KR20200023049A - Method and system for generating multi-turn conversation response using deep learing generation model and multi-modal distribution - Google Patents

Abstract

Disclosed is a technology for automatically generating a conversation response. A method for generating a conversation response comprises the steps of: learning a conversation model, which models data distribution, by learning a generative adversarial network (GAN) in a latent variable space for conversation context including past utterance; and generating a conversation response using a latent variable sampled from the data distribution through the conversation model.

Description

METHOD AND SYSTEM FOR GENERATING MULTI-TURN CONVERSATION RESPONSE USING DEEP LEARING GENERATION MODEL AND MULTI-MODAL DISTRIBUTION}

The description below relates to a technique for automatically generating a conversation response.

An interface operating based on voice, such as an AI speaker of a home network service, receives a user's voice request through a microphone and synthesizes the answer voice to provide response information corresponding to the voice request. It may provide or output the audio of the content included in the response information.

For example, Korean Patent Laid-Open Publication No. 10-2011-0139797 describes a home media device and a home network system and method using the same. In a home network service, a home network service using a second communication network such as Wi-Fi in addition to a mobile communication network is used. It is possible to provide, and discloses a technology that allows a user to control a plurality of multimedia devices in the home through a voice command without a separate button operation.

These prior arts automatically generate and provide a dialogue response to a given question. However, not only is there a lack of diversity of responses because only the same response is generated each time for the same question or utterance, but the content of the utterance and response is often semantically irrelevant. It is difficult to respond to the conversation.

It provides a method and system that can automatically generate conversational responses to a variety of expressions and dialogue-wide contexts using the Wassstein Generative Adversarial Network (WGAN) and the Multimodal Gaussian Mixture prior distribution.

A conversation response generation method performed in a computer system, comprising: a conversation modeling a data distribution by learning a GAN (Generative Adversarial Network) in a latent variable space for a dialogue context including past utterances Training the model; And generating a conversation response using a latent variable sampled from the data distribution through the conversation model.

According to an aspect, the learning may include modeling a prior distribution and a posterior distribution of latent variables using a feed-forward neural network (FFNN).

According to another aspect, the learning step comprises prior distribution of latent variables by converting context-dependent random noise into samples for latent variables using a neural network. And modeling a posterior distribution.

According to another aspect, the dialogue model may maximize the log probability of the reconstructed response from the latent variable while minimizing the divergence between the pre- and post-distribution.

According to another aspect, the learning may include mapping a pre-distribution and a post-distribution for latent variables using an adversarial discriminator that distinguishes the pre-sample from the post-sample.

According to another aspect, the context-dependent random noise may be derived from a normal distribution computed from the dialogue context via each of a prior network and a recognition network, which are feed-forward neural networks (FFNNs). have.

According to another aspect, the generating may include generating a sample of the latent variable from the context-dependent random noise through the neural network and then decoding the generated latent variable into the dialogue response.

According to another aspect, the learning may include learning a multimodal response by sampling random noise using a Gaussian mixture prior tetwork.

According to another aspect, the step of learning the multi-modal response may learn the multi-modal response in the latent variable space by capturing multiple modes in a Gaussian distribution having one or more modes.

In combination with a computer, a computer program stored on a computer readable recording medium for executing the method for generating a conversation response is provided.

Provided is a computer-readable recording medium, in which a program for causing the computer to execute the conversation response generating method is recorded.

A computer system, comprising: a memory; And at least one processor coupled with the memory and configured to execute computer readable instructions contained in the memory, wherein the at least one processor is configured to execute a GAN within a latent variable space for a dialogue context including past speech. Learning a dialogue model modeling the data distribution by learning the, and generates a dialogue response using the latent variables sampled from the data distribution through the dialogue model.

According to embodiments of the present invention, prior and post distributions for latent variables are obtained by transforming context-dependent random noise using a neural network. Conversation models can be implemented that sample from the posterior distribution and minimize the Wasserstein distance between the two distributions, thereby generating a dialogue response for the entire context of the dialogue.

According to embodiments of the present invention, a Gaussian mixture prior network (PriNet) for enriching the potential space may be used to implement a dialogue model in consideration of the multimodal nature of the dialogue response. This allows you to generate more logical, useful, and diverse conversation responses.

1 is a diagram illustrating an example of a service environment using a voice-based interface according to an embodiment of the present invention.
2 is a diagram illustrating another example of a service environment using a voice-based interface according to an embodiment of the present invention.
3 is a diagram illustrating an example of a cloud AI platform according to an embodiment of the present invention.
4 is a block diagram illustrating an internal configuration of an electronic device and a server according to an embodiment of the present invention.
FIG. 5 illustrates a schematic diagram of a DialogWAE dialogue model for generating a multimodal response using a Wasserstein Auto Encoder (WAE), in an embodiment of the invention.
6 illustrates a detailed learning algorithm of the DialogWAE dialogue model according to an embodiment of the present invention.
7 illustrates an example of a response generated by the DialogWAE dialogue model according to an embodiment of the present invention.

Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings.

Embodiments of the present invention relate to a technique for automatically generating a conversation response.

