call voice data acquired by the file server 10 is converted into call text data. The file server 10 transmits call text data to the analysis server 20 . When the analysis server 20 acquires the speech text data, it generates feature vectors in which word vectors and character vectors are integrated from the speech text data. In this specification and the like, a feature vector refers to a vector representing the content of a document based on the words contained therein and the degree of importance of each speech text data. When the feature vector is input to the classifier 24 , it outputs to the Web server 25 candidate phrases to be registered in the case management application corresponding to the call text data by means of a learned machine learning model. The candidate fixed phrases output to the web server 25 are displayed on the terminal device 32 via the thin client server 33 . When a fixed phrase corresponding to call text data is selected from the candidate fixed phrases in the terminal device 32, the selected fixed phrase can be input to the case management application.

A machine learning model of the classifier 24 is generated in advance by machine learning based on a data set of feature vectors generated from speech text data and a plurality of fixed phrases. Therefore, when the analysis server 20 acquires the call text data, the machine learning model can estimate candidate fixed phrases corresponding to the call text data. As a result, the operator can select a fixed phrase corresponding to the content of the call with the customer from among the candidate fixed phrases displayed on the terminal device 32, so that there are dozens or more or hundreds of fixed phrases. It is possible to shorten the search time for fixed phrases compared to searching for fixed phrases from among. As a result, regardless of the skill level of the operator, it is possible to significantly reduce the time required to search for the target fixed phrase, thereby improving operational efficiency.

The input support system 1 described above can be applied to various businesses that handle call voice data. The input support system 1 can be applied to, for example, insurance business, banking business, and sales business in which customers and operators talk over each other. Here, a call means a period from when the customer call terminal 41 and the operator call terminal 30 are connected to when they are disconnected.

Hereinafter, the configuration of the input support system 1 according to one embodiment of the present invention will be described in detail. Also, a call center service related to insurance will be described as an example of a service that handles call voice data.

[Configuration of input support system]
The input support system 1 shown in FIG. 1 has at least a file server 10 and an analysis server 20 . The input support system 1 may further include an exchange 31 , an operator call terminal 30 , a terminal device 32 and a thin client server 33 . In FIG. 1, file server 10 is connected to analysis server 20 via communication network 52 .

In the call center, the exchange 31 is communicably connected to the customer call terminal 41 via the communication network 51 . The communication network 51 is a public network such as the Internet or PSTN (Public Switched Telephone Networks), a wireless network, or the like. Also, the exchange 31 is connected to a plurality of operator call terminals 30 . When receiving a call from the customer call terminal 41 , the exchange 31 connects to one of the plurality of operator call terminals 30 .

In the data center, the file server 10 has a call data acquisition section 11 , a text data conversion section 12 and a storage section 13 . When the operator call terminal 30 and the customer call terminal 41 are connected, the call data acquisition unit 11 records the conversation between the customer and the operator in the storage unit 13 as call voice data. The call voice data may be given a call start time, a call end time, a tag indicating the start of the call, a tag indicating the start of the end of the call, and identification information for distinguishing the call voice data. The operator call terminal 30 may separately record the customer's call voice data and the operator's call voice data.

When the text data conversion unit 12 acquires the call voice data, it converts it into call text data by voice recognition processing. FIG. 2 is an example of call text data 400 . As shown in FIG. 2, in the call text data 400, customer tags are added to the customer's utterances, and operator's tags are added to the operator's utterances. Also, a call start time, a call end time, a tag indicating the start of a call, a tag indicating the start of the end of a call, and identification information for distinguishing call text data may be added. The speech recognition process may be executed in real time while the customer call terminal 41 and the operator call terminal 30 are connected, or may be executed after the customer call terminal 41 and the operator call terminal 30 are disconnected. processing may be performed. In this embodiment, a well-known technique may be used for this speech recognition processing, and the speech recognition processing itself and various speech recognition parameters used in the speech recognition processing are not particularly limited. Also, the call text data is stored in the storage unit 13 .

The analysis server 20 has a text data acquisition unit 21 , a preprocessing unit 22 , a storage unit 23 , a classifier 24 and a web server 25 .

