JP6183841B2 - Call center term management system and method for grasping signs of NG word - Google Patents

Description

  The present invention relates to a call center system, and more particularly, to a system for preventing a call center operator from issuing an NG word and a method for preventing the same.

  In mass media such as television and radio, so-called broadcast-prohibited terms that are prohibited from being broadcast are defined, and attention is always paid so that performers do not inadvertently speak broadcast-prohibited terms. And even if a performer says a broadcast-prohibited term, it is cut when editing unless it is a live broadcast.

  Similarly, in usage situations such as telephones used by individuals and terminals used by employees of companies, prohibited terms or semi-prohibited terms that should be refrained depending on the situation (hereinafter collectively referred to as “NG word”). Will be called). For example, terms that make the other person unpleasant or misleading, terms that should not be used depending on the situation, such as explaining to customers, terms related to personal information such as addresses and telephone numbers, confidential information, or sensitive information, etc. It is.

  Therefore, in order to prevent the transmission of the NG word, a method (for example, Patent Document 1) of detecting an NG word issued by a user such as a telephone in real time and replacing the detected NG word with another voice is disclosed. ing. Also disclosed is a method (for example, Patent Document 2) that issues an alarm of a situation when an NG word is issued to alert and correct the situation.

JP 2010-103751 A JP2013-25609A

  In the above-mentioned Patent Document 1, a server connected to a telephone or a device that emits voice receives voices uttered by a user and temporarily accumulates them, and the accumulated voices are recognized and analyzed at a word level. Then, it is determined whether or not the analyzed speech at the word level matches a predetermined prohibited phoneme model, and a portion determined to match the prohibited word phoneme model among the voices uttered by the user is specified and specified. A method of replacing the recorded portion with a dummy sound is described.

  Further, in Patent Document 2 described above, when an NG word is detected in a speech recognition result converted into text by speech recognition in an explanation support system for supporting explanation given to a customer or the like using a computer terminal. , Display a message including the detected NG word on the terminal display, extract the voice recognition result of the sound collection time, send the explanation status at that time to another backyard terminal to alert and correct Is described.

  As for NG words, the same thing exists in the call center of a company that accepts calls from a large number of customers (including prospective customers), and there are NG words that the operator must not issue to customers who make calls. It has been established. However, if the method disclosed in Patent Document 1 is applied to a call center system, even if the NG word transmission itself can be prevented, a dummy voice (such as a beep) converted from the NG word is transmitted. , Customers who heard it will feel very uncomfortable. Further, in the method of Patent Document 2, even if alerting or correction can be performed after the NG word has been transmitted, the transmission of the NG word itself cannot be prevented in advance.

  In view of the above-described problems, an object of the present invention is to provide a system and a method for preventing an operator from inadvertently issuing an NG word to a customer in advance in a call center.

  In order to solve the above problems, the system of the present invention provides the following solution.

The invention according to claim 1 is a system for managing terms issued by an operator to a customer at a call center, the voice capture server recording a conversation between the customer and the operator, and the voice capture server recording A speech recognition server that reads out speech data, converts it into conversation text that is text data, and stores it in a conversation text storage DB; and a conversation text analysis server that statistically processes and analyzes the conversation text in the conversation text storage DB. The conversation text analysis server reads the conversation text from the conversation text storage DB, and decomposes it into term units by morphological analysis, and converts the conversation text into term texts decomposed into term units by the morphological analysis unit. Detects whether a predetermined NG word is included , When the NG word is included, the term included in the previous utterance of the NG words, the operator is tied or signs word or not to be a sign that emits the NG word for each of the NG word A symptom word utterance determination unit for determining based on the NG word symptom DB in which the symptom word is registered .

According to the arrangement, it is a Turkey be detected in advance signs Oite operator generates an NG word call center.

  The invention according to claim 2 is the invention according to claim 1, further comprising a conversation content monitoring server for acquiring voice data during a call between a customer and an operator and checking the conversation content in real time, The server detects, based on data stored in the NG word sign DB, whether or not the sign word is included in the conversation during the call, and the sign word is in the call And an alarm generation unit that generates an alarm at the terminal of the supervisor who manages the operator when it is included in the conversation.

  According to the above configuration, when the conversation content monitoring server monitors the conversation in progress in real time and detects that the above-mentioned indication word is included in the conversation, an alarm is immediately generated. It is possible to prevent the operator from inadvertently issuing an NG word to the customer in advance. That is, the operator can pay attention to his / her conversation, and the supervisor who manages the operator can pay attention to the operator with priority.

  According to a third aspect of the present invention, in the second aspect of the present invention, the conversation content monitoring server includes the NG word, a symptom word associated with the service content of the call center, and the conversation of the conversation during the call. An operator based on at least one of time, information on whether the calling customer is designated as a specific customer, and information on whether the calling customer is a customer in a specific area NG word risk, which is the degree of risk of issuing an NG word, is calculated, and the alarm generation unit determines the level of the alarm according to the NG word risk.