Embodiments including those specifically disclosed herein may generate a multiturn communication dialogue response using a deep learning generation model and a multimodal distribution in a service environment using a voice-based interface, thereby allowing diversity and connectivity. Significant advantages are achieved in terms of accuracy, accuracy and efficiency.

1 is a diagram illustrating an example of a service environment using a voice-based interface according to an embodiment of the present invention. In the embodiment of FIG. 1, in a technology of connecting and controlling devices in a home such as a smart home or a home network service, an electronic device 100 that provides an interface operating based on voice may be used by the user 110. An example of controlling the light power of the indoor lighting device 120 connected to the electronic device 100 through the internal network by recognizing and analyzing the voice input “turn off the light” received according to the utterance is shown.

For example, in-house devices include home appliances such as televisions, personal computers (PCs), peripherals, air conditioners, refrigerators, robot cleaners, etc., as well as energy consumption such as water, electricity, and heating and heating devices, in addition to the above-described indoor lighting devices 120 It may include various devices that can be connected and controlled online such as security devices such as devices, door locks and surveillance cameras. Internal networks also include wired network technologies such as Ethernet, HomePNA, and IEEE 1394, and wireless network technologies such as Bluetooth, ultra wide band (UWB), ZigBee, Wireless 1394, and Home RF. Can be utilized.

The electronic device 100 may be one of devices in the home. For example, the electronic device 100 may be one of devices such as an AI speaker or a robot cleaner provided in the home. In addition, the electronic device 100 may be a mobile device of the user 110 such as a smart phone, a mobile phone, a notebook computer, a digital broadcasting terminal, a personal digital assistant (PDA), a portable multimedia player (PMP), a tablet PC, or the like. have. As such, the electronic device 100 is not particularly limited as long as the device includes a function capable of connecting with devices in the home to receive voice input of the user 110 and control the devices in the home. In addition, according to an embodiment, the above-described mobile devices of the user 110 may be included as devices in the home.

2 is a diagram illustrating another example of a service environment using a voice-based interface according to an embodiment of the present invention. FIG. 2 illustrates that the electronic device 100 providing an interface operating based on voice recognizes and analyzes a voice input “today's weather” received according to the user 110's speech from an external server 210 through an external network. This example shows how to obtain information about today's weather and output the obtained information as a voice such as "Weather in today is ...".

For example, external networks include personal area networks (PANs), local area networks (LANs), campus area networks (CANs), metropolitan area networks (MANs), wide area networks (WANs), broadband networks (BBNs), and the Internet. It may include any one or more of the network of.

In the embodiment of FIG. 2, the electronic device 100 may be one of the devices in the home or one of the mobile devices of the user 110. The electronic device 100 may include an external network and a function for receiving and processing a voice input of the user 110. The device is not particularly limited as long as the device includes a function for providing the user 110 with a service or content provided by the external server 210 by accessing the external server 210.

As such, the electronic device 100 according to the embodiments of the present invention may not be particularly limited as long as it is a device capable of processing a user command including a voice input received according to the speech of the user 110 through a voice-based interface. have. For example, the electronic device 100 may process a user command by directly recognizing and analyzing a voice input of the user and performing an operation suitable for the voice input. However, according to an embodiment, the electronic device 100 may recognize or recognize the voice input of the user. Analysis of voice input, synthesis of voice to be provided to the user, and the like may be performed through an external platform connected with the electronic device 100.

3 is a diagram illustrating an example of a cloud AI platform according to an embodiment of the present invention. 3 illustrates electronic devices 310, a cloud AI platform 320, and a content service 330.

For example, the electronic devices 310 may refer to devices provided in a home, and may include at least the electronic device 100 described above. Applications (hereinafter, referred to as apps) installed and driven on the electronic devices 310 or the electronic devices 310 may be associated with the cloud AI platform 320 through the interface connect 340. Here, the interface connect 340 may provide developers with a software development kit (SDK) and / or development documents for development of apps installed and driven on the electronic devices 310 or the electronic devices 310. In addition, the interface connect 340 includes an application program interface (API) that enables applications installed and driven on the electronic devices 310 or the electronic devices 310 to utilize functions provided by the cloud AI platform 320. Can provide. As a specific example, developers may develop a device or app using a software development kit (SDK) and / or a development document provided by the interface connect 340 using a cloud AI platform (API) provided by the interface connect 340. The functions provided by 320 may be utilized.

Here, the cloud AI platform 320 may provide a function for providing a voice-based service. For example, the cloud AI platform 320 recognizes the received voice, and a voice processing module 321 for synthesizing the output voice, and a vision processing module 322 for analyzing and processing the received image or video. A dialogue processing module 323 for determining an appropriate dialogue in order to output a suitable speech according to the received speech, a recommendation module 324 for recommending a function suitable for the received speech, and a sentence unit of artificial intelligence through data learning. It may include various modules for providing a speech-based service, such as neural machine translation (NMT, 325) to support the translation of the language.

For example, in the embodiments of FIGS. 1 and 2, the electronic device 100 may transmit the voice input of the user 110 to the cloud AI platform 320 using an API provided by the interface connect 340. have. In this case, the cloud AI platform 320 may recognize and analyze the received voice input by using the above-described modules 321 to 325, and synthesize and provide an appropriate answer voice according to the received voice input, Proper action can be recommended.

