KR102372642B1 - Apparatus and method for generating an automatic response - Google Patents

Abstract

Disclosed are a method for generating an automatic response and an apparatus for performing the same. According to one embodiment of the present invention, the apparatus for generating the automatic response comprises: a training data generating unit generating training data based on past conversation data including a first response request message and a first response message set corresponding to the first response request message, personality data, interruption data, and a second response message for a second response request message; and a model training unit training an automatic responding model so that the second response message generated by the automatic responding model becomes closer in a similarity with the personality data and becomes farther in a similarity with the interruption data.

Description

AUTOMATIC RESPONSE {APPARATUS AND METHOD FOR GENERATING AN AUTOMATIC RESPONSE}

Disclosed embodiments relate to techniques for generating an automatic response to a user's request for an answer.

Messenger is a service that has been in use since the mid-1990s. After the soft landing of services centered on PCs (Personal Computers), smartphones became a representative means of conversing with people with the advent of smartphones. Mobile messenger companies used 'Chatbot' technology that combines artificial intelligence and messenger to create a new platform for messenger influence.

However, the chatbot technology had a problem in that it did not generate a consistent response. Accordingly, a demand for a chatbot that generates a response with a consistent character has emerged.

Republic of Korea Patent Publication No. 10-2020-0042137 (published on April 23, 2020)

The disclosed embodiments are intended to provide a technical means for constructing learning data necessary for generating an automatic response to a user's response request.

According to an embodiment, the automatic response generating apparatus includes a first response request message and past conversation data including a first response message set corresponding to the first response request message, personality data, interruption data, and a second response request message. The second response message generated by the automatic response model on the basis of the training data generator and the training data for generating training data based on the second response message to the second response message to the similarity with the personality data is close , and a model learning unit for learning the automatic response model so that the degree of similarity with the interference data is farther away.

The personality data may include a plurality of sentences pre-classified as being uttered by a person having a specific personality.

The interference data includes a disturbing sentence generated based on a degree of similarity with words included in the plurality of sentences included in the personality data, and a label of whether the interference is authentic or not determined based on the similarity to the personality data. may include

The past conversation data may include the first response message corresponding to the first response request message generated based on a similarity with a plurality of sentences constituting the personality data.

The automatic response model includes a generator that receives the learning data and generates the second response message based on the personality data;

and a discriminator that receives the learning data and determines whether the interference of the second response message is authentic or not based on the similarity between the interference data and the second response message.

The model learning unit, a first loss function based on the similarity between the personality data and the second response message, and the disturbance authenticity label of the disturbance data are true and the second response message is false. Learning the discriminator using a second loss function to determine, based on the output value of the third loss function including the first loss function and the second loss function according to the similarity with the personality data and the disturbance authenticity label so you can learn

The response generating method according to an embodiment includes a first response request message and a first response message set corresponding to the first response request message for past conversation data, personality data, interruption data, and a second response request message. Generating training data based on a second response message and based on the training data, the second response message generated by the model has a similarity with the personality data close to, and with the interference data Similarity involves training the autoresponder model to move away from it.

The personality data may include a plurality of sentences pre-classified as being uttered by a person having a specific personality.

The interference data includes a disturbing sentence generated based on a degree of similarity with words included in the plurality of sentences included in the personality data, and a label of whether the interference is authentic or not determined based on the similarity to the personality data. may include

The past conversation data may include the first response message corresponding to the first response request message generated based on a similarity with a plurality of sentences constituting the personality data.

The automatic response model receives the learning data and generates the second response message based on the personality data. Based on the similarity between the interruption data and the second response message by receiving the learning data and receiving the learning data. Thus, it may include a discriminator for determining whether the interference of the second response message is authentic.

The model learning unit, the first loss function based on the similarity between the personality data and the second response message, and the disturbance authenticity label of the disturbance data are true, and the second response message is false. The automatic response model may be trained by using a third loss function based on the second loss function for determining.

According to the disclosed embodiments, by inputting personality data information, it is possible to maintain consistency of a response message based on the personality data.

According to the disclosed embodiments, it is possible to maintain the consistency of the response message by suppressing the generation of the interference data inconsistent with the personality data.