  According to the above configuration, not only the symptom word but also the conversation time, whether the customer is a specific customer (Creamer, important customer, etc.), or a customer in a specific region (if there is a sensitive term depending on the region) Since the NG word risk (degree of risk of issuing an NG word) is calculated based on the information on whether or not, an appropriate alarm according to the NG word risk can be generated.

  According to a fourth aspect of the present invention, in the third aspect of the present invention, the conversation content monitoring server allocates resources for processing voice / text conversion of the voice recognition server according to the NG word risk level. It is characterized by being changed.

  According to the above configuration, when processing the text of a conversation during a call in real time, the configuration of the resources of the voice recognition server (number of servers, number of CPUs, etc.) can be flexibly according to the load situation at that time Can be changed.

According the invention described in claim 5, before Symbol conversation candidate signs words extracted from the text, signs word registration determination unit that generates a determination screen for determining the determination of whether to register the administrator terminal, as the sign word It is characterized by providing .

  According to said structure, a human check and judgment can be added to determination of the symptom word which a conversation text analysis server performs.

A method for managing terms issued by an operator to a customer in a call center, wherein the voice capture step records a conversation between the customer and the operator, and the voice data recorded in the voice capture step is read out to be text data. Converting to conversational text; storing the conversational text in a conversational text storage DB;
A conversation text analysis step of statistically processing and analyzing the conversation text in the conversation text storage DB, wherein the conversation text analysis step reads the conversation text from the conversation text storage DB, and uses morphological analysis as a term unit. A morpheme analysis step that is decomposed into two, a step of detecting whether or not a predetermined NG word is included in the conversation text decomposed into term units in the morpheme analysis step, and an utterance before the NG word Determining whether the term included is a sign word that is a sign that the operator emits the NG word based on an NG word sign DB in which a sign word associated with each NG word is registered. It is characterized by that.

The invention of claim 6 regards the system of claim 1 as a method invention, and has the same effects as the invention of claim 1.
The invention according to claim 7 is a system for managing terms that an operator issues to a customer in a call center, wherein the voice capture server that records a conversation between the customer and the operator, and the voice capture server includes: A voice recognition server that reads recorded voice data, converts it into conversation text that is text data, and stores it in a conversation text storage DB; and a conversation text analysis server that statistically processes and analyzes the conversation text in the conversation text storage DB; The conversation text analysis server reads the conversation text from the conversation text storage DB and decomposes it into term units by morphological analysis, and the conversation decomposed into term units by the morphological analysis unit Whether the text contains a predetermined NG word If the NG word is detected and the NG word is included, a symptom word candidate that is a sign that the operator has issued the NG word when the number of occurrences of the term issued before the NG word is larger than a predetermined value. And a sign word registration determination unit stored in the NG word sign DB.
According to the configuration of the seventh aspect, since the text of the past conversation contents in the call center is statistically analyzed and extracted as a symptom word and stored in the database, the operator detects in advance a sign that an NG word is generated. Data can be obtained.

  According to the present invention, it is possible to capture the data of an indication that an NG word is issued, and by using the data to detect the indication word during a call, the operator can inadvertently issue an NG word to a customer. It can be prevented in advance.

It is a figure which shows the basic concept (NG word symptom data accumulation | storage stage) of this invention. It is a figure which shows the basic concept (NG word symptom detection step) of this invention. It is a figure which shows the functional block of the call center term management system which concerns on embodiment of this invention. It is a figure which shows the processing flow which accumulate | stores the data of a symptom word. It is a figure which shows the example of the determination screen of a sign word. It is a figure which shows typically the data stored in NG word indication DB133 in a table format. It is the figure which showed the flow of the process which checks the content of conversation in real time. It is a figure which shows an example of the operator's screen at the time of the alarm occurrence of a symptom word detection, and a supervisor's screen. It is a figure which shows the image of the resource change of a speech recognition server.

  DESCRIPTION OF EMBODIMENTS Hereinafter, embodiments for carrying out the present invention (hereinafter referred to as embodiments) will be described in detail with reference to the accompanying drawings. Note that the same numbers or symbols are assigned to the same elements throughout the description of the embodiment. In addition, arrows in a functional block diagram to be described later indicate a data flow direction or a process flow direction.

  1 and 2 are diagrams showing the basic concept of the present invention. As a basic concept, the present invention uses a call center speech recognition system and uses a term (hereinafter referred to as a “symptom word”) that is a symptom (warning factor) connected to the NG word instead of the NG word itself, and NG Stores word occurrence data and alerts the operator or supervisor (supervisor) of the operator or operator when a sign word is issued or when an NG word occurrence is detected It is what I did. FIG. 1 shows a conceptual diagram of the system in the accumulation stage of NG word sign data, and FIG. 2 shows a conceptual diagram of the system in the sign detection stage of NG word.

  First, FIG. 1 will be described. When a customer calls a call center, in general, first, an automatic voice response system (IVR) is connected, and an operator is connected only to a requirement that cannot be handled by the automatic response voice. The target is after the call with the operator is established.