In addition, the extension kit 350 may provide a development kit that allows third-party content developers or companies to implement new voice-based functions based on the cloud AI platform 320. For example, in the embodiment of FIG. 2, the electronic device 100 may transmit a voice input of the user 110 to the external server 210, and the external server 210 may be provided through the expansion kit 350. The voice input may be transmitted to the cloud AI platform 320 through the API. In this case, as described above, the cloud AI platform 320 recognizes and analyzes the received voice input, synthesizes an appropriate answer voice, and provides recommendation information on a function to be processed through the voice input. ) Can be provided. As an example, in FIG. 2, the external server 210 can send a voice input “today's weather” to the cloud AI platform 320 and through recognition of the voice input “today's weather” from the cloud AI platform 320. The extracted keywords "today" and "weather" can be received. In this case, the external server 210 generates text information such as "today's weather ..." through the keywords "today" and "weather" to send the text information generated back to the cloud AI platform 320. Can be. In this case, the cloud AI platform 320 may synthesize text information into a voice and provide it to the external server 210. The external server 210 may transmit the synthesized voice to the electronic device 100, and the electronic device 100 outputs the synthesized voice "Weather of the day ..." through the speaker, thereby outputting from the user 110. The received voice input "Weather today" can be processed.

At this time, the electronic device 100 is for providing a device operation or content based on a conversation with a user.

4 is a block diagram illustrating an internal configuration of an electronic device and a server according to an embodiment of the present invention. The electronic device 410 of FIG. 4 may correspond to the electronic device 100 described above, and the server 420 may correspond to one computer device that implements the external server 210 or the cloud AI platform 320 described above. Can be.

The electronic device 410 and the server 420 may include memories 411 and 421, processors 412 and 422, communication modules 413 and 423, and input / output interfaces 414 and 424. The memories 411 and 421 are computer-readable recording media. Non-volatile memory such as random access memory (RAM), read only memory (ROM), disk drive, solid state drive (SSD), flash memory, and the like. It may include a permanent mass storage device. In this case, the non-volatile mass storage device such as a ROM, an SSD, a flash memory, a disk drive, or the like may be included in the electronic device 410 or the server 420 as a separate permanent storage device that is distinct from the memories 411 and 421. In addition, the memory 411 and 421 may store an operating system and at least one program code (for example, a code installed in the electronic device 410 and driven by the electronic device 410 to provide a specific service). Can be. These software components may be loaded from a computer readable recording medium separate from the memories 411 and 421. Such a separate computer-readable recording medium may include a computer-readable recording medium such as a floppy drive, disk, tape, DVD / CD-ROM drive, memory card, and the like. In other embodiments, software components may be loaded into the memory 411, 421 through a communication module 413, 423 that is not a computer readable recording medium. For example, the at least one program may be electronically based on a computer program (for example, the application described above) installed by files provided through a network 430 by a file distribution system that distributes installation files of developers or applications. The memory 411 of the device 410 may be loaded.

Processors 412 and 422 may be configured to process instructions of a computer program by performing basic arithmetic, logic and input / output operations. Instructions may be provided to the processors 412 and 422 by the memory 411 and 421 or the communication modules 413 and 423. For example, processors 412 and 422 may be configured to execute instructions received in accordance with program code stored in a recording device such as memory 411 and 421.

The communication modules 413 and 423 may provide a function for the electronic device 410 and the server 420 to communicate with each other through the network 430, and the electronic device 410 and / or the server 420 may be different from each other. It may provide a function for communicating with an electronic device or another server. In one example, a request generated by the processor 412 of the electronic device 410 according to a program code stored in a recording device such as a memory 411 is controlled by the server 420 through the network 430 under the control of the communication module 413. Can be delivered. Conversely, control signals, commands, contents, files, and the like provided according to the control of the processor 422 of the server 420 are communicated with the communication module 413 of the electronic device 410 via the communication module 423 and the network 430. ) May be received by the electronic device 410. For example, the control signal, command, content, file, etc. of the server 420 received through the communication module 413 may be transmitted to the processor 412 or the memory 411, and the content, file, etc. may be transferred to an electronic device ( 410 may be stored as a storage medium (permanent storage device described above) that may further include.

The input / output interface 414 may be a means for interfacing with the input / output device 415. For example, the input device may include a device such as a keyboard, a mouse, a microphone, a camera, and the like, and the output device may include a device such as a display, a speaker, a haptic feedback device, and the like. As another example, the input / output interface 414 may be a means for interfacing with a device in which functions for input and output are integrated into one, such as a touch screen. The input / output device 415 may be configured as the electronic device 410 and one device. In addition, the input / output interface 424 of the server 420 may be a means for interfacing with an apparatus (not shown) for input or output that may be connected to or included in the server 420. More specifically, the processor 412 of the electronic device 410 processes a command of a computer program loaded in the memory 411, and includes a service screen configured by using data provided by the server 420 or another electronic device. The content may be displayed on the display via the input / output interface 414.