1 is a block diagram illustrating an apparatus for generating an automatic response according to an embodiment;
2 is an exemplary diagram for explaining a method of configuring learning data according to an embodiment;
3 is a block diagram illustrating an automatic response model for generating an automatic response according to an embodiment;
4 is a flowchart illustrating a method for generating an automatic response according to an embodiment;
5 is a block diagram illustrating an automatic response model for generating an automatic response according to an embodiment;
6 is a block diagram illustrating and explaining a computing environment including a computing device according to an embodiment;

Hereinafter, specific embodiments will be described with reference to the drawings. The following detailed description is provided to provide a comprehensive understanding of the methods, apparatus, and/or systems described herein. However, this is merely an example, and the disclosed embodiments are not limited thereto.

In the description of the embodiments, if it is determined that a detailed description of a related known technology may unnecessarily obscure the gist of the disclosed embodiments, the detailed description thereof will be omitted. And, the terms to be described later are terms defined in consideration of functions in the disclosed embodiments, which may vary according to intentions or customs of users and operators. Therefore, the definition should be made based on the content throughout this specification. The terminology used in the detailed description is for the purpose of describing the embodiments only, and should in no way be limiting. Unless explicitly used otherwise, expressions in the singular include the meaning of the plural. In this description, expressions such as “comprising” or “comprising” are intended to indicate certain features, numbers, steps, acts, elements, some or a combination thereof, one or more other than those described. It should not be construed to exclude the presence or possibility of other features, numbers, steps, acts, elements, or any part or combination thereof.

1 is a block diagram illustrating an answering machine according to an exemplary embodiment.

The automatic response apparatus 100 according to an embodiment receives an automatic response request to the response request message, and generates and provides a response message having a similarity to personality data.

The automatic answering device 100 is a computing system configured to receive an inquiry and return a response to the inquiry according to a preset response rule. In an embodiment, the automatic response apparatus 100 may include a response rule based on a rule or artificial intelligence, etc., and may generate a response to the query according to the response rule. there is.

As shown, the automatic response apparatus 100 according to an embodiment includes a training data generator 110 and a model learner 120 .

The learning data generation unit 110 includes a first response request message and a second response to past conversation data, personality data, interference data, and a second response request message including a first response message set corresponding to the first response request message. Generate training data based on the message.

According to an embodiment, the learning data generator 110 may acquire personality data that gives a consistent pattern to the response message of the automatic response device 100 in order to maintain consistency of responses corresponding to response requests. At this time, personality encompasses an individual's mood, attitude, and opinion, and is clearly revealed in interactions with others, and refers to a consistent behavioral pattern or internal mental process occurring inside an individual.

According to an embodiment, including a plurality of sentences pre-classified as being uttered by a person having a specific personality.

Specifically, personality data includes information related to social status such as occupation, family relationships, and wealth, as well as information that represents a unique adaptation to a situation encountered by an individual It may be composed of words expressing thoughts and emotions including a criterion for recognition, a criterion for classification of situations divided into thoughts/feelings, and a lifestyle divided into judgment/recognition.

According to an embodiment, the disturbance data includes a disturbance sentence generated based on a degree of similarity with words included in a plurality of sentences included in the personality data, and a label of whether the disturbance sentence is generated based on the similarity to the personality data includes In detail, the interruption data may include an interruption sentence composed of words having a far similarity to words included in a plurality of sentences constituting the personality data.

Specifically, when the disturbing sentence has a far similarity to the personality data, the interference authenticity label may be set to a true value.

According to an embodiment, the past conversation data includes a first response message corresponding to the first response request message generated based on a similarity with a plurality of sentences constituting the personality data.

According to an embodiment, the past conversation data is used to generate a response message requested by the automatic response device 100 and may include all conversations before the second response request message for which a response is requested.

According to an embodiment, the learning data generating unit 110 may generate the learning data by determining the similarity of words constituting the past conversation data, the interruption data, and the second answer request message based on the personality data.

Based on the training data, the model learning unit 120 trains the automatic response model so that the second response message generated by the automatic response model has a similarity to personality data and a similarity to interference data is farther away.

According to an embodiment, the automatic response model for automatic response generation may include a generative model and a discriminative model.

