KR102446294B1 - Utterance manipulation apparatus for retrieval-based response selection and method thereof - Google Patents

Abstract

The present invention relates to an utterance manipulation apparatus and method for selecting a response of a next utterance in a search-based dialog system. An utterance insertion module that first reconstructs the corresponding original dialogue in a separate form, and adds a specific insertion token before and after each utterance of the first reconstructed original dialogue so that the consistency of the original dialogue can be learned; , utterances with different meanings are randomly extracted from other dialogues other than the corresponding original dialogues, and the corresponding original dialogues are produced in the form of inserting the extracted utterances with different meanings before/after any one of the corresponding original dialogues. 2 After reconstructing, a utterance deletion module that adds a specific deletion token in front of each utterance of the second reconstructed original dialog so that consistency between utterances in the corresponding original dialog can be learned, and after extracting a random last utterance from the original dialog , a specific search token is provided in front of each utterance of the third reconstructed original dialogue so that the extracted last utterance is third reconstructed in a form separated from the corresponding original dialogue, and sequence information between utterances in the corresponding original dialogue can be learned. A preset business by applying a utterance search module to be added, a specific insertion token added from the utterance insertion module, a specific deletion token added from the utterance deletion module, and a specific search token added from the utterance search module as output features Based on the unsupervised learning method, by including a language learning model that learns whether the location of the specific deletion token, search token, and search token is the actual insertion, deletion, and search location based on the language model In designing the utterance prediction model, data augmentation effects can be achieved through the self-supervised learning method using a special token, and since it is a self-supervised learning method, a person constructs data separately It has an effect that you don't need to do.

Description

An utterance manipulation apparatus and method for selecting a response of the next utterance in a search-based dialog system

The present invention relates to an utterance manipulation apparatus and method for selecting a response of a next utterance in a retrieval-based dialog system.

Today, with the rise in the value of data utilization and the development of language processing technology, various convergence technologies and service markets using artificial intelligence-based technologies such as deep learning and natural language processing are being activated.

In the midst of the advent of the intelligent information society, as mobile-oriented interactive interfaces such as mobile messengers are popularized, chatbots, which are artificial intelligence-based services that allow users to feel the artificial intelligence ecosystem the most, are available for public/ It is spreading and growing across all sectors of private industry.

As the development of artificial intelligence-based technologies such as natural language processing, deep learning, and machine learning accelerated due to the chatbot craze across all industries, the level of chatbot service has changed from pattern matching and keyword-oriented text processing-based chatbots It is developing into an intelligent chatbot using learning, natural language processing, and machine learning technology, and is expected to develop into a personalized emotional chatbot that can exchange even human emotions/emotions as the technology matures.

With the advent of artificial intelligence and deep learning, computer-human chatbot systems have also developed a lot. The chatbot system aims to keep conversations with users natural and consistent.

There are two methods for creating such a chatbot: a search-based method and a generative-based method. The present invention relates to a search-based chatbot system.

The existing search-based dialog system uses a dialog-response matching method based on a recurrent neural network or an attention mechanism.

Recently, various language models such as BERT, ALBERT, RoBERTa, and ELECTRA, which are based on prior learning, have been proposed, and they show very good performance in various natural language related services or applications such as search, Q&A, and chatbot.

Most of the existing search-based methods design and train models in a method based on semantic matching between dialogue and response. However, when learning in this way, not only the utterances that will influence the decision of the next response, but also information on the utterances that interfere with the decision may be mixed together.

Accordingly, there arises a problem in that the prediction of the next utterance proceeds based on the semantic similarity with the entire simple conversation.

Domestic Registered Patent No. 10-2119468 (2020.06.05. Announcement)

The present invention has been devised to solve the above problems, and an object of the present invention is to design a speech prediction model based on a language model, using a specific token (Self Supervised Learning) ) method to provide a utterance manipulation device and method for selecting a response of a next utterance in a search-based dialog system capable of generating a data augmentation effect.