  As shown in the figure, when a call between a customer and an operator is established, first, the voice capture server records the voice of the conversation content in a database. Next, the voice recognition server reads the recorded voice data and converts the conversation contents into text. Up to this point, it is widely used in general call centers. In the present invention, the conversation text analysis server then analyzes the conversation content converted to text (referred to as “conversation text”), and accumulates sign data including sign words that are connected to a predetermined NG word. Process statistically.

  The “NG word” here means a term that is predetermined by the call center as not to be issued by the operator to the customer, and is different from a broadcast prohibited term. A term originated by a customer is not treated as an NG word, but is treated as an “indication word” connected to the NG word. For example, if an operator inadvertently issues an NG word triggered by a term originated by a customer, the term issued by the customer becomes a symptom word. The same applies to the case where the operator repeats the parrot after returning the term registered in the NG word.

  Here, the signs of the NG word include a specific keyword issued by the customer or operator, a keyword for a specific service, a conversation time between the customer and the operator, whether the customer is a specific person, and whether the customer is in a specific region. Analyzes are made from data such as whether or not a person (this is called an evaluation item for each element).

  Of the above-mentioned specific keywords and keywords for specific services, a keyword determined to be a sign or trigger for the operator to issue an NG word will be particularly referred to as a “sign word”. For indication words, for example, when the operator utters “Absolutely OK”, which is an NG word for product information, etc., the customer ’s remarks such as “Is it all right” or “Absolutely?” If there is a statistically large number, the “is it all right” or “is it absolutely” is registered as an indication word. In addition, “keywords for specific services” are generally terms that are difficult to consider as NG words or indication words, but may be NG words or indication words in specific services. Is taken into consideration. For example, in a card company's call center, such as a service for confirming a customer's card points, terms such as “less”, “insufficient”, “decreasing”, etc., can be an indication word specific to the service.

  As for the conversation time between the customer and the operator, if there is a higher tendency to issue an NG word when the conversation time is significantly longer than usual, the conversation time is associated with the inquiry content and used as indication data. sign up. In addition, regarding “whether or not it is a specific customer”, if it can be determined that it is a customer with a specific temperament such as a claimor or an important customer, it is necessary to capture the signs of the NG word more carefully. Is provided. In addition, regarding “whether or not it is a specific region”, if there is a term (for example, earthquake, disaster, nuclear power plant, etc.) that must be treated sensitively depending on the region, the term is used as the NG word in that region. It is provided because it is necessary to carefully capture the signs of emitting the NG word.

  The statistical data analyzed by the conversation text analysis server is stored in a database (referred to as “NG word sign DB”) for that purpose, and is used in the NG word sign detection stage of FIG. 2 described later. Note that the voice-to-text conversion of the speech recognition server, the statistical analysis of the conversation text analysis server, and the accumulation of data do not necessarily have to be performed online when the call center is operating. Batch processing may be performed at night. Of course, depending on the processing capability of the voice recognition server, it may be processed online even while the call center is operating.

  FIG. 2 is a diagram showing a basic concept of the NG word sign detection stage. As shown in the figure, the conversation content monitoring server checks the conversation content between the customer and the operator in real time and monitors whether or not an NG word sign has appeared. Of course, it may be checked at the same time whether the operator has not actually issued an NG word.

  When the conversation content monitoring server detects a sign of NG word in the conversation of the operator A, the voice recognition server is most suitable for preferentially processing the conversation of the operator A to increase the text conversion processing speed of the conversation of the operator A. The allocation of various resources (number of servers for voice / text conversion, number of CPUs, number of CPU built-in cores, memory capacity, etc.). Upon receiving this instruction, the voice recognition server immediately converts the conversation content of the operator A so far into text, and the conversation content in the future is converted into text in real time and transmitted to the conversation content monitoring server. The conversation content monitoring server transmits an alarm corresponding to the “NG word risk” to the supervisor and the operator A at the time of receiving or before receiving this data. The NG word risk level means the risk level of the signs connected to the NG word (the intensity level of the signs), and is based on the data stored in the NG word sign DB (each data of the evaluation items described in FIG. 1). Is calculated by the conversation content monitoring server. Details of this calculation method will be described later.

  As described above, according to the system of the present invention, the operator catches the sign that the NG word is issued as soon as possible, and generates and transmits an appropriate alarm according to the NG word risk to the supervisor and the operator (operator A). NG word transmission can be prevented in advance. In addition, the supervisor who received the alarm will send the conversation contents of the operator A and the customer as text as well as the conversation contents up to now, not only after the alarm occurs, but also listen to the recorded conversations so far. The operator A can be preferentially monitored without correction. Also, the operator A can be alerted to pay attention to the subsequent conversation.

  FIG. 3 is a functional block diagram of a call center term management system 100 (hereinafter simply referred to as the present system) which is a typical embodiment of the present invention. This system is a computer system of a company call center. Each server described in FIG. 1 includes a plurality of operator terminals 200 (from about 10 to several hundreds of scales), one or more supervisor terminals 210, One or more administrator terminals 220 are connected via a network. In general, a single supervisor manages the situation of about 10 operators. In a large-scale call center, there are a large number of sets of supervisors and a plurality of operators (operator groups).