The text data acquisition unit 21 acquires call text data from the text data conversion unit 12 via the communication network 52 .

The preprocessing unit 22 preprocesses the speech text data to generate feature vectors.

The feature vector generated from the speech text data is stored in the storage unit 23 as learning data. In addition, the storage unit 23 stores all kinds of fixed phrases to be registered in the matter management application. For example, several tens or more types of standard sentences are set for each of the first to third stages. The fixed phrases are combined into one set of the fixed phrases selected in the first step, the fixed phrases selected in the second step, and the fixed phrases selected in the third step. It should be noted that the set of candidate fixed phrases should include at least the fixed phrases selected in the first stage. Alternatively, it may be a combination of the fixed phrase selected in the first step and the fixed phrase selected in the second step. Here, it is assumed that there are 300 types of fixed phrase sets. Further, the storage unit 23 stores standard sentence data, which is annotated metadata (a label giving "correct answer") corresponding to the learning data. Further, the storage unit 23 stores the obtained call text data, the machine learning model, the program for executing the input support method according to the embodiment of the present invention, and the like.

When the feature vector is input, the classifier 24 outputs the probability of occurrence of each of a plurality of fixed phrases (sets) to be registered in the case management application by means of a learned machine learning model. Machine learning algorithms include logistic regression, neural networks, support vector machines, random forests, naive Bayes, and the like. In this embodiment, logistic regression is used as a machine learning algorithm. Logistic regression uses a sigmoid function as the output of the model. By using a sigmoid function that maps an arbitrary value between 0 and 1, the probability of whether the given data is positive (+1) or negative (0) is determined. The sigmoid function classifies positive examples and negative examples by a threshold. It is possible to estimate, as candidates, fixed phrases having occurrence probabilities equal to or higher than a predetermined threshold value (for example, 0.8) with respect to occurrence probabilities for each fixed phrase output by the sigmoid function. As another method, it is also possible to estimate a plurality of standardized sentences (for example, top five standardized sentences) with the highest probability of occurrence as candidates without setting a threshold.

The Web server 25 outputs candidate fixed form sentences. The web server 25 is connected to the thin client server 33 via the communication network 52 . The thin client server 33 holds various software (applications) used by the terminal device 32 . When the thin client server 33 runs a virtual desktop on the terminal device 32, it holds an application to run on the virtual desktop. One of the applications is a case management application. The terminal device 32 displays the results output from the web server 25 via the thin client server 33 .

When a fixed phrase is selected from the candidate fixed phrases displayed on the terminal device 32, the fixed phrase corresponding to the content of the call with the customer can be registered in the project management application held in the thin client server 33. can.

[Analysis server hardware configuration]
FIG. 3 is a block diagram for explaining the hardware configuration of the analysis server 20. As shown in FIG. As shown in FIG. 3 , the analysis server 20 has a control unit 201 , memory 202 , storage unit 203 and communication unit 204 .

The control unit 201, for example, CPU (for example, a multi-processor having a plurality of processor cores), GPUs (Graphics Processing Units), DSPs (Digital Signal Processors), FPGAs (Field-Programmable Gate Arrays), etc. can be combined. Consists of Further, it is preferable to use a GPGPU (General-Purpose Computing on Graphics Processing Units) as the control unit 201 in order to enable faster arithmetic processing.

The memory 202 uses memories such as ROM (Read Only Memory) and RAM (Random Access Memory). The ROM is a non-volatile storage unit in which control programs for causing the control unit 201 to execute various processes are stored in advance. The RAM is a volatile or nonvolatile memory that stores various types of information, and is used as a temporary storage memory (work area) for various types of processing executed by the control unit 201 .

The storage unit 203 is composed of a rewritable non-volatile recording medium such as an HDD (Hard Disk Drive) or an SSD (Solid State Drive). The storage unit 203 stores call text data, learning data, fixed phrase data, machine learning models, programs for causing a computer to execute the input support method according to one embodiment of the present invention, etc., which are stored in the storage unit 23. be done. These data may be stored in different storage media.