Also, in other embodiments, the electronic device 410 and server 420 may include fewer or more components than the components of FIG. 4. However, there is no need to clearly show most prior art components. For example, the electronic device 410 may be implemented to include at least some of the above-described input and output devices 415 or other components such as a transceiver, a global positioning system (GPS) module, a camera, various sensors, a database, and the like. It may also include more. More specifically, when the electronic device 410 is a smartphone, an acceleration sensor, a gyro sensor, a motion sensor, a camera module, various physical buttons, a button using a touch panel, an input / output port, and vibration are generally included in the smartphone. Various components such as a vibrator may be implemented to be further included in the electronic device 410.

In the present exemplary embodiment, the electronic device 410 may basically include a microphone for receiving a voice input of the user as the input / output device 415, and may output a sound such as an answer voice or audio content corresponding to the voice input of the user. A speaker for outputting may be further included as the input / output device 415.

The present invention proposes a dialogue model (hereinafter referred to as DialogWAE dialogue model) for generating a multimodal response using a conditional Wasserstein AutoEncoder (WAE).

Dialog response generation is an old topic in natural language research. Most recent approaches to data-driven neural network dialogue modeling are mainly based on seq2seq (sequence-to-sequence) learning or memory networks. However, in the seq2seq dialogue model, there is a difficulty in generating meaningful, diverse, and subjective responses, and in the case of memory network-based models, there are issues such as model size and speed due to memory increase.

VAE (Variational AutoEncoder) has shown promising results in solving problems with the seq2seq conversation model. VAE uses a cognitive network to calculate an approximate posterior distribution of latent variables that express high-level semantics for the response. Decode the response word by word, subject to samples of the distribution. For example, latent variables capture topics, tones, or high-level syntactic properties to produce more diverse responses. However, most VAE dialogue models limit the responses generated to a relatively simple (eg, single-modal) range by matching approximate posterior distributions for latent variables to a simple prior distribution, such as a standard normal distribution. do.

In addition to VAEs, a Generic Adversarial Network (GAN) -based dialogue model has emerged that directly models the distribution of responses, but this is due to the non-diffenrentiability of adversarial training on discrete tokens. The problem is that it is tricky.

In addition, a hybrid conversational model using reinforcement learning (RL) in the GAN has emerged, in which the discriminator uses the expected value as a reward for learning the generator. do. However, reinforcement learning is unstable by high fluctuations in the gradient estimates, which makes the GAN model differentiable with an approximate word embedding layer, adding only word-level variability. This is not suitable for expressing high-level response variability such as topics and situations.

Accordingly, the present invention proposes a dialog dialogue model, DialogWAE, which is a new variant of GAN for neural dialogue modeling. Unlike the existing VAE dialogue model, which applies a simple distribution to latent variables, the DialogWAE dialogue model according to the present invention models the data distribution by learning the GAN in a latent variable space. In particular, the DialogWAE dialogue model according to the present invention uses a neural network to transform context-dependent random noise to sample from pre- and post-distribution for latent variables, and to pre- and post-post Minimize the Wasserstein distance between distributions. In addition, the DialogWAE dialogue model according to the present invention considers the multimodal nature of the response by using a Gaussian mixture prior network (PriNet). Hostile learning through a Gaussian mixed dictionary network enables DialogWAE to capture rich potential space, which can generate more logical, useful and diverse responses.

The DialogWAE dialogue model according to the present invention is (1) a GAN-based model for neural dialogue modeling using GAN to generate samples for latent variables, and (2) a Gaussian mixture dictionary for sampling random noise from a multimodal prior distribution. Include a network. Accordingly, the DialogWAE dialogue model according to the present invention may be implemented as a GAN dialogue model using a multimodal latent structure.

Encoder-decoder variants: There are a number of variants to address the "safe response" problem for pure encoder-decoder dialogue models. The DialogWAE dialogue model according to the present invention is distinguished from the existing dialogue model because it does not require extra information such as situation and subject.

VAE conversation models: The Variable Auto Encoder (VAE) is one of the most popular frameworks for conversational modeling. In order to solve the "posterior collapse", which is the main problem of the VAE dialogue model, the model introduces auxiliary bag-of-words loss into the decoder, dialogue acts and speaker profiles ( Knowledge-guided CVAE model incorporating auxiliary dialogue information such as speaker profiles, using Gaussian to transform Gaussian noise, sampling from pre and post distributions of latent variables and KL divergence A collaborative CVAE model that corresponds to the pre and post distribution of Gaussian noise, and a variant hierarchical conversation RNN (VHCR) that integrates the hierarchical structure of latent variables and utterance drop regularization. Models appeared. The DialogWAE dialogue model according to the present invention solves the limitations of the VAE dialogue model by using a GAN architecture in the latent space.

GAN conversation models: Although GAN / CGAN has shown great success in image generation, applying it to natural language conversation generators is not an easy task. This is due to the non-differentiable nature of natural language tokens. The problem can be solved by combining the GAN with reinforcement learning, where the discriminator expects a reward to optimize the generator. However, reinforcement learning is unstable by the high variation of the sampled slope. In addition, the GAN dialogue model allows seq2seq GAN to be differentiated by directly multiplying word probabilities acquired by the decoder with corresponding word vectors, which are approximately vectorized to the target sequence. Derive an expression. However, the above approach merely guarantees diversity at the word level rather than the overall response level. The DialogWAE dialogue model according to the present invention is distinguished from the existing GAN dialogue model in that it forms a distribution of responses in a high level of latent space instead of direct tokens and does not rely on RL with large gradient fluctuations.