  In FIG. 3, the part surrounded by the solid line 110 corresponds to the voice capture server described in FIGS. 1 and 2. Similarly, the part surrounded by the solid line 120 is the voice recognition server, and the part surrounded by the solid line 130 is the conversation text. The part enclosed by the analysis server and the solid line 140 corresponds to the conversation content monitoring server. Each server is composed of a single computer or a plurality of computers. Particularly, the speech recognition server has a plurality of server configurations so that the processing capacity can be flexibly changed according to the load. It is desirable to provide a hardware device having a plurality of CPU cores. Hereinafter, each functional block of this system will be described in order.

  When a call from a customer arrives at the call center, a private branch exchange 101 (PBX: Private Branch eXchange) passes control to the automatic voice response unit 102 (IVR), and first responds with an automatic voice response. The automatic voice response device 102 allows the customer to input a requirement number corresponding to the inquiry content (can be omitted in a call center that divides the telephone number for each inquiry content). A terminal that is not currently in a call is found, and an incoming call is connected to the operator terminal 200 via the terminal control unit 103. At this time, when information that can identify the customer such as a requirement number (service code) and a card number or a mobile phone number is obtained, the information is also transmitted to the operator terminal 200 at the same time. Note that the operator terminal 200 is not necessarily in a call center facility, but may be a home PC of a home operator.

  In the terminal control unit 103, when the customer who has received a call can be identified, the customer information is read from the customer information DB 104 and displayed on the operator terminal 200. In this way, the call between the customer and the operator is established. In the case of a call center that does not employ IVR, or when the information transmitted from the IVR is insufficient, when the operator starts a conversation with the customer, the customer identification information and inquiry requirements are heard and input to the system. It is desirable.

(Voice capture server)
When the conversation between the customer and the operator starts, the voice recording unit 111 of the voice capture server 110 records the voice data of the conversation contents, and the voice data together with the conversation date / time, operator ID, customer ID, and service code (inquiry contents). Stored in the conversation recording DB 112. This is widely used in call centers as basic data for improving the quality of operators or as a countermeasure when there is a trouble with a customer.

(Voice recognition server)
The recorded data stored in the conversation recording DB 112 is converted into text data by the voice / text conversion unit 122 of the voice recognition server 120, and the converted and output text is stored in the conversation text stored inside or outside the voice recognition server 120. The data is sequentially stored in the DB 123. Since speech recognition is a known technique, the description thereof is omitted here. However, the voice recognition server 120 of this system, when it is necessary to process voice / text conversion during a call in real time, allows the resource allocation to be flexibly changed according to an instruction from the conversation content monitoring server 140. A resource management unit 121 that changes the allocation of the resource. This will be described in detail later.

(Conversation text analysis server)
The conversation text analysis server 130 includes a morpheme analysis unit 131, reads the conversation text from the conversation text accumulation DB 123, and decomposes it into morphemes (the smallest unit having meaning in language). Then, based on the analyzed morpheme, it is combined with a term (word) unit that is a word or a phrase, and all the terms that appear in the conversation are passed to the symptom word determination unit 132. Note that various known techniques may be used for the morphological analysis and the phrase analysis, but description of these known techniques is omitted here.

  The symptom word determination unit 132 checks whether or not the term corresponds to the NG word for all terms in the conversation. However, it is assumed that the NG word itself is registered in advance in the NG word DB 134 as an NG word list. Registration of the NG word is performed from the administrator terminal 220 via the NG word registration unit 135.

  The sign word determination unit 132 has two functions and is not particularly illustrated, but is further divided into a sign word utterance determination unit and a sign word registration determination unit. The symptom word utterance determination unit determines in real time whether the term being uttered corresponds to the symptom word list. That is, if an NG word is included in an utterance, the symptom word utterance determination unit rechecks the utterance term before the NG word appears and stores it in the symptom word list stored in the NG word symptom DB 133. To determine whether it corresponds to the specified term.

  The sign word registration determination unit usually extracts a sign word candidate from the content of the conversation by batch processing and determines whether or not to register as a sign word. That is, the sign word registration determination unit stores the extracted one or more terms as sign word candidates for the time being, and the stored sign word candidates are preferably registered as sign words after having been confirmed by an administrator. . Normally, a plurality of indication words are registered in association with one NG word. Similarly, a plurality of NG words may be associated with one indication word. Such mutual association registration of the NG word and the sign word is realized by using a known technique such as a relational database.

  When no indication word is associated, a conversation text immediately before the NG word appears or in a predetermined range is transmitted to the administrator terminal 220 to ask the system administrator to make a decision. The sign word determination unit 132 statistically processes the appearance frequency of the term before the NG word appears, and displays the term having a high appearance frequency before the NG word as the “sign word candidate” on the administrator terminal 220. It has a function to do. The administrator terminal 220 can also determine whether or not to register a sign word candidate as a sign word. For example, if the NG word that appears is the above-mentioned “absolutely ok”, the term “absolute” or “ok” uttered by the customer before that is also displayed as a candidate for the sign word. The manager determines whether or not the symptom word candidate should be the symptom word of the NG word and inputs it to the system. By doing so, statistical accumulation of the symptom word and human check are possible at the same time.