The communication unit 204 can use a communication device such as a network card for wired communication, a wireless communication device that connects to the file server 10, or a wireless communication device that supports connection to an access point. The control unit 201 is capable of communication connection or information transmission/reception with the file server 10 or the thin client server 33 via the communication network 51 using the communication unit 204 .

The analysis server 20 may have at least the above hardware configuration, and cloud services provided by companies, such as AWS (registered trademark) (Amazon Web Services (registered trademark)), Azure (registered trademark) (Microsoft ( (registered trademark)), GCP (registered trademark) (Google Cloud Platform (registered trademark)).

[Flowchart of input support method]
Next, an input support method according to one embodiment of the present invention will be described with reference to FIGS. 4 to 10. FIG. The analysis server 20 causes the control unit 201 to execute each process.

[Learning stage]
First, the learning stage of fixed phrases will be described. FIG. 4 shows functional blocks of the analysis server 20 used in the learning stage of the machine learning model. FIG. 4 shows functional blocks of the text data acquisition unit 21, the preprocessing unit 22, the storage unit 23, and the classifier 24 of the analysis server 20. As shown in FIG. FIG. 5 is a flow chart illustrating the learning stage.

In the learning stage, first, the text data acquisition unit 21 acquires learning call text data (step S301). The training call text data is call text data accumulated from the past. The learning speech text data is acquired from the storage unit 13 of the file server 10, for example. In this embodiment, it is assumed that there are 20,000 learning call text data. Here, if the speech text data includes numerical values having characteristics, it is preferable to convert all the numerical values to "0" in order to make the numerical information into a single characteristic.

Next, the preprocessing unit 22 performs preprocessing on the speech text data for learning and converts it into a word vector (step S302). FIG. 6 shows a method of generating word vectors from learning speech text data.

First, as shown in FIG. 6, morphological analysis is performed on the acquired speech text data to perform part-of-speech decomposition (step S311). Morphological analysis can be performed using various programs such as MeCab and JUMAN++, and libraries as appropriate. An example of sentences included in the speech text data before morphological analysis is shown below.

An example of sentences included in the speech text data after morphological analysis is shown below. By performing morphological analysis on the speech text data, the sentence is divided into morphemes and the parts of speech and variations of each are determined.

Next, word data is extracted by cleaning the morphologically analyzed speech text data (step S312). Cleaning refers to the process of removing preset characters or words based on predetermined selection criteria. Here, words of nouns, verbs and adjectives are extracted by removing particles and symbols contained in the speech text data. The cleaning process may also remove words that are too common to be useful in classifying natural language as stop words. Stop words are, for example, words such as "do" and "masu". Also, stop words may be registered in a dictionary according to the results of machine learning. Stopwords registered in the dictionary may be removed during the cleaning process. In addition to these, noise and the like may be removed by cleaning processing. Data obtained by removing particles and symbols from the speech text data is shown below. "Masu" and "Do" shown below correspond to stop words.

Below is the data with the stopwords removed from the data shown. This extracts the necessary word data.

Next, the word data is converted into word vectors (step S313). First, the word data is divided. Word segmentation is performed by, for example, N-grams. N-gram is a text division method for dividing an arbitrary character string or document into consecutive n characters or words. For example, a case where n is 1 is called a uni-gram, a case where n is 2 is called a bi-gram, and a case where n is 3 is called a tri-gram. An example of dividing word data by tri-grams is shown below.

The results of vectorization by counting the frequency of appearance of one or a plurality of consecutive words divided by tri-grams are shown below.

Next, a weighting process is performed on one or a plurality of consecutive words to generate a weighted word vector (step S314). Weighting processing is performed by, for example, tf-idf. tf-idf is an index used in fields such as information search and text mining to identify how characteristic a specific word appears in a document. Here, tf (term frequency) represents the number of occurrences of a particular word in the document, and idf (inverse document frequency) represents the natural number of documents including that document in the entire corpus. It represents the logarithm (rarity), and "tf×idf" is the tf-idf value of that word in that document.

