KR20220042103A - Method and Apparatus for Providing Hybrid Intelligent Customer Consultation - Google Patents

Abstract

Disclosed are an apparatus and method for providing a hybrid intelligent customer consultation for efficient order creation and response prediction. In an embodiment of the present invention, when a counselor conducts a consultation, a consultation task is partially automated based on the prediction of a customer's intent by analyzing conversation progress. Provided are an apparatus and method for providing a hybrid customer consultation, which can automatically create and modify an order and a response sentence based on the analysis of the conversation progress.

Description

Hybrid Intelligent Customer Consultation Apparatus and Method {Method and Apparatus for Providing Hybrid Intelligent Customer Consultation}

The present invention relates to a hybrid intelligent customer consultation apparatus and method for efficient order creation and response prediction.

The content described below merely provides background information related to the present invention and does not constitute the prior art.

In a general customer counseling system, when a customer accesses a counseling service by phone or messenger, the entire counseling process may be performed based on direct conversation with a counselor or information retrieval by the counselor. 10 is a block diagram of an existing general counseling system. Here, the Interactive Voice Response (IVR) system is a counseling system that depends on voice conversation, and the messaging system is a counseling system that depends on a messenger service. The customer service system may include information available to the counselor for information retrieval in the counseling process.

A general counseling system requires a high skill level of a counselor, so there is a difficulty in processing a vast amount of services with high quality. That is, since each counselor has different skill levels, inconsistent services may be provided or problems with incorrect counseling may occur.

On the other hand, as AI (Artificial Intelligence) technology develops recently, a consultation system incorporating dialogue understanding technology has been developed and used. 11 is a block diagram of an existing intelligent counseling system based on a dialog system. The intelligent counseling system includes a conversation understanding system based on AI technology that is linked with a general counseling system.

A typical dialogue understanding system includes a language understanding module, a task prediction and execution module, a response generation module, and a customer service assistance model. The language understanding module infers the intention of the customer, the task prediction and execution module retrieves information that matches the intention and recommends a service item, and the response generation module generates a response to the intention. The customer service support model may include deep learning based models capable of supporting the operation of each module.

Due to language understanding errors or response generation errors that often occur in AI conversation systems, intelligent counseling systems often fail in their consulting work. In addition, since the intelligent counseling system often fails to correct or recover an error situation, it may cause great inconvenience to customers.

Therefore, there is a need for a hybrid intelligent counseling system that can overcome the frequent error situations shown by the AI conversation system while alleviating the high proficiency requirements for counselors.

Non-Patent Document 1: Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. 2018. Improving language understanding with unsupervised learning. Technical report, OpenAI.

According to the present disclosure, when a counselor conducts a consultation, a part of the consultation task is automated based on the prediction of the customer's intent by analyzing the conversation progress. Its main purpose is to provide a hybrid customer consultation device and method that enables automatic creation and modification of order form and response sentences based on the analysis of the conversation progress.

According to an embodiment of the present invention, a language understanding unit for extracting keywords from dialogues with customers; a task prediction and execution unit that analyzes the context of the conversation based on the keyword and predicts and executes tasks necessary for the consultation with the customer; a response generator configured to generate at least one candidate response to the conversation by using a response prediction model based on the execution result of the task; and a service review generation unit that generates an order form that meets the customer's intent by using an order generation model based on the execution result of the task. Counseling equipment is provided.

According to another embodiment of the present invention, in a hybrid counseling method of a hybrid counseling device, the process of extracting keywords from a conversation with a customer; analyzing the context of the conversation based on the keyword, predicting and executing a task required to proceed with the consultation with the customer; A process of generating at least one candidate response to the conversation by using a pre-trained deep-learning-based response prediction model based on the execution result of the task ; And based on the execution result of the task, using a deep learning-based order generation model trained in advance to include the process of generating an order book that meets the customer's intent (intent) characterized in that it comprises to provide a hybrid counseling method implemented on a computer.

According to another embodiment of the present invention, there is provided a computer program stored in a computer readable recording medium in order to execute each step included in the hybrid counseling method.