  Furthermore, it is desirable that the sign word determination unit 132 has a function of learning the determination result of the administrator according to the accumulation of the sign word data. For example, when a new NG word is registered, a function of analogizing a sign word of a new NG word from a result determined as a similar NG word and a sign word is provided. Moreover, it is desirable that a sign word once registered is also provided with a function of appropriately updating it so as to receive feedback from an operator or a supervisor.

(Conversation monitoring server)
The NG word sign detection unit 141 has a function of checking whether or not a sign word appears in the conversation contents during the call between the operator and the customer while referring to the NG word sign DB 133. Further, the NG word detection unit 144 also checks whether or not the NG word itself appears. The NG word sign detection unit 141 and the NG word detection unit 144 may be integrated. The check word word or the NG word may be checked using a prohibited word phoneme model as shown in the prior patent document 1, or a voice / text in real time (for example, within 1 second) by a voice recognition resource changing function described later. If conversion is possible, the text after the conversion may be checked. When the inquiry content (service content) can be acquired, the use of the indication word list or the NG word list categorized for each service content can improve the processing efficiency.

  In addition to the detection of the sign word, the NG word sign detection unit 141 also checks the evaluation value of each element of the conversation time, the service content, the specific customer, and the specific area, and calculates the NG word risk. The NG word risk can be calculated by, for example, the following mathematical formula using a method of multiple regression analysis (one of multivariate analysis).

NG word risk = (b ₁ × N ₁ + b ₂ × N ₂ + b ₃ × N ₃ +...) + A ₁ × T + a ₂ × S + a ₃ × C + a ₄ × R
However,
N ₁ , N ₂ , N ₃ ,... Are the number of occurrences of the sign word that appeared during the conversation, b ₁ , b ₂ , b ₃ are weighting factors for each sign word,
T is an elapsed time value (a value related to the conversation time from the start of the conversation to the present time) described later,
S is a predetermined value determined for each service content,
C is a specific customer value described later,
R is a specific area value described later,
a _{1 to} a ₄ are weighting factors of the respective evaluation values.

  In addition, when using the number of occurrences of the sign word to detect the sign of the NG word, it is not necessary to use the counted number as it is, but to use the number of times considering the state of the conversation before the sign word is uttered. Good. For example, if the operator replied “Yes, it ’s okay” to the customer ’s “Are you okay?”, The symptom word “I ’m okay” is “1”, If the operator replied “Yes, it ’s okay” to the customer ’s “Looks like it ’ll take a long time to make an inquiry,” the symptom word “I ’m okay” The number of times may be “0”.

  Further, when calculating the NG word risk calculated by a predetermined mathematical formula, the skill level of the operator may be taken into consideration. If the operator is highly skilled, the NG word is not issued while issuing the sign word. However, if the operator is low in skill, there is a high risk of issuing the NG word after issuing the sign word. Because.

  The voice recognition resource change instruction unit 142 is activated when the NG word sign detection unit 141 detects an NG word occurrence sign, and raises the priority of the voice recognition processing of the operator to the resource management unit 121 of the voice recognition server 120 in real time. Instruct to perform voice / text conversion. Upon receiving this instruction, the resource management unit 121 has a voice recognition resource changing function, and provides the voice / text conversion unit 122 with allocation of resources (server, CPU, CPU core, memory capacity, etc.) as much as possible. To improve processing speed. The resource provided at this time is determined by the NG word risk. Needless to say, when the highest level of NG word risk is detected, the resources allocated to the speech / text converter 122 are also maximized. In addition, when the operator who has been alerted by the alarm generation unit 143 to be described later no longer detects the sign word from the content of the subsequent conversation, the priority of the voice recognition of the operator whose priority has been increased is lowered. You may do it. This is because system resources can be effectively used in this way.

  The alarm generation unit 143 transmits a warning or alarm message to the operator terminal 200 and the supervisor terminal 210 in accordance with the detection result of the NG word sign detection unit 141 or the NG word detection unit 144. The content of the message may be defined for each alarm level by defining an alarm level corresponding to the NG word risk described above. A specific example of this message will be described later with reference to the drawings.

  The functional configuration of the system described above is merely an example, and one functional block (database and function processing unit) may be further divided, or a plurality of functional blocks may be configured as one functional block. . Each function processing unit reads a computer program stored in a storage device such as a ROM (Read Only Memory) or a hard disk by a CPU (Central Processing Unit) built in the computer device, and is executed by the CPU. It is realized by. That is, in each function processing unit, this computer program reads and writes necessary data such as a table from a database (DB; Data Base) stored in a storage device or a storage area on a memory, and may be related in some cases. This is realized by controlling hardware (for example, an input / output device, a display device, and a communication interface device). Further, the database (DB) in the embodiment of the present invention may be a commercial database, but it simply means a collection of tables and files, and the internal structure of the database itself is not questioned. The database may be considered as one server (database server).