According to an embodiment, the model learning unit 120 receives training data and generates a second response message based on personality data, and based on the similarity between the interference data and the second response message by receiving the training data. and a discriminator for determining whether the second response message is disturbed or not.

According to an embodiment, the model learning unit 120 sets the first loss function based on the similarity between the personality data and the second response message, the interference authenticity label of the interference data to true, and sets the second response message to false. ) to learn the classifier using the second loss function to determine the learn

2 is an exemplary diagram for explaining a method of configuring learning data according to an embodiment.

As in the example shown in FIG. 2 , the sentence 200 for generating exemplary learning data includes a plurality of sentences 210 constituting personality data, a sentence 220 constituting past conversation data, and a second response message 230 . ) and a disturbance sentence 240 constituting the disturbance data.

According to an embodiment, the learning data 200 may be generated based on embedding vectors for each of the sentences 210 constituting the personality data, the sentences included in the past conversation data, and the words included in the disturbing sentences.

According to an embodiment, the learning data may be expressed by Equation (1).

은 학습 데이터, p 는 성격 데이터를 구성하는 복수의 문장,

은 과거 대화 데이터에 포함된 제1 답변 요청 메시지와 제1 응답 메시지를 포함하는 과거 문장,

은 방해 데이터에 포함된 방해 문장,

은 제2 응답 메시지를 나타낸다. At this time,

is the learning data, p is a plurality of sentences constituting the personality data,

is a past sentence including the first response request message and the first response message included in the past conversation data,

is the disturbance sentence included in the disturbance data;

indicates the second response message.

According to an embodiment, the past sentence may be expressed by Equation (2).

은 제1 답변 요청 메시지와 제1 응답 메시지를 포함하는 과거 문장,

은 i번째 제1 답변 요청 메시지,

은 i번째 제1 응답 메시지를 나타낸다. At this time,

is a past sentence including a first response request message and a first response message,

is the i-th first response request message,

denotes an i-th first response message.

)는 성격 데이터에 포함된 복수의 문장(p), 과거 대화 데이터에 포함된 문장(

), 방해 문장 (

) 및 제2 응답 메시지(

)로 구성될 수 있다.Specifically, according to the training data (

) is a plurality of sentences (p) included in personality data, sentences (p) included in past conversation data

), interfering sentences (

) and the second response message (

) may consist of

Referring back to FIG. 2 , the plurality of sentences 210 constituting the personality data are expressed by p in the equation, 'hobbies are watching TV dramas', 'I like to rest at home on weekends', and 'I like Christmas. .' It may include a plurality of sentences such as

), 'A: 1년 중에서 가장 좋아하는 연휴는 언제야?'와 첫 번째 제1 답변 요청 메시지(

)에 대응하여 자동 응답 장치(100)가 생성한 첫 번째 제1 응답 메시지(

), 'B: 나는 크리스마스를 좋아해'의 문장(

)으로 표현될 수 있다.Referring back to FIG. 2 , the sentence 220 constituting the past conversation data is the first first answer request message (

), 'A: When is your favorite holiday of the year?' and the first response request message (

) in response to the first first response message generated by the automatic response device 100 (

), the sentence of 'B: I like Christmas' (

) can be expressed as

)일 수 있다. 이때, 방해 데이터는 방해 문장(

)의 성격 데이터와의 유사도가 먼 방해 문장 여부를 판단하는 방해 진위 여부 레이블(TRUE)을 포함할 수 있다.Referring back to FIG. 2 , the interruption sentence 240 constituting the interruption data may be composed of words that are distant from the sentence 'I like Christmas', which is a sentence included in the personality data, and 'I have white eyes. The sentence 'I really like it' is a disturbing sentence (

) can be At this time, the disturbance data is a disturbance sentence (

) may include a disturbance authenticity label (TRUE) for determining whether a distant disturbance sentence has a similarity with the personality data.

According to an embodiment, the second response message may be generated based on co-occurrence probability of words constituting the second response message. In this case, the simultaneous appearance probability (P(k/i)) may mean a conditional probability calculated by counting the number of times that the k word appears when the i word appears.

According to an embodiment, the generation probability of the second response message generated by the answering machine 100 may be expressed by Equation (3).