The DialogWAE dialogue model according to the present invention may be implemented in a computer system such as the electronic device 410 or the server 420 described above, and may generate a multiturn type dialogue response based on a deep learning generation model and a multimodal distribution. Can be. In this case, the processor 412 or 422 of the computer system 410 or 420 may be implemented to execute a control instruction according to a code of an operating system included in the memory 411 or 421 or a code of at least one program. . Here, the processor 412 or 422 performs a method of generating a dialogue response based on the DialogWAE dialogue model, which will be described below by the computer system 410 or 420 according to a control command provided by code stored in the computer system 410 or 420. Computer system 410 or 420 can be controlled to do so.

The DialogWAE dialogue model according to the present invention will be described in detail as follows.

Problem Statement

가

개의 발화(utterance)에 대한 대화 발화(dialogue utterance)를 나타낸다고 하자. 여기서,

는 하나의 발화를 나타내고,

는

내의 n번째 단어(word)를 나타낸다.

end

Suppose a dialogue utterance of the dog's utterance. here,

Represents one utterance,

Is

Represents the nth word in.

는

개의 과거 발화(historical utterances)인 대화 문맥(dialogue context)을 나타내고,

가 다음 발화를 의미하는 응답(response)이라고 하자.And,

Is

Represents the dialogue context, which is the historical utterances of dogs,

Is a response that means the next utterance.

를 추정하는 것이다.The goal in the DialogWAE dialogue model is to be a conditional distribution of the current response given a past utterance.

To estimate.

와

가 이산 토큰(discrete tokens)에 대한 배열(sequence)이기 때문에 이것들 간의 직접적인 결합을 찾는 것은 쉽지 않다. 대신에, 응답에 대한 높은 레벨의 표현식을 나타내는 연속적인 잠재 변수

를 도입한다.

Wow

Since is a sequence of discrete tokens, finding a direct combination between them is not easy. Instead, a series of latent variables that represent a high level expression of the response.

Introduce.

는 잠재 공간

상의 분포

로부터 샘플링되며, 그 후 응답

는

를 사용하여

로부터 디코딩 된다. DialogWAE 대화 모델 하에서 응답의 확률은 방정식 1과 같이 정의될 수 있다.Response generation can be thought of as a two-step procedure, where latent variables

Latent space

Distribution

Is sampled from and then responded to

Is

use with

Decoded from Under the DialogWAE dialogue model, the probability of the response can be defined as in Equation 1.

Equation 1

를 주변화(marginalize out)하는 것은 다루기 힘들기 때문에 정확한 로그 확률을 계산하기가 어렵다. 그러므로, 본 발명에서는 잠재 변수

에 대한 사후 분포를

로 근사화하며, 이는 인지 네트워크(recognition network, RecNet)라고 불리는 신경망에 의해 계산될 수 있다. 이러한 근사적인 사후 분포를 사용하여 증거 하한 값(evidence lower bound, ELBO)을 대신 계산할 수 있다(방정식 2).Latent variables

Marginalize out is difficult, so it is difficult to calculate the exact log probability. Therefore, in the present invention, latent variables

Post-distribution for

Approximation can be calculated by a neural network called a cognition network (RecNet). Using this approximate posterior distribution, the lower bound of evidence (ELBO) can be calculated instead (Equation 2).

[Equation 2]

는

가 주어졌을 때

에 대한 사전 분포를 나타내며, 사전 네트워크(prior network)라고 불리는 신경망을 통해 모델링될 수 있다.here,

Is

Is given

Represents a prior distribution for and can be modeled through a neural network called a prior network.

Conditional Wasserstein  Auto-Encoders for Dialogue Modeling

가 정규 분포과 같이 단순한 사전 분포를 따른다고 가정한다. 그러나, 실제 응답의 잠재 공간은 더 복잡하며 그렇게 단순한 분포로 추정하기는 어렵다. 이것은 대게 사후 붕괴(posterior collapse) 문제를 야기한다.The existing VAE dialogue model is a latent variable

Assume that follows a simple dictionary distribution like a normal distribution. However, the latent space of the actual response is more complex and difficult to estimate with such a simple distribution. This usually causes a problem of posterior collapse.

에 대한 분포를 모델링한다.The DialogWAE dialogue model according to the present invention is based on learning GAN in latent space based on GAN and Adversarial Auto-Encoder (AAE).

Model the distribution for.

를 변환함으로써 잠재 변수에 대한 사전 및 사후 분포로부터 샘플링한다.In the present invention, random noise using a neural network.

Sample from the pre and post distributions for latent variables by transforming

은 생성자

에 의해 문맥-의존 랜덤 노이즈(context-dependent random noise)

로부터 생성되는 반면, 근사 사후 샘플

은 생성자

에 의해 문맥-의존 랜덤 노이즈

로부터 생성된다.