In order to achieve the above object, a first aspect of the present invention is to extract an arbitrary utterance from an original dialogue input from the outside, and separate the extracted utterances from the corresponding original dialogue in a first manner. a speech insertion module for reconstructing and adding specific insertion tokens before and after each utterance of the first reconstructed original dialogue so that consistency of the corresponding original dialogue can be learned; An utterance with a different meaning is randomly extracted from any other dialogue other than the corresponding original dialogue, and the corresponding original dialogue is produced in a form in which the extracted utterance with a different meaning is inserted before/after any one of the corresponding original dialogues. 2 a speech deletion module for reconstructing and adding a specific deletion token in front of each utterance of the second reconstructed original conversation so that consistency between utterances in the corresponding original conversation can be learned; After extracting any last utterance from the corresponding original conversation, the extracted last utterance is third reconstructed in a form separated from the corresponding original dialogue, and the third reconstruction is performed to learn sequence information between utterances in the corresponding original dialogue. an utterance search module for adding a specific search token in front of each utterance of the original conversation; and a predetermined unsupervised learning ( The response of the next utterance in a search-based dialog system including a language learning model that learns whether the specific deletion token, search token, and the location of the search token are actually inserted, deleted, and searched locations based on the Unsupervised Learning) method An object of the present invention is to provide an ignition manipulation device for selection.

A second aspect of the present invention provides an utterance manipulation method for selecting a response of a next utterance in a retrieval-based dialog system, wherein (a) an arbitrary utterance in an original conversation inputted from the outside through an utterance insertion module after extracting, configuring each of the extracted utterances in a form separated from the corresponding original dialogue; (b) adding a specific insertion token before each utterance and after the last utterance of the original dialogue configured in step (a) so that the consistency of the corresponding original dialogue can be learned through the utterance insertion module; (c) After randomly extracting utterances with different meanings from other conversations other than the corresponding original conversations through the utterance deletion module, the extracted utterances with different meanings are inserted before/after any one of the original conversations. Constructing in a put form; (d) adding a specific deletion token in front of each utterance of the original conversation configured in step (c) so that consistency between utterances in the corresponding original conversation can be learned through the utterance deletion module; (e) extracting a last utterance from the corresponding original conversation through the utterance search module, and then configuring the extracted last utterance in a form separated from the original conversation; (f) adding a specific search token in front of each utterance of the original conversation configured in step (e) so that order information between utterances in the corresponding original conversation can be learned through the utterance search module; and (g) outputting a specific insertion token added in step (b), a specific deletion token added in step (d), and a specific search token added in step (f) through the language learning model Applied as a feature, based on a preset unsupervised learning method, the specific deletion token, the search token, and the location of the search token include the step of learning whether the location is actually inserted, deleted, and searched. An object of the present invention is to provide an utterance manipulation method for selecting a response of a next utterance in a search-based dialog system characterized by the present invention.

A third aspect of the present invention provides a computer-readable recording medium in which a program capable of executing a utterance manipulation method for selecting a response of a next utterance in the above-described search-based conversation system is recorded.

The utterance manipulation method for selecting a response of the next utterance in the search-based dialog system according to the present invention may be implemented as computer-readable codes 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.

For example, computer-readable recording media include ROM, RAM, CD-ROM, magnetic tape, hard disk, floppy disk, removable storage device, and non-volatile memory (Flash Memory). , and optical data storage devices.

According to the speech manipulation apparatus and method for selecting the response of the next speech in the search-based dialogue system of the present invention as described above, when designing a speech prediction model based on a language model, a specific token (Special Token) ) can produce data augmentation effect through the self-supervised learning method using the

In addition, according to the present invention, it can be applied to techniques for predicting the next utterance in a retrieval-based dialog system, as well as unsupervised learning based on a large-capacity corpus. It has the advantage of being able to apply a language model to a conversational system.

In addition, according to the present invention, there is an advantage that can be applied to all language models and can also be applied to existing speech prediction models.

In addition, according to the present invention, if there is a language model (eg, BERT, ELECTRA, etc.) of each language that is not limited to one language, there is an advantage that can be easily applied.