은 자동 응답 장치(100)에 의해 학습 데이터에 포함된 단어에 기초하여 제2 응답 메시지 생성 확률,

은 학습 데이터,

은 학습 데이터를 구성하는 복수의 문장에 포함된 단어,

는 각각의 단어(

)가 선택될 확률을 나타낸다.

is the second response message generation probability based on the word included in the learning data by the automatic response device 100,

is the training data,

is a word included in a plurality of sentences constituting the learning data,

is each word (

) represents the probability of being selected.

)은 각각의 단어(

)가 선택될 확률(

)의 곱으로 계산될 수 있다. 구체적으로 제2 응답 메시지는 성격 데이터에 포함된 단어와의 유사도가 가까운 단어가 선택되는 경우 확률(

)은 커지고, 방해 문장과 가까운 단어가 선택되는 경우 확률(

)은 작아질 수 있다. At this time, the second response message generation probability (

) for each word (

) will be chosen (

) can be calculated by multiplying Specifically, the second response message is a probability (

) increases, and the probability (

) can be small.

)은 커질 수 있으며, 방해 데이터에 포함된 '눈'이 선택되는 경우 확률(

)은 작아 질 수 있다. For example, when the word 'holiday' included in the second response request message 'When is your favorite holiday in a year?'

) can be large, and the probability (

) can be small.

)을 수학식 4를 통해 구할 수 있다.According to an embodiment, the automatic response apparatus 100 may generate a probability (

) can be obtained through Equation 4.

은 전체 대화의 생성 확률,

은 학습 데이터, N은 전체를 문장의 수를 나타낸다. At this time,

is the generation probability of the entire conversation,

is the training data, and N is the total number of sentences.

)은 제2 응답 메시지 생성 확률(

)과 학습 데이터(

)의 조건부 확률 곱으로 계산할 수 있다. 즉, 성격 데이터에 포함된 단어와의 유사도가 가까운 단어가 선택되는 경우 전체 대화의 생성 확률(

)은 커지고, 방해 문장과 가까운 단어가 선택되는 경우 전체 대화의 생성 확률(

)은 작아질 수 있다. Specifically, the generation probability of the entire conversation (

) is the second response message generation probability (

) and training data (

) can be calculated as the conditional probability product of That is, when a word with a close similarity to the word included in the personality data is selected, the generation probability of the entire conversation (

) increases, and the probability of generation of the entire conversation (

) can be small.

3 is a block diagram illustrating an automatic answering model for generating an automatic answering apparatus according to an exemplary embodiment. As shown, the automatic response model according to an embodiment includes a generator 310 and a delimiter 320 .

The automatic response model may receive training data and generate a second response message. In this case, the second response message may generate a second response message such that the similarity to the personality data is close and the similarity to the interference data is farther away.

In this case, according to an embodiment, the automatic response model may be, for example, a Masked Self Attention-based model.

According to an embodiment, for example, the automatic response model may be learned based on Equation 5 below.

은 자동 응답 모델의 전체 손실 함수,

은 자동 응답 모델의 생성 손실 함수,

은 자동 응답 모델의 방해 손실,

는 방해 손실 함수 조절 파라미터를 나타낸다. At this time,

is the overall loss function of the autoresponse model,

is the generation loss function of the autoresponse model,

is the disturbance loss of the autoresponse model,

is the disturbance loss function adjustment parameter.

According to an embodiment, the second response message generated by the generator 310 reduces the loss of the generation loss function so that the similarity to a plurality of sentences included in the personality data is close, and the separator 320 prevents the generation of the disturbing sentence. It is possible to reduce the loss of the disturbance loss function. Thereafter, the automatic response model can be trained using the overall loss function using the generation loss function and the disturbance loss function.

Meanwhile, the generator 310 may calculate the generation loss based on Equation 6 below.

은 자동 응답 모델의 생성 손실 함수,

은 정답으로 입력된 제2 응답 메시지를 구성하는 단어의 성격 데이터에 기초한 유사도 확률을 나타내고,

은 예측된 제2 응답 메시지를 구성하는 단어의 성격 데이터에 기초한 유사도 확률, M은 단어 집합의 크기를 나타낸다. At this time,

is the generation loss function of the autoresponse model,

represents the similarity probability based on the personality data of the words constituting the second response message input as the correct answer,

is a similarity probability based on personality data of words constituting the predicted second response message, and M represents a size of a word set.