(Sign word data storage processing flow)
FIG. 4 is a diagram showing a processing flow for accumulating data of a sign word. This process is a process performed by the conversation text analysis server 130 (hereinafter abbreviated as an analysis server). Note that the following processing is not necessarily performed in the order shown in the flowchart, and the processing order may be changed as long as the relationship between the input data and the output data of each processing block is not impaired.

  First, in step S10, the analysis server reads the NG word list from the NG word DB 134, and in step S11, reads the unprocessed conversation text from the conversation text storage DB 123. In step S12, it is checked whether the read conversation text includes an NG word. If the NG word is not included, the processed flag is attached to the conversation text, and then the process jumps to step S26 and sequentially checks until there is no unprocessed conversation text.

  If the conversation text contains an NG word, in step S13, the service content or inquiry content and conversation time of the conversation are acquired. It is assumed that a service code is pre-defined in the service content or inquiry content, and the service code is acquired by the IVR or operator at the start of the conversation with the customer and added to the header of the conversation text. . However, since multiple service contents may be included in one conversation, a keyword that can identify the service contents is detected from the entire conversation, and if multiple service contents are included, the service code is added. To do.

  Next, in step S14, the analysis server checks whether the customer is a registered customer. If it is a registered customer, the customer can be identified from the mobile phone number, customer number, and information that the operator has heard since the start of the conversation. If the customer is not registered, the processing of the next steps S15 to S16 is skipped.

  In step S15, it is checked whether or not it is a “specific customer”. The specific customer refers to a customer who has special temperament such as a claimer or a customer registered as an important customer (VIP) and should pay more attention than a normal customer. If it is a specific customer, an NG word list for specific customers is added for checking (step S16).

  In step S17, the analysis server further checks whether the customer's area can be specified. If the incoming call from the customer is from a landline, it can be easily identified from the area code, but if the conversation text contains keywords related to the region (prefecture, city, etc.) May be estimated using natural language analysis technology or the like. If the area can be specified, it is determined whether the NG word is unique to the area (step S18). If the region is unique, the NG word is linked to the specific region. Specifically, the area name is written in the specific area column of the NG word table (step S19).

  When the process for the specific customer and the specific area is completed, the analysis server proceeds to the process of step S20. In step S20, the number of appearances of a term (referred to as a preceding term) that occurs before the detected NG word in the conversation text is counted and added to the statistical data. In step S21, terms that are previous terms and the number of occurrences of the NG word in the statistical data are larger than a predetermined value are listed as indication word candidates. If the previous term has already been registered as an indication word, information to that effect is also added.

  Sets whether or not to request the judgment of a human (system administrator or a person assigned as a judge) whether or not to immediately register the sign word candidates listed in this way as a sign word of the NG word It is desirable to be able to do so. If it is set that human judgment is necessary (step S22: Y), the process proceeds to step S23, and a judgment screen for that purpose is displayed on the administrator's terminal or the like to request judgment.

  FIG. 5 shows an example of this determination screen. This figure shows information displayed on the screen, and the layout in the screen is free. A portion indicated by reference numeral 301 in the drawing is a portion for displaying basic information of the conversation text being processed. A portion denoted by reference numeral 302 is an area for displaying the full text of the conversation and the detected term in time series. Here, among the detected terms, the sign word candidates are shown in bold, the terms determined to be the sign words by the determiner are shown in bold + underline, and the NG words are shown in encircled ruled lines. In fact, it goes without saying that the operability is improved by displaying characters in color or changing the size. The number in parentheses after the term indicates that the term appears multiple times in the conversation. The determination person makes a final determination as to whether or not to register as a sign word on such a screen, and the sign word specified by pressing the registration button 303 is registered in the system.

  Returning to FIG. 4, if it is determined in step S24 that a new sign word needs to be registered, the determined term is stored in the NG word sign DB 133 as a sign word associated with the NG word in step S25. To do.

  FIG. 6 schematically shows data stored in the NG word indication DB 133 in a table format. In this example, a call center of a credit card company is assumed. In this table, NG words are divided into categories, and the terms before occurrence of NG words are listed in order of the number of occurrences for all NG words belonging to the category, and an evaluation value as a symptom word is added. The higher the evaluation value, the higher the possibility of the sign word. However, it should be noted that the evaluation value does not necessarily increase because the number of occurrences of the term is high. The sign word may be registered as zero or one, but by providing an evaluation value in this way, depending on the service content, the evaluation value is determined depending on the scene, such as a sign word above a predetermined evaluation value. You can change the criteria. This table can be realized by a relational database that joins a plurality of tables.

(Conversation content check processing flow)
FIG. 7 is a diagram showing a flow of processing for checking the conversation contents in real time. When the operator starts a conversation (call) with the customer, the conversation content monitoring server 140 (hereinafter simply referred to as the monitoring server) first acquires the service content (step S30), and the NG word corresponding to the service content is obtained. An indication word list (step S31) is acquired. If the service contents cannot be acquired, an indication word list of NG words corresponding to all service contents is acquired. Furthermore, the monitoring server may acquire information on the proficiency level of the operator (step S32).