In this embodiment, for example, a corpus is created from 200,000 call text data. tf-idf can suppress the influence of non-call-specific words such as "me" and "yesterday". In addition, by analyzing the entire 200,000 call text data, it is possible to select words that are important for classification of fixed phrases from each call text data. In addition, even if the word has a high appearance frequency, the word appearing in any speech text data is adjusted so that the rarity level is low. Below is an example of a weighted word vector.

Next, among the weighted word vectors, the word vectors with higher importance are extracted (step S315). In this embodiment, for example, 20,000 word vectors with high importance are extracted. As described above, speech text data can be converted into a word vector.

Next, the preprocessing unit 22 performs preprocessing on the speech text data for learning and converts it into a character vector (step S303). FIG. 7 shows a method of generating a character vector from learning speech text data. It should be noted that, if the speech text data includes numerical values having characteristics, it is preferable to convert all the numerical values to "0" at this stage in order to make the numerical information into a single characteristic.

First, character data is extracted from the speech text data for learning by data cleaning (step S321). Data cleaning for character data is performed in a manner similar to cleaning word data.

Next, the character data is converted into a character vector (step S322). Character segmentation is performed by N-grams in the same manner as word segmentation. In this embodiment, character data is divided into tri-grams. Then, the appearance frequency of one or a plurality of consecutive characters is counted to convert to a character vector.

Next, a weighting process is performed on the character vector to generate a weighted character vector (step S323). Weighting processing is performed by tf-idf in the same way as word vectors.

Next, among the weighted character vectors, character vectors with higher importance are extracted (step S324). In this embodiment, for example, 60,000 character vectors with high importance are extracted. As described above, a character vector can be extracted from the speech text data. Note that the order of the conversion into word vectors in step S302 and the conversion into character vectors in step S303 may be reversed.

Next, a feature vector is generated by integrating the word vector and the character vector (step S304). In this embodiment, the feature vector has 80,000 columns, all of which have a fixed length. Through the processing described above, feature vectors can be generated from learning speech text data. By converting all the learning speech text data into feature vectors, the feature vectors of 20,000 rows and 80,000 columns are used as learning data 27 .

Next, the feature vector obtained in step S304 is stored in the storage unit 23 as learning data 27 (step S305).

Next, the machine learning model is machine-learned based on the learning data 27 and the standard sentence data 28 (step S306). The standard sentence data 28 is stored in the storage unit 23 . The fixed phrase data 28 is annotated metadata (a label giving "correct answer"). The learning data 27 and fixed phrase data 28 are also called a data set 25 . 20,000 data sets 25 are stored in the storage unit 23 corresponding to the training speech text data. A learned machine learning model can be generated by performing machine learning using a data set 25 composed of learning data 27 and fixed phrase data 28 as teacher data.

Finally, the trained machine learning model is stored in the storage unit 23 (step S307).

[Estimation stage]
Next, the step of estimating fixed phrases will be described. FIG. 8 shows functional blocks of the analysis server 20 used in the estimation stage. FIG. 8 shows functional blocks of the text data acquisition unit 21, the preprocessing unit 22, and the classifier 24 of the analysis server 20. As shown in FIG. FIG. 9 is a flow chart for explaining the step of estimating fixed phrases.

In the estimation stage, first, the text data acquisition unit 21 acquires call text data from the file server 10 (step S331). Here, call text data is call text data newly generated by a call between a customer and an operator. In this embodiment, the call text data is analyzed after the call between the customer and the operator is completed.

Next, the preprocessing unit 22 performs preprocessing on the call text data and converts it into a feature vector (step S332). The method of converting speech text data into feature vectors may follow the flowcharts described in FIGS. 6 and 7. FIG.

Next, the classifier 24 inputs the feature vector as an input vector to the machine learning model 29 (step S333). The machine learning model 29 is the learned machine learning model described in the learning stage. As input vectors, a total of 80,000 feature vectors including 20,000 word vectors and 60,000 character vectors are input to the machine learning model 29 . In this embodiment, a sigmoid function is used as the output of the machine learning model. As a result, occurrence probabilities are output as output vectors for all fixed phrases. For example, the probability of occurrence of fixed phrase 1 is 3%, the probability of occurrence of fixed phrase 2 is 32%, the probability of occurrence of fixed phrase 3 is 98%, and so on. A standard sentence having an occurrence probability equal to or greater than a predetermined threshold value (for example, 0.8) may be estimated as a candidate for the occurrence probability of each standard sentence output by the sigmoid function. In this embodiment, no threshold is set for the output of the sigmoid function, and a plurality of standardized sentences (for example, top five standardized sentences) with the highest probability of occurrence are estimated as candidates.