As described above, according to the present embodiment, it is possible to increase the efficiency of consulting work by providing a hybrid intelligent customer counseling device and method capable of automatically creating and modifying order form and response sentences based on the analysis of the conversation progress. It has a destructive effect.

In addition, according to this embodiment, by providing a hybrid intelligent customer counseling device and method capable of automatically generating and modifying order form and response sentences based on analysis of the conversation progress, customer satisfaction according to prompt and accurate processing of extensive counseling tasks It has the effect of making it possible to increase.

1 is a configuration diagram of a hybrid intelligent customer consultation system according to an embodiment of the present invention.
2 is an exemplary diagram of an order form review screen according to an embodiment of the present invention.
3 is a block diagram of an order book creation model according to an embodiment of the present invention.
4 is an exemplary diagram of a response modification process according to an embodiment of the present invention.
5 is an exemplary diagram of an AR language modeling method according to an embodiment of the present invention.
6 is a block diagram of a response prediction model according to an embodiment of the present invention.
7 is an exemplary diagram of DB search-based response modification according to another embodiment of the present invention.
8 is a flowchart for generating a response in a counseling support method according to an embodiment of the present invention.
9 is a flowchart for generating an order form in a counseling support method according to an embodiment of the present invention.
10 is a block diagram of an existing general counseling system.
11 is a block diagram of an existing intelligent counseling system based on a dialog system.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, embodiments of the present invention will be described in detail with reference to exemplary drawings. In adding reference numerals to the components of each drawing, it should be noted that the same components are given the same reference numerals as much as possible even though they are indicated on different drawings. In addition, in the description of the present embodiments, if it is determined that a detailed description of a related well-known configuration or function may obscure the gist of the present embodiments, the detailed description thereof will be omitted.

Also, in describing the components of the present embodiments, terms such as first, second, A, B, (a), (b), etc. may be used. These terms are only for distinguishing the elements from other elements, and the essence, order, or order of the elements are not limited by the terms. Throughout the specification, when a part 'includes' or 'includes' a certain element, this means that other elements may be further included, rather than excluding other elements, unless otherwise stated. . In addition, the '... Terms such as 'unit' and 'module' mean a unit that processes at least one function or operation, which may be implemented as hardware or software or a combination of hardware and software.

DETAILED DESCRIPTION The detailed description set forth below in conjunction with the appended drawings is intended to describe exemplary embodiments of the present invention and is not intended to represent the only embodiments in which the present invention may be practiced.

The present embodiment discloses the contents of a hybrid customer consultation apparatus and method for efficient order form creation and response prediction. More specifically, when a counselor conducts a consultation, a part of the consultation task is automated based on the prediction of the customer's intent by analyzing the conversation progress. Based on the analysis of the conversation progress, we provide a hybrid customer consultation device and method that enables automatic creation and modification of order form and response sentences.

1 is a configuration diagram of a hybrid customer consultation system according to an embodiment of the present invention.

In an embodiment of the present invention, the hybrid customer consultation system 100 analyzes the progress of a conversation between the counselor and the customer to predict the customer's intention, and prepares an order form and predicts a response. The hybrid customer consultation system 100 (hereinafter referred to as a hybrid consultation system) includes all or a part of the access system 110 , the customer service system 120 , and the consultation support device 130 .

The access system 110 links between the customer and the customer service system 120 . As the access system, an IVR (Interactive Voice Response) system or a messaging system may be used. The IVR system is based on a voice conversation using a telephone, and the messaging system is based on a text conversation using a messenger service.

There is no significant difference between the two access systems in terms of communication with customers. However, in terms of manpower management, the characteristic of the messaging system is that, unlike the IVR system, it is possible for one counselor to deal with multiple customers at the same time.

On the other hand, when the IVR system is used, if a module for recognizing voice conversations and converting them into text or vice versa is added in the hybrid consultation system 100, it will be possible to operate the process after text conversion in the same way as the messaging system. can

Therefore, in the description of the present embodiment, it is assumed that the conversation means a text conversation.