In addition, according to the present invention, in the case of the existing open domain dialog system, there are many difficulties in collecting data on each dialog, but by using a language model learned in advance with a large corpus, even with a small amount of dialog data, it is It has the advantage of better performance than designing a model and proceeding with training.

1 is an overall block diagram illustrating a speech manipulation apparatus for selecting a response of a next utterance in a search-based conversation system according to an embodiment of the present invention.
2 and 3 are overall flowcharts for explaining a utterance manipulation method for selecting a response of a next utterance in a search-based conversation system according to an embodiment of the present invention.
4 is a structural diagram illustrating an example of an utterance manipulation method for selecting a response of a next utterance in a search-based conversation system according to an embodiment of the present invention.

The above-described objects, features and advantages will be described below in detail with reference to the accompanying drawings, and accordingly, those skilled in the art to which the present invention pertains will be able to easily implement the technical idea of the present invention. In describing the present invention, if it is determined that a detailed description of a known technology related to the present invention may unnecessarily obscure the gist of the present invention, the detailed description will be omitted.

Terms including ordinal numbers such as first, second, etc. may be used to describe various elements, but the elements are not limited by the terms. The above terms are used only for the purpose of distinguishing one component from another. For example, without departing from the scope of the present invention, a first component may be referred to as a second component, and similarly, a second component may also be referred to as a first component. The terms used in the present application are only used to describe specific embodiments, and are not intended to limit the present invention. The singular expression includes the plural expression unless the context clearly dictates otherwise.

The terms used in the present invention have been selected as currently widely used general terms as possible while considering the functions in the present invention, but these may vary depending on the intention or precedent of a person skilled in the art, the emergence of new technology, and the like. In addition, in a specific case, there is a term arbitrarily selected by the applicant, and in this case, the meaning will be described in detail in the description of the corresponding invention. Therefore, the term used in the present invention should be defined based on the meaning of the term and the overall content of the present invention, rather than the name of a simple term.

In the entire specification, when a part "includes" a certain component, it means that other components may be further included, rather than excluding other components, unless otherwise stated. In addition, terms such as "...unit" and "module" described in the specification 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. .

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings. However, the embodiments of the present invention illustrated below may be modified in various other forms, and the scope of the present invention is not limited to the embodiments described below. The embodiments of the present invention are provided to more completely explain the present invention to those of ordinary skill in the art.

Each block in the accompanying block diagram and combinations of steps in the flowchart may be executed by computer program instructions (execution engine), which computer program instructions may be executed by a processor of a general-purpose computer, special-purpose computer, or other programmable data processing equipment. It may be mounted so that its instructions, which are executed by the processor of a computer or other programmable data processing equipment, create means for performing the functions described in each block of the block diagram or in each step of the flowchart. These computer program instructions may also be stored in a computer-usable or computer-readable memory, which may direct a computer or other programmable data processing equipment to implement a function in a particular manner, the computer-usable or computer-readable memory. It is also possible for the instructions stored in the block diagram to produce an article of manufacture containing instruction means for performing a function described in each block of the block diagram or each step of the flowchart.

And, since the computer program instructions may be mounted on a computer or other programmable data processing equipment, a series of operating steps are performed on the computer or other programmable data processing equipment to create a computer-executed process to create a computer or other program It is also possible that instructions for performing the possible data processing equipment provide steps for carrying out the functions described in each block of the block diagram and each step of the flowchart.

Additionally, each block or step may represent a module, segment, or portion of code comprising one or more executable instructions for executing specified logical functions, and in some alternative embodiments the blocks or steps referred to in the block or step. It should be noted that it is also possible for functions to occur out of sequence. For example, it is possible that two blocks or steps shown one after another may be performed substantially simultaneously, and also the blocks or steps may be performed in the reverse order of the corresponding functions, if necessary.

1 is an overall block diagram illustrating a speech manipulation apparatus for selecting a response of a next utterance in a search-based conversation system according to an embodiment of the present invention.