)를 통해 성격 데이터를 구성하는 복수의 문장과 유사도 차이에 의한 손실을 계산할 수 있다. Specifically, the generator 310 generates a generation loss function (

), it is possible to calculate the loss due to differences in similarity with multiple sentences constituting personality data.

Meanwhile, the delimiter 320 may calculate the disturbance loss based on Equation 7 below.

은 입력된 학습 데이터에 포함된 방해 진위 여부 레이블,

은 생성자(310)가 생성한 제2 응답 데이터의 성격 데이터에 기초한 유사도에 기초하여 방해 진위 여부의 확률을 나타낸다. At this time,

is the authenticity label of the disturbance contained in the input training data;

denotes the probability of whether the interference is authentic or not based on the similarity based on the personality data of the second response data generated by the generator 310 .

)이 참인지 거짓인지 판단하며, 구체적으로 방해 데이터의 방해 진위 여부 레이블(

)은 참(true), 정답으로 입력된 제2 응답 메시지의 방해 진위 여부를 거짓(false)으로 판단하도록 방해 손실 함수를 통해 방해 손실을 계산할 수 있다. The delimiter 320 is a label (

) is true or false, and specifically a label (

) may be true, and the disturbance loss may be calculated through the disturbance loss function to determine the authenticity of the disturbance of the second response message input as the correct answer as false.

)이 참인 경우, 방해 손실 함수(

)가 커지도록 하여 방해 문장을 생성하지 못하도록 할 수 있다. Specifically, the delimiter 320 is the interference authenticity label (

) is true, the disturbance loss function (

) can be increased to prevent the generation of interfering sentences.

4 is a flowchart illustrating an automatic response generating method according to an exemplary embodiment.

The method shown in FIG. 4 may be performed, for example, by the aforementioned automatic answering machine 100 .

In step 410, the automatic response device 100 for the first response request message and the first response request message for the past conversation data including the first response message set corresponding to the response request message, personality data, interruption data and the second response request message It generates training data based on the second response message.

In step 420, based on the learning data, the automatic response model is trained so that the second response message generated by the automatic response model has a similarity with the personality data and a similarity with the interference data is farther away.

In the illustrated flowchart, the method is described by dividing the method into a plurality of steps, but at least some of the steps are performed in a different order, are performed together in combination with other steps, are omitted, are performed in sub-steps, or are not shown. One or more steps may be added and performed.

5 is a block diagram illustrating and describing a computing environment 10 including a computing device according to an embodiment. In the illustrated embodiment, each component may have different functions and capabilities other than those described below, and may include additional components in addition to those described below.

The illustrated computing environment 10 includes a computing device 12 . In one embodiment, computing device 12 may be automatic response generating device 100 .

Computing device 12 includes at least one processor 14 , computer readable storage medium 16 , and communication bus 18 . The processor 14 may cause the computing device 12 to operate in accordance with the exemplary embodiments discussed above. For example, the processor 14 may execute one or more programs stored in the computer-readable storage medium 16 . The one or more programs may include one or more computer-executable instructions that, when executed by the processor 14, configure the computing device 12 to perform operations in accordance with the exemplary embodiment. can be

Computer-readable storage medium 16 is configured to store computer-executable instructions or program code, program data, and/or other suitable form of information. The program 20 stored in the computer readable storage medium 16 includes a set of instructions executable by the processor 14 . In one embodiment, computer-readable storage medium 16 includes memory (volatile memory, such as random access memory, non-volatile memory, or a suitable combination thereof), one or more magnetic disk storage devices, optical disk storage devices, flash It may be memory devices, other forms of storage medium accessed by computing device 12 and capable of storing desired information, or a suitable combination thereof.

Communication bus 18 interconnects various other components of computing device 12 , including processor 14 and computer readable storage medium 16 .