The customer service system 120 holds information about the customer and information on the product or service desired by the customer, and may include a DB (Database) inside or be linked with an external DB for data retrieval targeted for customer service. can

While receiving support from the counseling support device 130, By using the customer service system 120 , the counselor may perform services for the customer, such as responding and filling out an order form. The counselor may use the screen on the device on which the customer service system 120 is mounted to be connected to the access system 110 and the counseling support apparatus 130 , and use information inside the customer service system 120 .

The counseling support apparatus 130 according to the present embodiment predicts the customer's intent based on the analysis of the text conversation process, and prepares an order form and predicts a response. The counseling support device 130 includes a language understanding unit 131 , a task prediction and execution unit 132 , a response generation unit 133 , a service review generation unit 134 , and a service assistance model pool 135 . All or part of the model pool). Here, components included in the counseling support apparatus 130 according to the present embodiment are not necessarily limited thereto. For example, a training unit (not shown) for training an inference model may be additionally provided on the counseling support device 130 , or may be implemented in the form of interworking with an external training unit (meaning a server).

The language understanding unit 131 extracts keywords from the conversation with the customer. Here, the keyword may include a proper noun (eg, a customer name or a product name) necessary for customer consultation, a word containing the intention of the customer, and the like.

The language understanding unit 131 may extract a keyword by inputting the conversation into a language understanding model. The language understanding model is implemented as a neural network based on deep-learning, but may be trained in advance to be suitable for natural language processing.

The task prediction and execution unit 132 analyzes the context of a conversation based on a keyword, and predicts and executes a task required for consulting with a customer. The task includes searching for a consultation system, recommending a product or service item, and the like, and the execution result of the task may be used as basic information for predicting a response or writing an order form.

The task prediction and execution unit 132 stores information about a conversation between a counselor and a customer and a task execution history of the consultation system, and inputs the stored information into a task prediction model to provide a task required for consultation. predictable. The task prediction model is implemented as a deep learning-based neural network, and may be trained in advance to be suitable for task prediction.

The response generator 133 generates response information to be delivered to the customer based on the execution result of the task. Correspondence information may include a response to a conversation in progress, a table or figure in an organized form, and the like. Since the table or figure organized here may be included as a part of the response, it may be assumed that the generation of correspondence information generates a response to the conversation. Hereinafter, it is assumed that the response generator 133 generates a response.

The response generator 133 uses a deep learning-based response prediction model based on the execution result for the task. Multiple candidate responses can be generated. The generated plurality of candidate responses may be reviewed by the counselor. The response prediction model, the candidate response generation process, and the review stage by the counselor will be described later.

According to the counselor's preset setting, the counseling support device 130 may directly exchange a conversation with the customer using the response predicted by the response generating unit 133 . However, in the case of a direct conversation between the counseling support device 130 and the customer, the counselor should check the conversation history to see if there is any problem. The verification process consumes a lot of time and causes inconvenience of checking details in detail. Therefore, rather than a direct conversation, it may be more efficient to provide the customer with the counselor after reviewing and correcting the candidate response predicted by the response prediction model as described above.

The service review generation unit 134 generates an order book that meets the customer's intention based on the execution result of the task. The service review generation unit 134 may generate an order form using a deep learning-based order generation model based on the execution result for the task.

When the automated consultation support device 130 directly executes an important task such as an order, there is a risk due to a malfunction (especially in the case of an expensive product or service). Accordingly, the order book predicted by the counseling support device 130 may be passively reviewed by the counselor. The order form creation model and order form creation process will be described later.

The service support model pool 135 stores information on an inference model and an inference model required to perform a consultation on a task domain that the consultation support device 130 is responsible for. The inference model is a language comprehension model, a task prediction model, an order writing model, and/or a response prediction model.

The configuration of the counseling support device 130 shown in FIG. 1 is an exemplary configuration according to this embodiment, and the structure of the language understanding model, task prediction model, order form writing model, and response prediction model, and the counselor and counseling support device 130 ), an implementation including other components or other connections between components is possible according to the division of roles between them.