Referring to FIG. 1 , the utterance manipulation device for selecting a response of the next utterance in the search-based dialog system according to an embodiment of the present invention is largely a utterance insertion module 100 , a utterance deletion module 200 , and a utterance search module. 300 , and a language learning model 400 . Meanwhile, since the components shown in FIG. 1 are not essential, the speech manipulation apparatus for selecting a response of the next speech in the search-based conversation system according to an embodiment of the present invention has more components or fewer components. It may have elements.

Hereinafter, components of a speech manipulation apparatus for selecting a response of a next speech in a search-based dialogue system according to an embodiment of the present invention will be described in detail as follows.

The utterance insertion module 100 extracts an arbitrary utterance from an original dialogue input from the outside, and performs a function of first reconstructing the corresponding original dialogue in a form separated from the corresponding original dialogue by extracting each extracted utterance.

In addition, the utterance insertion module 100 performs a function of adding a specific insertion token before each utterance and after the last utterance of the first reconstructed original dialogue to learn the consistency of the corresponding original dialogue.

The utterance deletion module 200 randomly extracts utterances having different meanings from any other dialogue other than the corresponding original dialogue, and sets the extracted utterances with different meanings before and after any one of the original dialogues. / or performs a function of second reconstructing the corresponding original dialogue in the interleaved form.

In addition, the utterance deletion module 200 performs a function of adding a specific deletion token in front of each utterance of the second reconstructed original conversation so that consistency between utterances in the corresponding original conversation can be learned.

The utterance search module 300 extracts any last utterance from the corresponding original conversation, and then performs a function of reconstructing the extracted last arbitrary utterance in a form separate from the corresponding original conversation.

In addition, the utterance search module 300 performs a function of adding a specific search token in front of each utterance of the third reconstructed original conversation so that order information between utterances in the corresponding original conversation can be learned.

In addition, the language learning model 400 includes a specific insertion token added from the utterance insertion module 100 , a specific deletion token added from the utterance deletion module 200 , and a specific search added from the utterance search module 300 . By applying the token as an output feature, the function of learning whether the location of the specific deletion token, search token, and search token is actually inserted, deleted, or searched based on a preset unsupervised learning method carry out

Hereinafter, a utterance manipulation method for selecting a response of a next utterance in a search-based conversation system according to an embodiment of the present invention will be described in detail.

2 and 3 are overall flow charts for explaining a utterance manipulation method for selecting a response of a next utterance in a search-based conversation system according to an embodiment of the present invention, and FIG. 4 is a search according to an embodiment of the present invention. It is a structural diagram for explaining an utterance manipulation method for selecting a response of a next utterance in a conversation-based conversation system.

1 to 4 , the utterance manipulation method for selecting a response of the next utterance in a search-based dialog system according to an embodiment of the present invention includes, first, an original inputted from the outside through the utterance insertion module 100 . After extracting one utterance from the dialogue, each extracted utterance is configured in a form separated from the corresponding original dialogue (S100).

Thereafter, a specific insertion token is added before and after the last utterance of each utterance of the original conversation configured in step S100 so that the consistency of the corresponding original conversation can be learned through the utterance insertion module 100 ( S200 ).

At this time, the utterance manipulation process (X _INS ) through the utterance insertion module 100 can be performed by the following Equation 1, and it is a process of learning where an arbitrarily extracted utterance should be inserted in the original conversation. make it possible

(Equation 1)

는 삽입의 대상이 되는 발화이고, 각 삽입토큰([INS])의 위치는 삽입이 될 수 있는 위치이다. [CLS]는 분류를 위한 단어 단위 토큰이고, [SEP]는 문장 단위 구분을 위한 토큰이다.Here, a specific insertion token ([INS]) is added before each utterance, and also added after the last utterance.

is an utterance to be inserted, and the location of each insertion token ([INS]) is a location that can be inserted. [CLS] is a token for word unit classification, and [SEP] is a token for sentence unit classification.