  Next, in step S33, it is checked whether or not the customer is a specific customer from the basic conversation information. If it is a specific customer, a specific customer value is set (step S34). The specific customer value is a degree of attention of the customer, and is a value for raising the degree of attention and performing an alarm as the specific customer value is higher (in the simplest case, it is 0 or 1). As described above, the specific customer value is used when calculating the NG word risk level.

  Next, in step S35, it is checked whether or not it is a specific area from the basic conversation information. If it is a specific area, a specific area value is set (step S36). The specific area value is the degree of caution required for the area, and the higher the specific area value is, the higher the degree of caution is required for alarming (simply 0 or 1 may be used). The specific area value is also used when calculating the NG word risk level.

  Next, in step S37, the conversation time from the start of the conversation to the present is obtained, and it is checked whether or not the predetermined time determined for each service content has elapsed. If the predetermined time has elapsed, an elapsed time value is set (step S38). The elapsed time value is a value that indicates whether the conversation does not significantly exceed the normal conversation time of the service content, and the higher the elapsed time value is, the higher the caution level is to give an alarm. (For example, 1 is determined when a predetermined time exceeds 20%, 2 is determined when 50% or more is exceeded, etc.). The elapsed time value is also used when calculating the NG word risk level.

  In step S39, it is checked whether the service content has changed during the conversation (including addition) or the operator has been changed. If there are any changes during the conversation, the processing from step S30 is repeated. In step S40, it is checked whether the customer or the customer's area is found during the conversation. If these pieces of information are known, the processing from step S33 is performed again. However, in either case, the already processed part is stored as it is.

  Next, in step S41, it is checked whether or not the sign word acquired in step S31 is included in the conversation content. In step S42, the NG word risk is calculated. Since the specific method for calculating the NG word risk has already been described, it is omitted here. However, the proficiency level of the operator acquired in step S32 may be taken into consideration.

  In step S43, it is checked whether the calculated NG word risk does not exceed a predetermined value (threshold value). If the threshold is exceeded, an alarm corresponding to the NG word risk is transmitted to the terminal of the corresponding operator and the supervisor terminal managing and instructing the operator. FIG. 8 shows a screen when an alarm is generated at this time.

  FIG. 8 is a diagram illustrating an example of an operator screen and a supervisor screen when an alarm for detecting a sign word is generated. As shown in the figure, the terminal screen 400 of the operator A displays the symptom word issued by the customer and himself and the number of times. In addition, an NG word (in this example, “absolutely ok”) associated with the indication word is also displayed.

  Returning to FIG. 7, after the alarm is generated or simultaneously with the monitoring server, in step S45, the voice of the conversation in which the symptom word is detected to the voice recognition server 120 according to the calculated NG word risk level. / Increase resources for text conversion, further increase the priority, and send an instruction to convert the conversation in real time. Upon receiving this instruction, the speech recognition server 120, as shown in FIG. 9, resources for speech recognition (number of servers executed online, number of CPUs, number of CPU cores, memory capacity) according to the NG word risk and priority. To increase the processing speed.

  FIG. 9 is a diagram showing an image of resource change of the voice recognition server at this time. For example, when the priority is low (normal state), only the voice recognition server 1 is operating online, and the voice recognition server 2 is in a standby state for batch processing or backup. At this time, it is assumed that the voice / text conversion of the conversation of the operator A is processed only by the CORE1 of the CPU1.

  When the NG word risk level increases and the priority level becomes medium, all the cores of the CPU 1 are assigned to the above processing, and if the processing speed is still insufficient, the CPU 2, CPU 3 and CPU 4 are also assigned to this processing. .

  When the NG word risk level further increases and the priority level becomes high, not only the voice recognition server 1 but also the backup voice recognition server 2 is assigned. The number of servers actually allocated, the number of CPUs or COREs is determined by other load conditions. When there are a plurality of terminals that detect sign words, resource allocation to be prioritized is determined according to the NG word risk.

  Returning to FIG. 7, finally, in step S46, it is checked whether or not the conversation has ended. If the conversation has ended, the monitoring server ends monitoring for this call. If the conversation continues, the process returns to step S37 and the subsequent processing is repeated. In this way, a real-time check process of the conversation content is performed.

  In the examples of FIGS. 5, 6, and 8 described above, the call center system of a card company such as a credit card has been described as a specific example. However, it is naturally applicable to call centers of other companies. .

  In addition to the signs of the NG word, this system captures signs that the customer expresses emotional words such as satisfaction, dissatisfaction and frustration, joy and anger. It can also be applied to terms that require special attention.

  Also, depending on the call center, an operator who has a conversation with the customer within a certain period may reply to the inquiry by email as a follow-up email. In that case, it is natural that the NG word should not be included in the response sentence, but even if the inquiry sentence from the customer includes the NG word or the symptom word, it should be inadvertently included in the response sentence. It is necessary to be careful not to have any. For this reason, it is possible to calculate the NG word risk level of the response sentence before sending the mail and check it to further increase the alert of the operator to the NG word.