Hereinafter, an order creation model for creating order details and an order form that can be derived from a conversation in progress will be described.

2 is an exemplary diagram of an order form review screen according to an embodiment of the present invention. An example of FIG. 2A is a chatting window displaying an ongoing conversation, and an example of FIG. 2B is an order book displayed on the screen of the customer service system 120 .

The counseling support device 130 displays a highlight for a part that needs to be reviewed in the order form based on the keyword inferred from the conversation (in the case of (a) of FIG. 2 , italicized word) (FIG. In case of 2(b), thick solid line box). On the screen, "GG phone" highlights the model name according to the analysis result, "device change" and "24-month installment payment" according to the conversation sentence interpretation, indicating that confirmation by the counselor is required. After confirming the order, the counselor may forward it to the relevant department or conduct additional consultation with the customer.

에 대하여, 신뢰도는 예측 타겟(prediction target)

에 대한 사후 확률(posterior probability)이며 수학식 1로 표현할 수 있다. The counseling support device 130 calculates the confidence score of the order book detail items corresponding to the keywords extracted from the language understanding unit 131. there is. keyword sequence

For , the reliability is a prediction target

It is a posterior probability for , and it can be expressed by Equation 1.

는 결합확률(joint probability)을 이용하여 산정할 수 있다. Here, the prediction target represents an individual order item in the order book. The prediction target may have a form of a finite class or an output sequence. As illustrated in (b) of FIG. 2 , the order book is composed of a plurality of detail items, and it is possible to assume independence between the order details. As shown in Equation 2, the reliability of the entire order history

can be calculated using joint probability.

The counseling support device 130 may calculate the reliability of the prediction target by using the order book writing model based on the keyword column.

3 is a block diagram of an order book creation model according to an embodiment of the present invention.

The order creation model can be implemented as a deep learning-based neural network. In this embodiment, an order creation model based on a transformer decoder is used, but a transformer decoder in which an encoder-related part is deleted is used (refer to Non-Patent Document 1). As shown in Fig. 3, the order creation model includes N (N is a natural number, usually 12 or more) decoder modules and detection modules. Hereinafter, the entire stacked decoder module is referred to as a representation layer.

는 입력을 워드 임베딩(word embedding) 형태로 변환시키는 임베딩 행렬이며,

는 단어의 위치 정보에 대한 포지션 임베딩(position embedding)이다.

은 표현층의 출력으로서 예측 타겟에 대한 신뢰도를 추정 검출 모듈로 전달된다. 검출 모듈은 피드포워드(feedforward) 망 및 소프트맥스 층(softmax layer)을 포함할 수 있다.3 in the city

is an embedding matrix that transforms the input into word embeddings ,

is a position embedding for position information of a word.

is transmitted to the estimation detection module of the prediction target confidence as an output of the presentation layer. The detection module may include a feedforward network and a softmax layer.

The expression layer can be pre-trained by applying unsupervised learning to a large-capacity dialogue corpus that has no correct answer label. Wikipedia and Bookscorpus can be used as the large-capacity corpus, and hundreds of millions to billions of words extracted from the large-capacity corpus are used for dictionary training.

In the order book creation model including the pre-trained expression layer and the detection module, parameters can be fine-tuned based on supervised learning using training data and correct answer labels.

An order book creation model including a pre-trained expression layer and a detection module can efficiently detect a large number of order details. When the detection range is expanded or reduced, only the necessary parts of the order creation model can be precisely adjusted. In addition, since it is possible to estimate the posterior probability using the softmax layer without a separate module for reliability calculation, there is an effect that unnecessary calculation and modeling errors can be reduced.

Hereinafter, a response prediction model that generates a response to a conversation in progress and a process in which the counselor corrects the predicted response will be described.

There are cases in which the counselor needs to manually correct the response predicted by the counseling support device 130 . In this case, the counseling support apparatus 130 may predict a plurality of candidate responses in real time and display the candidate responses on the screen of the customer service system 120 so that the counselor conveniently executes the task of correcting the response.