Then, after randomly extracting an utterance having a different meaning from any other conversation other than the corresponding original conversation through the utterance deletion module 200, the extracted utterance with a different meaning is used before any one of the original conversations. and/or configured to be inserted later (S300).

Next, a specific deletion token is added in front of each utterance of the original conversation configured in step S300 so that consistency between utterances in the corresponding original conversation can be learned through the utterance deletion module 200 ( S400 ).

At this time, the utterance manipulation process (X _DEL ) through the utterance deletion module 200 can be performed by Equation 2 below, and learning where an utterance having a different meaning in another conversation is located in the original conversation and whether to be deleted The process enables learning of coherence between utterances within a conversation.

(Equation 2)

은 임의의 대화에서 추출한 발화이며 원본 대화 내에 임의로 삽입된 형태이다. 특정의 삽입토큰([DEL])이 각 발화 앞에 추가되며, 각 특정의 삽입토큰([DEL])의 위치는 삭제될 수 있는 위치이다.here,

is an utterance extracted from a random dialogue and is arbitrarily inserted into the original dialogue. A specific insertion token ([DEL]) is added before each utterance, and the position of each specific insertion token ([DEL]) is a position that can be deleted.

Thereafter, after extracting any last utterance from the corresponding original conversation through the utterance search module 300, the extracted last utterance is configured in a form separated from the corresponding original conversation (S500).

Then, a specific search token is added in front of each utterance of the original conversation configured in step S500 so that order information between utterances in the corresponding original conversation can be learned through the utterance search module 300 ( S600 ).

At this time, the utterance manipulation process (X _SRCH ) through the utterance search module 300 may be performed by Equation 3 below, and after arbitrarily shuffling the original dialogue, it is possible to find where the dialogue just before the separated last utterance is located. As a process, order information between utterances can be learned, and it can be applied as a process that can assist in predicting the next response.

(Equation 3)

는 추출된 마지막 발화이며,

발화들은 순서가 섞인 발화들로

가

의 바로 이전 발화가 된다. 그리고, 특정의 검색토큰([SRCH])은 각 발화 앞에 추가되며, 특정의 검색토큰([SRCH])의 위치는 바로 이전의 발화가 될 수 있는 위치이다.here,

is the last utterance extracted,

The utterances are utterances in a mixed order.

go

is the immediately preceding utterance of In addition, a specific search token ([SRCH]) is added before each utterance, and the location of the specific search token ([SRCH]) is a position that can be the previous utterance.

Finally, through the language learning model 400, the specific insertion token added in step S200, the specific deletion token added in step S400, and the specific search token added in step S600 are applied as output features. , learns whether the positions of the specific deletion token, the search token, and the search token are actually inserted, deleted, and searched positions based on a preset unsupervised learning method (S700).

The utterance manipulation method for selecting a response of the next utterance in the search-based dialog system according to an embodiment of the present invention described above has a multi-task learning structure in which learning is performed together with a utterance prediction method using a pre-learned language model. can

In this case, by adding a fully connected layer to the pre-learned language model, the correspondence between the conversation content and the next utterance can be converted into a score as shown in Equation 4 below.

(Equation 4)

Here, X _TASK is output of specific tokens (ie, insertion token ([INS]), deletion token ([DEL]), and search token ([SRCH]) proposed in the above-mentioned speech manipulation processes through the language model. Represented as a feature, it is a feature that is used to learn whether each location is actually a location to be inserted, deleted, and retrieved.

In the utterance manipulation method for selecting a response of the next utterance in the search-based conversation system according to the embodiment of the present invention described above, a special token suitable for each utterance manipulation process (ie, insertion, deletion, and search) ) (ie, insertion token, deletion token, and search token) are newly added, and these specific tokens can help to maintain semantic information of expressions of utterances entering the input of the language model.

Meanwhile, the utterance manipulation method for selecting a response of the next utterance in the search-based conversation system according to an embodiment of the present invention may also be implemented as computer-readable codes 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.