와

는 평균과 공분산 행렬(대각선 행렬로 가정)이 피드 포워드 신경망(feed-forward neural networks)인 사전 네트워크(prior network) 및 인지 네트워크(recognition network) 각각을 통해

로부터 계산되는 정규 분포에서 도출된다(방정식 3과 방정식 4).Especially, advance sample

Silver constructor

Context-dependent random noise

Approximate post-sample, while generated from

Silver constructor

Context-dependent random noise by

Is generated from

Wow

Is derived from each of the prior and recognition networks where the mean and covariance matrices (assuming a diagonal matrix) are feed-forward neural networks.

It is derived from the normal distribution calculated from (Equations 3 and 4).

[Equation 3]

[Equation 4]

및

는 피드 포워드 신경망이다. 본 발명에 따른 DialogWAE 대화 모델의 목표는

과

간의 다이버전스(divergence)를 최소화하는 반면,

로부터 재구성된(reconstructed) 응답의 로그 확률을 최대화하는 것이다.here,

And

Is a feedforward neural network. The goal of the DialogWAE dialog model according to the present invention is

and

Minimize divergence between livers,

To maximize the log probability of the reconstructed response.

The DialogWAE dialogue model according to the invention relates to the problem of equation (5).

[Equation 5]

및 사후 분포

는 각각 방정식 3과 방정식 4를 구현하는 신경망이다.

는 디코더이고,

는 두 분포 간의 바세르슈타인 거리(Wasserstein distance)를 의미한다.Where the dictionary distribution

And posterior distribution

Are neural networks that implement equations 3 and 4.

Is a decoder,

Is the Wasserstein distance between the two distributions.

5 shows a schematic diagram of the DialogWAE dialogue model according to the present invention.

를 포함하는) 각 발화를 실수 벡터(real-valued vector)로 변환한다.An utterance encoder (RNN) 501 may be used to

Each utterance (including) is converted to a real-valued vector.

를 계산한다. 문맥 인코더(502)의 마지막 숨겨진 상태는 문맥 표현식(context representation)으로서 사용된다.A context encoder (RNN) 502 receives as input a concatenation of the encoding vector and the conversation floor 504 at the i th speech in the context, the hidden state.

Calculate The last hidden state of the context encoder 502 is used as the context representation.

를 변환하는 사전 네트워크(prior network, PriNet)(510)로부터 랜덤 노이즈

(511)를 도출한다. 그 후, 생성자

(512)는 피드-포워드 네트워크를 통해 노이즈(511)로부터 잠재 변수

(513)의 샘플을 생성한다. 디코더(decoder) RNN은 생성된

(513)를 응답으로 디코딩한다.At the time of creation, the DialogWAE dialogue model uses a feed-forward network, followed by two matrix multiplications that result in mean and diagonal matrix covariances, respectively.

Random noise from a prior network (prior network, PriNet) 510 for converting

(511) is derived. After that, the constructor

512 is a latent variable from noise 511 via a feed-forward network.

A sample of 513 is generated. Decoder RNN is generated

Decode 513 as a response.

와 응답

를 조건으로 잠재 변수에 대한 사후 분포를 추론한다. 인지 네트워크(recognition network, RecNet)(520)는

와

의 연결을 입력으로 받아서 정규 평균과 대각선 행렬 공분산 각각을 정의하는 두 가지 행렬 곱셈이 뒤따르는 피드-포워드 네트워크를 통해 변환한다. 가우시안 노이즈(Gaussian noise)

(521)는 재매개화 트릭(re-parametrization trick)을 사용해 인지 네트워크(520)로부터 도출된다. 그 후, 생성자(generator)

(522)는 피드-포워드 네트워크를 통하여 가우시안 노이즈

(521)를 잠재 변수

(523)에 대한 샘플로 변환한다. 응답 디코더(RNN)(503)는 재구성 손실(reconstruction loss)을 방정식 6을 통해 계산한다.At the time of learning, the DialogWAE dialogue model

And response

Infer the posterior distribution of the latent variables on the condition of Recognition network (RecNet) 520

Wow

Takes a concatenation of, and converts it through a feed-forward network followed by two matrix multiplications that define each of the normal mean and the diagonal matrix covariance. Gaussian noise

521 is derived from cognitive network 520 using a re-parametrization trick. Then the generator

522 is Gaussian noise over a feed-forward network

521 latent variables

Convert to sample for 523. Response decoder (RNN) 503 calculates the reconstruction loss through equation (6).

[Equation 6]

(530)를 도입함으로써

에 대한 사전 분포와 근사 사후 분포를 대응시킨다.

(530)는 입력으로

와

의 연결을 받고 실수 값(real value)을 출력하는 피드-포워드 신경망으로서 구현된다.Adversarial discriminator that distinguishes a pre-sample from a post-sample

By introducing 530

Correlate the pre- and approximate post-distribution for.

530 is the input

Wow

It is implemented as a feed-forward neural network that accepts a connection of and outputs a real value.

(530)를 학습한다.By minimizing the discriminator loss, as in equation 7

Learn 530.