4 is an exemplary diagram of a response modification process according to an embodiment of the present invention. A number of candidate sentences are proposed step by step as the sentences expressed in the first "courier delivery, today's arrival, immediate pickup, visiting opening ..." are slightly modified. As the sentence correction progresses, the type, rank, or number of candidate responses may change. The counselor can select from among the recommended phrases, or manually correct the recommendation when there is no suitable response to the recommended phrase.

In this embodiment, an AR (Auto Regressive) language model-based real-time response probability calculation method is used for real-time response recommendation. As expressed in Equation 3, the AR language model is a method of generating a sentence sequence with high probability based on a sequential word sequence.

A deep layered neural network, such as a deep neural network (DNN) or a recurrent neural network (RNN), is known as a structure suitable for generating time series and character sequences in an AR method.

이 출력되며 이 출력은 다시 신경망의 입력으로 사용되어 다음 단어 출력

가 얻어지는 재귀적(recurrent) 과정을 기반으로 응답이 생성된다. 여기서 토큰은 단어, 단어 취급되는 문구, 문장 구분을 위한 구분자(예컨대, <START>, <SEP> 등) 등을 의미한다.5 is an exemplary diagram of an AR language modeling method according to an embodiment of the present invention. The AR language modeling method as in the example of FIG. 5 may be applied to real-time response sentence generation. Print word for sentence start token <START>

is output, and this output is again used as an input for the neural network to output the next word

A response is generated based on a recurrent process in which . Here, the token means a word, a phrase treated as a word, a separator (eg, <START>, <SEP>, etc.) for separating sentences.

The AR language model using a simple recursive neural network as shown in the exemplary diagram of FIG. 5 may be difficult to apply in practice due to its low performance. Therefore, in this embodiment, a deep neural network based on a transformer decoder is used as a response prediction model.

6 is a block diagram of a response prediction model according to an embodiment of the present invention.

The response prediction model can be implemented as a deep learning-based neural network. In this embodiment, a response prediction model based on a transformer decoder is used (refer to Non-Patent Document 1). As shown in FIG. 6 , the response prediction model includes N (N is a natural number, usually 12 or more) decoder modules.

는 입력을 워드 임베딩 형태로 변환시키는 임베딩 행렬이며,

는 단어의 위치 정보에 대한 포지션 임베딩이다. 또한

은 최종 디코더 모듈의 출력으로서 워드 임베딩으로 변환된 후, 응답에 대한 확률을 추정할 수 있는 소프트맥스 층(softmax layer)으로 전달된다. 또한

이 디코더 모듈의 입력 쪽으로 전달됨으로써 응답예측 모델의 재귀 구조가 형성될 수 있다.In the city of Figure 6

is an embedding matrix that transforms the input into a word embedding form,

is a position embedding for position information of a word. also

is converted into word embedding as the output of the final decoder module, and then transferred to a softmax layer that can estimate a probability for a response. also

A recursive structure of the response prediction model can be formed by being transmitted to the input side of the decoder module.

The response prediction model can be pre-trained by applying an unsupervised learning method to a large-capacity corpus without an answer label. In the pre-trained response prediction model, parameters can be fine-tuned based on supervised learning using training data and correct answer labels.

에 대한 로그 라이클리후드(log likelihood)를 최대화시키는 손실 함수(loss function)를 기반으로 응답예측 모델에 대한 사전 트레이닝이 진행될 수 있다. 수학식 4는 사전 트레이닝 과정에서 이용되는 손실 함수를 나타낸다. 사전 트레이닝 과정에서의 입력 토큰열은 말뭉치로부터 추출한 임의의 문장일 수 있다.input token sequence

Pre-training of the response prediction model may be performed based on a loss function that maximizes the log likelihood of . Equation 4 represents a loss function used in the pre-training process. The input token sequence in the pre-training process may be an arbitrary sentence extracted from a corpus.