Finally, the classifier 24 outputs the candidate fixed phrases to the Web server 25 (step S334). Candidate fixed phrases to be output to the Web server are, for example, output in descending order of the probability of occurrence of fixed phrases.

When the candidate fixed phrases are output to the Web server 25 , at least one of the high occurrence probabilities is output from the terminal device 32 to the screen of the terminal device 32 via the thin client server 33 .

FIG. 10 shows an example of fixed phrases displayed on the screen of the terminal device 32. As shown in FIG. As shown in FIG. 10, window 300 displays five candidate fixed phrases. It should be noted that the candidate fixed phrases displayed in window 300 are examples of fixed phrases that may correspond to the call text data shown in FIG.

A window 300 shown in FIG. 3 displays a web page address 301, additional information 302 of call text data, a case number 303, a registration/update button 303, a scroll bar 304, and a page forward button 306. FIG. Further, in the window 300, candidate fixed sentences 311 to 315 are displayed together with their occurrence probabilities. Candidate standardized sentences 311 to 315 are five candidate standardized sentences selected in descending order of probability of occurrence. Candidate fixed phrases are displayed as one set by combining the fixed phrases selected in the first step, the fixed phrases selected in the second step, and the fixed phrases selected in the third step. It should be noted that at least the fixed phrases selected in the first stage should be included in the candidate fixed phrases. Alternatively, it may be a combination of the fixed phrase selected in the first step and the fixed phrase selected in the second step.

The operator selects a candidate selection sentence with the cursor 305 . The operator selects one of the candidate fixed phrases displayed in the window 300 and selects the register/update button 303 to register the selected fixed phrase in the matter management application. can be done. Further, if there is no appropriate fixed phrase among the presented candidate fixed phrases, a plurality of next candidate fixed phrases may be displayed by selecting the page turn button 306 . In this way, the terminal device 32 can display an appropriate candidate fixed phrase according to the content of the call text data.

[Relearn]
In the classifier 24, re-learning may be performed every predetermined period to update the machine learning model 29. FIG. The storage unit 23 stores the acquired call text data in the storage unit 23 . Further, feature vectors converted from speech text data are accumulated in the storage unit 23 as learning data 27 . At this time, the fixed phrase selected as the fixed phrase corresponding to the call text data is given a label giving "correct answer" and stored in the storage unit 23 as fixed phrase data 28 .

The classifier 24 updates the machine learning model 29 by re-learning based on the accumulated learning data 27 and fixed phrase data 28 . As a result, the updated machine learning model 29 can improve the accuracy of estimating candidate fixed phrases that could not be accurately estimated by the previous machine learning model 29 . Note that the timing of updating the machine learning model 29 is not limited to every predetermined period, and may be executed at the timing when a predetermined number of data sets 25 of the learning data 27 and fixed phrase data 28 are accumulated in the storage unit 23. .

(Second embodiment)
In the first embodiment, the method of transmitting call text data to the analysis server 20 after the call between the customer and the operator is finished and estimating a fixed phrase indicating the progress has been described, but one implementation of the present invention. The form is not limited to this. Even during a call between the customer and the operator, call voice data may be converted into call text data in real time, and the call text data may be analyzed sequentially.

In this embodiment, in the middle of a call between a customer and an operator, by converting call voice data into call text data and analyzing the call text data, a method of updating candidate phrases as time elapses, Description will be made with reference to FIGS. 1, 11 and 12. FIG.