Computing device 12 may also include one or more input/output interfaces 22 and one or more network communication interfaces 26 that provide interfaces for one or more input/output devices 24 . The input/output interface 22 and the network communication interface 26 are coupled to the communication bus 18 . Input/output device 24 may be coupled to other components of computing device 12 via input/output interface 22 . Exemplary input/output device 24 may include a pointing device (such as a mouse or trackpad), a keyboard, a touch input device (such as a touchpad or touchscreen), a voice or sound input device, various types of sensor devices, and/or imaging devices. input devices, and/or output devices such as display devices, printers, speakers and/or network cards. The exemplary input/output device 24 may be included in the computing device 12 as a component constituting the computing device 12 , and may be connected to the computing device 12 as a separate device distinct from the computing device 12 . may be

Meanwhile, an embodiment of the present invention may include a program for performing the methods described in this specification on a computer, and a computer-readable recording medium including the program. The computer-readable recording medium may include program instructions, local data files, local data structures, etc. alone or in combination. The medium may be specially designed and configured for the present invention, or may be commonly used in the field of computer software. Examples of the computer-readable recording medium include hard disks, magnetic media such as floppy disks and magnetic tapes, optical recording media such as CD-ROMs and DVDs, and program instructions specially configured to store and execute program instructions such as ROMs, RAMs, flash memories, etc. Hardware devices are included. Examples of the program may include high-level language codes that can be executed by a computer using an interpreter or the like as well as machine language codes such as those generated by a compiler.

Although representative embodiments of the present invention have been described in detail above, those of ordinary skill in the art to which the present invention pertains will understand that various modifications are possible within the limits without departing from the scope of the present invention with respect to the above-described embodiments. . Therefore, the scope of the present invention should not be limited to the described embodiments, but should be defined by the following claims as well as the claims and equivalents.

10: Computing Environment
12: computing device
14: Processor
16: computer readable storage medium
18: communication bus
20: Program
22: input/output interface
24: input/output device
26: network communication interface
100: automatic response generating device
110: training data generation unit
120: model training unit
121: constructor
123: delimiter

Claims (12)

Learning data ( training data generating unit for generating training data; and
On the basis of the learning data, the second response message generated by the automatic response model includes a model learning unit that trains the automatic response model so that the similarity with the personality data is close, and the similarity with the interference data is farther away,
The automatic response model receives the learning data and generates the second response message based on the personality data. Based on the similarity between the interruption data and the second response message by receiving the learning data and receiving the learning data. to include a discriminator for determining the authenticity of the interference of the second response message,
The model learning unit sets the first loss function based on the degree of similarity between the personality data and the second response message and the disturbance authenticity label of the disturbance data to be true and the second response message to false. An apparatus for generating an automatic response that trains the automatic response model by using a third loss function based on a second loss function for determining. The method according to claim 1,
The personality data includes a plurality of sentences pre-classified as being uttered by a person having a specific personality. 3. The method of claim 2,
The interruption data includes an interruption sentence generated based on a degree of similarity with words included in the plurality of sentences included in the personality data and a label of whether the interruption sentence is authentic or not. 3. The method of claim 2,
The past conversation data includes the first response message to the first response request message generated based on a degree of similarity with a plurality of sentences included in the personality data. delete delete Learning data ( generating training data); and
On the basis of the learning data, the second response message generated by the automatic response model comprising the step of learning the automatic response model so that the similarity with the personality data is close, and the similarity with the interference data is far away,
The automatic response model receives the learning data and generates the second response message based on the personality data. Based on the similarity between the interruption data and the second response message by receiving the learning data and receiving the learning data. to include a discriminator for determining the authenticity of the interference of the second response message,
In the learning step, the first loss function based on the similarity between the personality data and the second response message and the interference authenticity label of the interference data are true, and the second response message is false. Learning the automatic response model using a third loss function based on the second loss function for determining as , automatic response generation method. 8. The method of claim 7,
The personality data includes a plurality of sentences pre-classified as being uttered by a person having a specific personality. 9. The method of claim 8,
The interruption data includes a disturbing sentence generated based on a degree of similarity with words included in the plurality of sentences included in the personality data, and a label of whether the interference sentence is generated based on the similarity to the personality data. A method for generating an automatic response, comprising: 9. The method of claim 8,
The past conversation data includes the first response message to the first response request message generated based on a degree of similarity with a plurality of sentences constituting the personality data. delete delete

Images