에 대하여, 정답 라벨 y를 기반으로 응답예측 모델의 출력에 대한 로그 라이클리후드(log likelihood)를 최대화시키는 손실 함수(loss function)를 기반으로 응답예측 모델에 대한 정밀 조정이 진행될 수 있다. 수학식 5는 정밀 조정 과정에서 이용되는 손실 함수를 나타낸다. 정밀 조정 과정에서의 입력 토큰열은 라벨이 부착된 학습용 문장일 수 있다.Also the input token sequence

, fine adjustment of the response prediction model may be performed based on a loss function that maximizes the log likelihood of the output of the response prediction model based on the correct answer label y. Equation 5 represents a loss function used in the fine tuning process. The input token string in the fine-tuning process may be a labeled learning sentence.

를 나타내고 있다. 여기서 k는 시간에 대한 인덱스(index)이다.By obtaining a token string that is the execution result of the task, the response prediction model infers a candidate response. Here, the token string may include a number of words that may constitute a response. Since the softmax function is used at the output stage of the response prediction model, m (m is a natural number) token outputs can be obtained for one token input. Equation 6 is a set containing m output tokens.

represents where k is an index with respect to time.

를 나타낸다. The response prediction model can generate a plurality of candidate responses at every time k to broaden the choice of counselors. Equation 7 is a set of candidate responses at time k

indicates

When the counselor modifies the response, the response prediction model may regenerate a plurality of candidate responses in real time by using the modified response as a start token.

In another embodiment of the present invention, instead of generating a candidate response, the response prediction model may use a database retrieval method for selecting a candidate response from a predefined response pool. Since one of the predefined responses is selected, the DB search method can generate a stable response. The search-based response has a characteristic that the candidate response is fixed.

When the counselor modifies a part of the response, only the response having the same part as the phrase modified by the counselor from among the responses retrieved from the response pool is first selected. Next, the selected response is applied to the response prediction model to obtain a probability for each response, and then re-rank between the selected candidate responses is performed.

7 is an exemplary diagram of DB search-based response modification according to another embodiment of the present invention. The example of FIG. 7 (a) is a list of responses retrieved from the response pool, and the example of FIG. 7 (b) is a list of candidate responses whose ranking has been re-arranged. As described above, since the counselor selected "Courier is now" for the response illustrated in (a) of FIG. 7, in the example of FIG. 7 (b), based on the probability inferred by the response prediction model, Candidate responses beginning with "now" are listed with reranking.

The device (not shown) on which the counseling support apparatus 130 according to the present embodiment is mounted may be a programmable computer and includes at least one communication interface that can be connected to a server (not shown).

Training for the deep learning-based neural network included in the counseling support device 130 according to the present embodiment may be performed in the server. The training unit (not shown) of the server can perform training for the deep learning model having the same structure as the language understanding model, task prediction model, order form writing model, and response prediction model of the consultation support device 130 mounted on the device. there is. Using a communication interface connected to the device, the server transmits the trained parameters to the device, and using the received parameters, the counseling support apparatus 130 provides parameters of the language understanding model, task prediction model, order form writing model, and response prediction model. can be updated. In addition, parameters of the language comprehension model, task prediction model, order form creation model, and response prediction model may be set at the time of device shipment or when the consultation support apparatus 130 is mounted on the device.

In another embodiment of the present invention, the counseling support apparatus 130 may be mounted on a server or a programmable system having a computing capability equivalent to that of the server. It is assumed that the server acquires a conversation from a plurality of devices (eg, an IVR system or a messenger system, etc.) capable of providing a conversation with a customer using a wired or wireless transmission method.

8 is a flowchart for generating a response in a counseling support method according to an embodiment of the present invention.

The consultation support apparatus 130 according to the present embodiment extracts keywords from the conversation with the customer (S801). Here, the keyword may include a proper noun (eg, a customer name or a product name) necessary for customer consultation, a word containing the customer's intent, and the like.

The counseling support apparatus 130 analyzes the context of the conversation based on the keyword, and predicts and executes a task required for the consultation with the customer (S802). The task includes searching for a consultation system, recommending a product or service item, and the like, and the execution result of the task may be used as base information for predicting a response or writing an order form.