[Equation 7]

와 함께 화자 정보를 학습시킴으로써 화자 스타일을 고려하여

에 대한 분포를 모델링할 수 있다. 따라서, 본 발명에 따른 DialogWAE 대화 모델은 주어진 문맥에 대하여 해당 화자의 대화 스타일에 맞는 응답을 생성하여 제공할 수 있다.Although the specific illustration is omitted, the DialogWAE dialogue model does not have the dialogue context in the latent space.

By considering the speaker style by learning the speaker information with

You can model the distribution for. Accordingly, the DialogWAE dialogue model according to the present invention can generate and provide a response that matches the dialogue style of the speaker for a given context.

Multimodal  Response generation with Gaussian  Mixture Prior Network

In conditional hostile auto-encoder (AAE) architectures, it is a common application that the pre-distribution is normal. However, responses usually have a multimodal nature that reflects multiple situations, subjects, and emotions that are equally likely. Random noise with a normal distribution can limit the generator to creating latent space in a single dominant mode by the single-modal nature of the Gaussian distribution. As a result, the generated responses may follow a simple prototype.

)이라고 불리는 가우시안 분포의 혼합을 캡쳐하도록 한다. 여기서,

,

및

는 k 번째 구성요소의 파라미터들이다. 이는 두 단계의 생성 절차로 잠재 변수 공간에서 다중모달 매니폴드(multimodal manifold)를 학습하게 한다. 첫 번째 단계는

로 구성요소

를 선택하는 것이고, 다음은 선택된 구성요소로 방정식 8과 같이 가우시안 노이즈를 샘플링하는 것이다.In order to capture multiple modes in a probability distribution over latent variables, the present invention uses a distribution that may have more than one mode K. Each time, the noise that produces the latent variable is selected in one of these modes. In order to achieve this, in the DialogWAE dialogue model according to the present invention, the dictionary network is called GMM (

Capture a mixture of Gaussian distributions called). here,

,

And

Are parameters of the k th component. This is a two-step creation process that allows you to learn multimodal manifolds in latent variable spaces. First step

Component

Next, we sample the Gaussian noise with the selected component, as shown in equation (8).

[Equation 8]

는 클래스 확률

를 갖는 구성요소 지시자(indicator)이고,

는 GMM의 k 번째 구성요소의 혼합 계수(mixture coefficient)이다.here,

Is the class probability

Is a component indicator with

Is the mixing coefficient of the k-th component of the GMM.

는 방정식 9와 같이 계산된다.

Is calculated as in equation 9.

Equation 9

에 대한 인스턴스를 샘플링하기 위하여 방정식 10과 같이 Gumbel-softmax 재매개화를 사용한다.Instead of accurate sampling, the component indicators in the present invention

Use Gumbel-softmax remediation as shown in Equation 10 to sample the instances for.

[Equation 10]

는 방정식 11과 같이 계산되는 Gumbel 노이즈이다.here

Is the Gumbel noise calculated as

Equation 11

는 모든 실험에서 0.1로 설정된 softmax 온도이다.

Is the softmax temperature set to 0.1 in all experiments.

Training

An example of a detailed learning procedure of the DialogWAE dialogue model according to the present invention is the same as Algorithm 1 shown in FIG. 6.

Referring to FIG. 6, the DialogWAE dialogue model learns in epochwise terms until convergence is reached. The dialogue model is trained by iteratively alternating between the Auto Encoder (AE) stage where the reconstruction loss of the decoded responses at each epoch is minimized and the GAN stage where the total post-distribution of latent variables matches the conditional pre-distribution. In one example, the detailed learning procedure of the DialogWAE dialogue model is the same as Algorithm 1 shown in FIG. 6.

7 shows an example of a response generated by the DialogWAE dialogue model according to the present invention in a daily dialogue data set. In the table of FIG. 7, '_eou_' represents a change in turn, and 'Eg.i' represents an i th response.

7 is a context-response pair consisting of responses generated by the dialogue model for a given context, generated by the existing model (CVAE-CO) and by the DialogWAE dialogue model (DialogWAE-GMP) according to the present invention. Compare the responses.

As shown in FIG. 7, it can be seen that the DialogWAE Dialog Model (DialogWAE-GMP) generates more consistent and diverse responses while addressing various possible aspects. In addition, it can be seen that the DialogWAE Dialog Model (DialogWAE-GMP) provides a response containing longer and more beneficial contents than the response of the conventional model (CVAE-CO).

The response generated by the existing model (CVAE-CO) shows a relatively limited change and some variation in the response content, but most of the similar expressions (eg, “how much”, etc.) are repeated.

As described above, according to the embodiments of the present invention, a neural network can be used to transform the context-dependent random noise to sample a pre- and post-distribution distribution of latent variables and to implement a dialogue model that minimizes the Wasserstein distance between the two distributions. This allows you to generate a dialogue response for the entire context of the conversation. Furthermore, Gaussian mixed dictionary networks to enrich the potential space can be used to enhance conversation models that take into account the multimodal nature of the conversation responses, thereby generating more logical, useful and diverse conversation responses.