For example, computer-readable recording media include ROM, RAM, CD-ROM, magnetic tape, hard disk, floppy disk, removable storage device, and non-volatile memory (Flash Memory). , and optical data storage devices.

In addition, the computer-readable recording medium may be distributed in computer systems connected through a computer communication network, and stored and executed as readable codes in a distributed manner.

Although a preferred embodiment of an apparatus and method for utterance manipulation for selecting a response of a next utterance in the search-based conversation system according to the present invention has been described above, the present invention is not limited thereto, and the claims and detailed description of the invention It is possible to carry out various modifications within the scope of the description and the accompanying drawings, and this also belongs to the present invention.

100: speech insertion module,
200: utterance deletion module,
300: utterance search module,
400: language learning model

Claims (3)

A random utterance is extracted from the original dialog input from the outside, and the extracted utterance is first reconstructed in a form separated from the original dialog, and the randomly extracted utterance should be inserted in the original dialog. an utterance insertion module for adding a specific insertion token before and after each utterance of the first reconstructed original conversation so as to learn the consistency of the corresponding original dialogue as an utterance manipulation process for learning whether to do so;
An utterance with a different meaning is randomly extracted from any other dialogue other than the corresponding original dialogue, and the corresponding original dialogue is produced in a form in which the extracted utterance with a different meaning is inserted before/after any one of the corresponding original dialogues. 2 It is a speech manipulation process that reconstructs and learns where utterances with different meanings in the corresponding random other dialogue are located in the corresponding original dialogue. a utterance deletion module for adding a specific deletion token in front of each utterance;
After extracting a last utterance from the original dialogue, the extracted last utterance is third reconstructed in a form separated from the original dialogue, and after randomly mixing the original dialogue, where is the dialogue immediately before the separated last utterance? a utterance search module for adding a specific search token in front of each utterance of the third reconstructed original dialog so that order information between utterances in the corresponding original dialog can be learned as an utterance manipulation process for finding the location; and
Pre-set unsupervised learning (Unsupervised) by applying the specific insertion token added from the utterance insertion module, the specific deletion token added from the utterance deletion module, and the specific search token added from the utterance search module as output features Learning) method to select the response of the next utterance in a search-based dialog system including a language learning model that learns whether the specific deletion token, search token, and the location of the search token are the actual insertion, deletion, and search locations fire control device for
A utterance manipulation method for selecting a response of a next utterance in a retrieval-based dialog system, comprising:
(a) extracting an arbitrary utterance from an original dialog input from the outside through a utterance insertion module, and then configuring each extracted utterance in a form separate from the corresponding original dialog;
(b) each utterance of the original dialogue constructed in step (a) so as to learn the consistency of the corresponding original dialogue as a speech manipulation process for learning where the randomly extracted speech should be inserted in the corresponding original dialogue through the speech insertion module adding a specific insertion token before and after the last utterance of ;
(c) After randomly extracting utterances with different meanings from other conversations other than the corresponding original conversations through the utterance deletion module, the extracted utterances with different meanings are inserted before/after any one of the original conversations. Constructing in a put form;
(d) through the utterance deletion module, an utterance manipulation process of learning where utterances with different meanings in the corresponding arbitrary other dialogues are located in the corresponding original dialogue, the step ( adding a specific deletion token in front of each utterance of the original conversation constructed in c);
(e) extracting a last utterance from the corresponding original conversation through the utterance search module, and then configuring the extracted last utterance in a form separated from the original conversation;
(f) arbitrarily shuffle the original dialogue through the speech search module and then find the location of the dialogue just before the separated last utterance. adding a specific search token in front of each utterance of the original conversation constructed in (e); and
(g) output the specific insertion token added in step (b), the specific deletion token added in step (d), and the specific search token added in step (f) through the language learning model Applied to, based on a preset unsupervised learning method, the specific deletion token, the search token, and the search token location are actually inserted, deleted, and searched. A utterance manipulation method for selecting the response of the next utterance in a search-based dialog system.
A computer-readable recording medium in which a program capable of executing the method of claim 2 by a computer is recorded.

Images