When the customer call terminal 41 and the operator call terminal start connection, the call data acquiring unit 11 starts acquiring call voice data. When acquisition of call voice data is started, a tag indicating call start (hereinafter referred to as a start tag) is added to the call voice data. Next, in the middle of the call between the customer and the operator, a tag indicating a break in the call (hereinafter referred to as a break tag) is attached to the call voice data. The tag indicating the delimitation of the call given to the call voice data may be given at the timing of switching between the customer's speech and the operator's speech, for example. Alternatively, a delimiter tag may be added every time a predetermined time elapses after a call start tag is added to the call voice data. Alternatively, a delimiter tag may be added after a predetermined period of time has elapsed since the voice was interrupted. When the connection between the operator call terminal 30 and the operator call terminal is terminated, a tag indicating the end of the call (hereinafter referred to as an end tag) is added to the call voice data.

The text data conversion unit 12 sequentially performs speech recognition processing on the call voice data and converts it into call text data. A start tag and a delimiter are added to the converted speech text data as well. The call text data is stored in the storage unit 13 until the end tag is added. FIG. 11 shows call text data 400A to which start tags, delimiter tags, and end tags have been added. In FIG. 11, the delimiter tag from the start tag to 30 seconds is call text data 401, the delimiter tag from the start tag to 1 minute is call text data 402, and the delimiter tag from the start tag to the end tag is call text data 403. and

Next, a method in which the analysis server 20 sequentially analyzes the acquired call text data and continues to output candidate fixed phrases until the end of the call will be described with reference to FIG. 12 . FIG. 11 is a flow chart for explaining a method for the analysis server 20 to estimate fixed phrases.

First, the text data acquisition unit 21 acquires the call text data 401 from the file server 10 (step S341). The text data acquisition unit 21 acquires the call text data during the conversation between the customer and the operator, that is, while the call text data is being accumulated in the file server 10 . A start tag and a delimiter tag are added to the acquired call text data 401 .

Next, the preprocessing unit 22 performs preprocessing on the call text data 401 and converts it into a feature vector (step S342). Note that even if the obtained speech text data is text data after the delimiter tag, the preprocessing is executed with the speech text data from the start tag to the delimiter tag. Upon detecting the delimiter tag, the preprocessing unit 22 converts the speech text data 401 from the start tag to the delimiter tag into a feature vector. Next, the classifier 24 inputs the feature vector as an input vector to the machine learning model 29 (step S343). Next, the classifier 24 outputs the candidate fixed phrases to the Web server 25 (step S344). By the above processing, it is possible to output a candidate fixed phrase during the conversation between the customer and the operator.

Next, the text data acquisition unit 21 acquires the call text data 402 from the file server 10 (step S341). The text data acquisition unit 21 acquires call text data each time a predetermined time elapses. A start tag, a delimiter tag, and an added delimiter tag are added to the acquired call text data 402 .

Next, the preprocessing unit 22 performs preprocessing on the call text data 402 and converts it into a feature vector (step S342). Upon detecting the added delimiter tag, the preprocessing unit 22 converts the speech text data 402 from the start tag to the added delimiter tag into a feature vector. Next, the classifier 24 inputs the feature vector as an input vector to the machine learning model 29 (step S343). Next, the classifier 24 outputs the candidate fixed phrases to the Web server 25 (step S344). With the above processing, candidate fixed phrases can be updated during a call between the customer and the operator.

The analysis server 20 repeats the estimation process shown in FIG. 11 until the conversation between the customer and the operator ends, that is, until the preprocessing unit 22 detects an end tag from the call text data. The amount of call text data acquired by the text data acquisition unit 21 increases as time passes. Therefore, as the amount of speech text data increases, the accuracy of output candidate fixed sentences can be improved. Furthermore, when the call between the customer and the operator ends, it is possible to output final candidate fixed phrases. Therefore, when the call ends, the operator can immediately enter a fixed phrase corresponding to the content of the call into the progress management application, thereby shortening the time required to register the fixed phrase in the progress management application. can.

(Modification 1)
In the above-described embodiment, the method of outputting a candidate standard sentence using the text data of the conversation between the customer and the operator has been described, but one embodiment of the present invention is not limited to this. Machine learning may be performed by extracting the operator's uttered text data from the call voice data between the customer and the operator. In this case, the preprocessing unit 22 extracts the operator's uttered text data, and then preprocesses the operator's uttered text to generate a feature vector. In a conversation between a customer and an operator, the operator often speaks more about the current status of the item than the customer. Therefore, the analysis processing time can be shortened by extracting only the text data spoken by the operator from the call text data and analyzing it.