  As mentioned above, although this invention was demonstrated using embodiment, it cannot be overemphasized that the technical scope of this invention is not limited to the range as described in the said embodiment. It will be apparent to those skilled in the art that various modifications or improvements can be added to the above-described embodiments. Further, it is apparent from the scope of the claims that the embodiments added with such changes or improvements can be included in the technical scope of the present invention.

100 Call center term management system (this system)
101 Private Branch Exchange 102 Automatic Voice Response Device 103 Terminal Control Unit 104 Customer Information DB
110 voice capture server 111 voice recording unit 112 conversation recording DB
120 speech recognition server 121 resource management unit 122 voice / text conversion unit 123 conversation text storage DB
130 Conversation Text Analysis Server 131 Morphological Analysis Unit 132 Sign Word Determination Unit 133 NG Word Sign DB
134 NG Word DB
135 NG word registration unit 200 Operator terminal 210 Supervisor terminal 220 Administrator terminal 300 Sign word determination screen 400 Screen of the operator who detected the sign word 410 Screen of the supervisor managing the operator

Claims (7)

A system for managing terms issued by operators to customers in a call center,
A voice capture server that records conversations between the customer and the operator;
A voice recognition server that reads voice data recorded by the voice capture server, converts the voice data into conversation text that is text data, and stores the voice data in a conversation text storage DB;
A conversation text analysis server that statistically analyzes and analyzes the conversation text in the conversation text storage DB;
With
The conversation text analysis server includes:
A morpheme analyzer that reads the conversation text from the conversation text storage DB and decomposes it into term units by morphological analysis;
It is detected whether or not a predetermined NG word is included in the conversation text decomposed into term units by the morphological analysis unit, and when the NG word is included, the utterance before the NG word A sign word utterance determination that determines whether or not the term included in the word is a sign word that is a sign that the operator emits the NG word, based on an NG word sign DB in which a sign word associated with each NG word is registered And
A call center term management system comprising: It further includes a conversation content monitoring server that acquires voice data during a call between the customer and the operator and checks the conversation content in real time.
The conversation content monitoring server includes:
Based on data stored in the NG word indication DB, an NG word indication detection unit for detecting whether or not the indication word is included in the conversation during the call;
An alarm generating unit for generating an alarm on the terminal of the operator and a terminal of a supervisor managing the operator when the indication word is included in the conversation during the call;
The call center term management system according to claim 1, further comprising: The conversation content monitoring server is configured to determine whether the NG word, an indication word associated with the service content of the call center, conversation time of the conversation during the call, and whether the customer during the call is designated as a specific customer. Based on the information and at least one of the information indicating whether the customer in a call is a customer in a specific area, an operator calculates an NG word risk that is a risk of issuing an NG word,
The call center term management system according to claim 2, wherein the alarm generation unit determines the level of the alarm according to the NG word risk level.   4. The call center term management system according to claim 3, wherein the conversation content monitoring server changes allocation of resources for processing voice / text conversion of the voice recognition server according to the NG word risk level. Claims a candidate before Symbol signs words extracted from the conversation text, characterized in that it comprises a sign word registration determination unit that generates a determination screen for determining determines whether to register as the sign word to the administrator terminal call center terminology management system according to any one of 1 to 4. A method for managing terms issued by an operator to a customer in a call center,
A voice capture step of recording a conversation between the customer and the operator;
Reading the voice data recorded in the voice capture step and converting it into conversational text that is text data;
Storing the conversation text in a conversation text storage DB;
A conversation text analysis step of statistically processing and analyzing the conversation text in the conversation text storage DB;
Have
The conversation text analysis step includes:
A morphological analysis step of reading the conversation text from the conversation text storage DB and decomposing the conversation text into terms by morphological analysis;
Detecting whether or not a predetermined NG word is included in the conversation text decomposed into term units in the morphological analysis step;
Based on the NG word indication DB in which the indication word associated with each NG word is registered as to whether or not the term included in the utterance before the NG word is an indication word that is an indication that the operator issues the NG word. And judging step,
Call center term management method, which comprises a. A system for managing terms issued by operators to customers in a call center,
A voice capture server that records conversations between the customer and the operator;
A voice recognition server that reads voice data recorded by the voice capture server, converts the voice data into conversation text that is text data, and stores the voice data in a conversation text storage DB;
A conversation text analysis server that statistically analyzes and analyzes the conversation text in the conversation text storage DB;
With
The conversation text analysis server includes:
A morpheme analyzer that reads the conversation text from the conversation text storage DB and decomposes it into term units by morphological analysis;
It is detected whether or not a predetermined NG word is included in the conversation text decomposed into term units by the morphological analysis unit, and when the NG word is included, it is issued before the NG word. A vocabulary word registration determination unit that stores a term in which the number of occurrences of the generated term is greater than a predetermined value as a symptom word candidate that is a symptom of the operator issuing the NG word ;
A call center term management system comprising:

Images