The counseling support apparatus 130 generates at least one candidate response to the conversation by using a response prediction model based on the execution result of the task (S803).

The response prediction model can be implemented as a deep learning-based neural network. In this embodiment, a response prediction model based on a transformer decoder is used.

The response prediction model can be pre-trained by applying unsupervised learning to a large-capacity dialogue corpus that has no correct answer label. In the pre-trained response prediction model, parameters may be fine-tuned based on supervised learning using training data and correct answer labels.

The counseling support device 130 provides a candidate response to the counselor (S804).

The counselor checks whether the response to the conversation is completed using the candidate response (S805). Until the counselor completes the response, the counseling support device 130 generates at least one candidate response to the conversation using the response prediction model based on the counselor's selection (S806) and provides the process to the counselor (S804). Can be repeated.

Using the IVR system or the messenger system, the counselor provides a completed response to the customer (S807)

9 is a flowchart for generating an order form in a counseling support method according to an embodiment of the present invention.

In generating the order form, the steps of extracting keywords (S901), predicting and executing tasks (S902) by the counseling support apparatus 130 according to the present embodiment are part of the steps of generating responses (S801 and S802) step) is the same.

The consultation support device 130 generates an order book that meets the customer's intent by using an order generation model based on the execution result of the task (S903). The generated order form for review is provided to the consultant.

The order creation model can be implemented as a deep learning-based neural network. In this embodiment, an order book creation model based on a transformer decoder is used.

By applying the unsupervised learning method to a large-capacity corpus without an answer label, the order-writing model can be pre-trained. In the pre-trained order book creation model, parameters can be fine-tuned based on supervised learning using training data and correct answer labels.

After confirming the order, the counselor delivers it to the relevant department or proceeds with additional consultation with the customer (S904).

As described above, according to the present embodiment, it is possible to increase the efficiency of consulting work by providing a hybrid intelligent customer counseling device and method capable of automatically creating and modifying order form and response sentences based on the analysis of the conversation progress. It has a destructive effect.

In addition, according to this embodiment, by providing a hybrid intelligent customer counseling device and method capable of automatically generating and modifying order form and response sentences based on analysis of the conversation progress, customer satisfaction according to prompt and accurate processing of extensive counseling tasks It has the effect of making it possible to increase.

Although it is described that each process is sequentially executed in each flowchart according to the present embodiment, the present invention is not limited thereto. In other words, since it may be applicable to change and execute the processes described in the flowchart or to execute one or more processes in parallel, the flowchart is not limited to a time-series order.

Various implementations of the systems and techniques described herein may be implemented in digital electronic circuitry, integrated circuitry, field programmable gate array (FPGA), application specific integrated circuit (ASIC), computer hardware, firmware, software, and/or combination can be realized. These various implementations may include being implemented in one or more computer programs executable on a programmable system. The programmable system includes at least one programmable processor (which may be a special purpose processor) coupled to receive data and instructions from, and transmit data and instructions to, a storage system, at least one input device, and at least one output device. or may be a general-purpose processor). Computer programs (also known as programs, software, software applications or code) contain instructions for a programmable processor and are stored on a "computer-readable recording medium".

The computer-readable recording medium includes all types of recording devices in which data readable by a computer system is stored. These computer-readable recording media are non-volatile or non-transitory, such as ROM, CD-ROM, magnetic tape, floppy disk, memory card, hard disk, magneto-optical disk, and storage device. media, and may further include transitory media such as carrier waves (eg, transmission over the Internet) and data transmission media. In addition, the computer-readable recording medium is distributed in network-connected computer systems, and computer-readable codes may be stored and executed in a distributed manner.

Various implementations of the systems and techniques described herein may be implemented by a programmable computer. Here, the computer includes a programmable processor, a data storage system (including volatile memory, non-volatile memory, or other types of storage systems or combinations thereof), and at least one communication interface. For example, a programmable computer may be one of a server, a network appliance, a set-top box, an embedded device, a computer expansion module, a personal computer, a laptop, a Personal Data Assistant (PDA), a cloud computing system, or a mobile device.