(Modification 2)
In the above-described embodiment, the method of converting call voice data into call text data and then transmitting the data to the text data acquisition unit 21 of the analysis server 20 has been explained, but one embodiment of the present invention is not limited to this. The call voice data acquired by the file server 10 may be transmitted to the analysis server 20 and then converted into call text data by voice recognition processing. After that, the analysis server 20 converts the call text data into a feature vector and inputs it to the machine learning model 29 .

(Modification 3)
In the above-described embodiment, an example has been described in which a combination of fixed phrases selected in the first step to fixed phrases selected in the third step is set as one set, but one embodiment of the present invention. The form is not limited to this. The fixed phrase does not have to be a combination of the fixed phrases of the first to third stages, and may be one fixed phrase. Also, the standard sentence set is not limited to the first to third stages, and may be a combination of standard sentences of the fourth and higher stages.

(Modification 4)
In the above-described embodiment, an example in which the file server 10 and the analysis server 20 are separate in the input support system 1 has been described, but one embodiment of the present invention is not limited to this. As shown in FIG. 13, the input support system 1A may include a text data conversion section 12, a text data acquisition section 21, a preprocessing section 22, and a classifier 24. FIG.

1: Input Support System, 10: File Server, 11: Call Data Acquisition Unit, 12: Text Data Conversion Unit, 20: Analysis Server, 21: Text Data Acquisition Unit, 22: Preprocessing Unit, 23: Storage Unit, 24: Classifier, 25: Web server, 27: Learning data, 28: Fixed phrase data, 29: Machine learning model, 30: Operator call terminal, 31: Exchange, 32: Terminal device, 33: Thin client server, 41: Customer call terminal, 51: communication network, 52: communication network, 201: control unit, 202: memory, 203: storage unit, 204: communication unit

Claims (9)

Words and characters included in call text data are extracted, and feature vectors are weighted by frequency of appearance and rarity for each of a single word, a plurality of consecutive words, and a single character and a plurality of consecutive characters. to generate
outputting a plurality of selectable fixed form sentences corresponding to the contents of the call text data to the application for registration in an application for managing the case, by calculation of the machine learning model with the feature vector as input; Input assistance method. The input support method according to claim 1, wherein said machine learning model includes logistic regression and neural network. 3. The input support method according to claim 1, wherein said machine learning model is updated based on said feature vector and fixed form sentences selected for said speech text data. The feature vector is
Text data spoken by an operator is extracted from the call text data,
Words and characters included in the text data uttered by the operator are respectively extracted, and one word, a plurality of consecutive words, and one character and a plurality of consecutive characters are weighted by appearance frequency and rarity. 4. The input support method according to any one of claims 1 to 3, wherein: 5. The input support method according to any one of claims 1 to 4, wherein said call text data includes a tag indicating the start of a call and a tag indicating the end of said call. The call text data includes a tag indicating the start of the call and a delimiter tag indicating the middle of the call,
Each time the call text data is obtained, the call text data is converted into the feature vector, and a selection corresponding to the call text data output to the application is performed by a machine learning model operation using the feature vector as an input. 6. The input support method according to any one of claims 1 to 5, wherein a plurality of possible candidate fixed phrases are updated. The input support method according to any one of claims 1 to 6,
A program that makes a computer run. a text data conversion unit that converts call voice data into text;
a text data acquisition unit for acquiring call text data;
Extracting words and characters contained in the call text data, and weighting features of one word, a plurality of consecutive words, and a single character and a plurality of consecutive characters by appearance frequency and rarity a preprocessing unit that generates a vector;
Classification for outputting a plurality of selectable fixed form sentences corresponding to the content of the call text data to the application for registering in the application for managing the matter by calculation of the machine learning model with the feature vector as input. an input support system, including The text data conversion unit is included in a file server,
9. The input support system according to claim 8, wherein said text data acquisition unit, said preprocessing unit, and said classifier are included in an analysis server.

Images