The apparatus described above may be implemented as a hardware component, a software component, and / or a combination of hardware components and software components. For example, the devices and components described in the embodiments may include a processor, a controller, an arithmetic logic unit (ALU), a digital signal processor, a microcomputer, a field programmable gate array (FPGA), a programmable PLU (programmable). It can be implemented using one or more general purpose or special purpose computers, such as logic units, microprocessors, or any other device capable of executing and responding to instructions. The processing device may execute an operating system (OS) and one or more software applications running on the operating system. The processing device may also access, store, manipulate, process, and generate data in response to the execution of the software. For convenience of explanation, one processing device may be described as being used, but one of ordinary skill in the art will appreciate that the processing device includes a plurality of processing elements and / or a plurality of types of processing elements. It can be seen that it may include. For example, the processing device may include a plurality of processors or one processor and one controller. In addition, other processing configurations are possible, such as parallel processors.

The software may include a computer program, code, instructions, or a combination of one or more of the above, and may configure the processing device to operate as desired, or process independently or collectively. You can command the device. The software and / or data may be embodied in any type of machine, component, physical device, computer storage medium or device in order to be interpreted by or provided to the processing device or to provide instructions or data to the processing device. have. The software may be distributed over networked computer systems so that they may be stored or executed in a distributed manner. The software and data may be stored on one or more computer readable recording media.

The method according to the embodiment may be embodied in the form of program instructions that can be executed by various computer means and recorded in a computer readable medium. In this case, the medium may be to continuously store a computer executable program or to temporarily store the program for execution or download. In addition, the medium may be a variety of recording means or storage means in the form of a single or several hardware combined, not limited to a medium directly connected to any computer system, it may be distributed on the network. Examples of media include magnetic media such as hard disks, floppy disks and magnetic tape, optical recording media such as CD-ROMs and DVDs, magneto-optical media such as floptical disks, And ROM, RAM, flash memory, and the like, configured to store program instructions. In addition, examples of another medium may include a recording medium or a storage medium managed by an app store that distributes an application, a site that supplies or distributes various software, a server, or the like.

Although the embodiments have been described by the limited embodiments and the drawings as described above, various modifications and variations are possible to those skilled in the art from the above description. For example, the described techniques may be performed in a different order than the described method, and / or components of the described systems, structures, devices, circuits, etc. may be combined or combined in a different form than the described method, or other components. Or, even if replaced or substituted by equivalents, an appropriate result can be achieved.

Therefore, other implementations, other embodiments, and equivalents to the claims are within the scope of the claims that follow.

Claims (15)

In the method of generating a conversation response performed in a computer system,
Learning a dialogue model modeling data distribution by learning a GAN (Generative Adversarial Network) in a latent variable space for a dialogue context including past utterances; And
Generating a dialogue response using latent variables sampled from the data distribution through the dialogue model
Conversation response generation method comprising a. The method of claim 1,
The learning step,
Modeling prior and posterior distributions of latent variables using a feed-forward neural network (FFNN)
Conversation response generation method comprising a. The method of claim 1,
The learning step,
Modeling prior and posterior distributions of latent variables by converting context-dependent random noise into samples of latent variables using neural networks
Conversation response generation method comprising a. The method of claim 3,
The dialogue model maximizes the log probability of the reconstructed response from the latent variable while minimizing the divergence between the pre and post distributions.
Method for generating a conversation response characterized in that. The method of claim 3,
The learning step,
Mapping the pre- and post-distribution of latent variables using an adversarial discriminator that distinguishes the pre-sample from the post-sample
Conversation response generation method comprising a. The method of claim 3,
The context-dependent random noise is derived from a normal distribution computed from the dialogue context via each of a prior network and a recognition network, which are feed-forward neural networks (FFNNs).
Method for generating a conversation response characterized in that. The method of claim 3,
The generating step,
Generating samples of the latent variables from the context-dependent random noise via the neural network and then decoding the generated latent variables into the dialogue response.
Conversation response generation method comprising a. The method of claim 1,
The learning step,
Learning multimodal responses by sampling random noise using a Gaussian mixture prior tetwork
Conversation response generation method comprising a. The method of claim 8,
Learning the multimodal response,
Learning multimodal responses in the latent variable space by capturing multiple modes in a Gaussian distribution with one or more modes
Method for generating a conversation response characterized in that. A computer program stored in a computer readable recording medium, coupled with a computer, for causing a computer to execute the method of generating a conversational response of any one of claims 1 to 9. A computer-readable recording medium having a program recorded therein for executing a method of generating a conversation response according to any one of claims 1 to 9. In a computer system,
Memory; And
At least one processor coupled with the memory and configured to execute computer readable instructions contained in the memory
Including,
The at least one processor,
We learn a dialogue model that models the distribution of data by learning the GAN in a latent variable space for dialogue contexts that include past utterances,
Generating a dialogue response using latent variables sampled from the data distribution through the dialogue model
Computer system characterized in that. The method of claim 12,
The at least one processor,
Modeling pre and post distributions of latent variables using FFNN
Computer system characterized in that. The method of claim 12,
The at least one processor,
Modeling the pre and post distribution of latent variables by transforming context-dependent random noise into samples of latent variables using neural networks
Computer system characterized in that. The method according to claim 12,
The at least one processor,
Learning multimodal responses by sampling random noise using a Gaussian mixed dictionary network
Computer system characterized in that.

Images