The above description is merely illustrative of the technical idea of this embodiment, and a person skilled in the art to which this embodiment belongs may make various modifications and variations without departing from the essential characteristics of the present embodiment. Accordingly, the present embodiments are intended to explain rather than limit the technical spirit of the present embodiment, and the scope of the technical spirit of the present embodiment is not limited by these embodiments. The protection scope of this embodiment should be interpreted by the following claims, and all technical ideas within the scope equivalent thereto should be interpreted as being included in the scope of the present embodiment.

100: Hybrid customer service system
110: access system 120: customer service system
130: counseling support device 131: language understanding unit
132: task prediction and execution unit 133: response generation unit
134: service review generation unit 135: service support model pool

Claims (12)

a language understanding unit that extracts keywords from dialogues with customers;
a task prediction and execution unit that analyzes the context of the conversation based on the keyword and predicts and executes tasks necessary for the consultation with the customer;
a response generator configured to generate at least one candidate response to the conversation by using a response prediction model based on the execution result of the task; and
Based on the execution result of the task, a service review generation unit that generates an order form that meets the customer's intent by using an order generation model
including,
The service review generation unit,
Calculate a confidence score for detail items of the order book using an order book creation model based on the sequence of the keyword, and highlight detail information on the order book based on the reliability ) A hybrid (hybrid) counseling device, characterized in that. According to claim 1,
Further comprising a service assistance model pool (service assistance model pool), wherein the service assistance model pool stores an inference model necessary for consulting with the customer and information on the inference model, the inference model is A hybrid counseling apparatus, characterized in that it is a language understanding model for extracting the keyword, a task prediction model for predicting the task, the order form writing model, and/or the response prediction model. According to claim 1,
The keyword is
Hybrid counseling apparatus, characterized in that it includes a proper noun and/or a word implying the intention of the customer required to proceed with the consultation with the customer. According to claim 1,
The task is
A hybrid counseling apparatus comprising searching for information on the customer, searching for a product or service meeting the intention of the customer, and/or recommending the product or service. According to claim 1,
The response prediction model and the order writing model are implemented as a deep learning-based neural network, and pre-training based on an unsupervised learning method using a large-capacity dialogue corpus without an answer label. After training), the hybrid counseling device, characterized in that the parameters are fine-tuned based on supervised learning using the training data and the correct answer label. According to claim 1,
The response generating unit,
At least one candidate response corresponding to the token sequence by generating a softmax output by executing a recurrent operation of the response prediction model based on a token sequence that is the execution result of the task Hybrid counseling device, characterized in that for generating. According to claim 1,
The response generator provides the at least one candidate response to the counselor, and the service review generator provides the order book to the counselor. 8. The method of claim 7,
The response generating unit,
Based on the counselor's selection or modification of the at least one candidate response, iteratively generates the at least one candidate response using the response prediction model until a response to the conversation with the customer is completed A hybrid counseling device with In the hybrid counseling method of a hybrid counseling device,
the process of extracting keywords from conversations with customers;
analyzing the context of the conversation based on the keyword, predicting and executing a task required to proceed with the consultation with the customer;
A process of generating at least one candidate response to the conversation using a pre-trained deep-learning-based response prediction model based on the execution result of the task ; and
A process of generating an order book that meets the customer's intent using a pre-trained deep learning-based order generation model based on the execution result of the task
including,
The process of creating the order form is:
Calculate a confidence score for detail items of the order book using an order book creation model based on the sequence of the keyword, and highlight detail information on the order book based on the reliability ), characterized in that, hybrid counseling method. 10. The method of claim 9,
Hybrid counseling method, characterized in that it further comprises the step of providing the at least one candidate response and the order book to the counselor. 11. The method of claim 10,
The process of generating the candidate response includes:
Based on the counselor's selection or modification of the at least one candidate response, iteratively generates the at least one candidate response using the response prediction model until a response to the conversation with the customer is completed , a hybrid counseling method. 12. A computer program stored in a recording medium, which is readable by a computer to execute each step included in the hybrid counseling method according to any one of claims 9 to 11.

Images