KR102464121B1 - Apparatus for providing query answering based on relation between query and response and method there of - Google Patents

Abstract

The present disclosure relates to a method of providing an answer to a user's query and an electronic device for performing the same. According to an embodiment, a method for an electronic device to provide an answer to a user's query includes: acquiring a query from the user; When the query is input, the first neural network model determines the importance of each token of the query, and outputs the similarity of the query and all answers to the query based on the determined importance of each token of the query. acquiring importance for each token; and by inputting the importance of each token of the query and the obtained query to a second neural network model that outputs an answer to the query when the importance of each token of the query and the obtained query is input, the response to the query is generated to do; may include.

Description

{APPARATUS FOR PROVIDING QUERY ANSWERING BASED ON RELATION BETWEEN QUERY AND RESPONSE AND METHOD THERE OF}

The present disclosure relates to a method and apparatus for providing an answer to a query. More particularly, it relates to a method and apparatus for providing an answer to a query using a neural network-based model.

As the amount of data existing on the Internet increases exponentially, the development of data processing technology to provide information desired by users is being actively studied today. Accordingly, a technology for providing a response related to a user's query is also being actively developed.

Recently, techniques for providing an answer to a query using a sequence-to-sequence model are being studied. The answer generator using the sequence-to-sequence model has a high limitation in generating a general answer that can be answered to any query. This is because the conventional sequence-to-sequence model has a limitation in that information on the entire contents of answers to queries is not reflected when decoding in token units.

Therefore, technology development is required to effectively provide answers to questions by reflecting information on the contents of the entire answer.

Korean Patent Publication No. 2020-0023620

According to an embodiment, a method for providing an answer to a query and an electronic device for performing the same may be provided.

In more detail, a method of providing an answer to a query based on the importance of each token of the query and an electronic device for performing the same may be provided.

As a technical means for achieving the above-described technical problem, according to an embodiment, a method for providing an answer to a user's query by an electronic device includes: acquiring a query from the user; When the query is input, the first neural network model determines the importance of each token of the query, and outputs the similarity of the query and all answers to the query based on the determined importance of each token of the query. acquiring importance for each token; and by inputting the importance of each token of the query and the obtained query to a second neural network model that outputs an answer to the query when the importance of each token of the query and the obtained query is input, the response to the query is generated to do; may include.

According to another embodiment as a technical means for achieving the above-described technical problem, an electronic device for providing an answer to a user's query, comprising: a network interface; a memory storing one or more instructions; and at least one processor executing the one or more instructions. wherein the at least one processor obtains a query from the user, and when the query is input, determines the importance of each token of the query, and based on the determined importance of each token of the query, the query and the A second neural network that obtains the importance of each token of the query from a first neural network model that outputs the similarity of all answers to a query, and outputs an answer to the query when the importance of each token of the query and the obtained query is input An electronic device may be provided that generates a response to the query by inputting the query and the importance of each token of the obtained query to a model.

According to another embodiment as a technical means for achieving the above technical problem, there is provided a method for providing an answer to a user's query by an electronic device, the method comprising: acquiring a query from the user; When the query is input, the first neural network model determines the importance of each token of the query, and outputs the similarity of the query and all answers to the query based on the determined importance of each token of the query. acquiring importance for each token; and by inputting the importance of each token of the query and the obtained query to a second neural network model that outputs an answer to the query when the importance of each token of the query and the obtained query is input, the response to the query is generated to do; A computer-readable recording medium in which a program for performing the method is stored, including a computer-readable recording medium may be provided.

According to an embodiment, it is possible to effectively provide an answer to a query.

According to an embodiment, there may be provided a method of providing an answer to a query that reflects information on the entire content of the answer to the query.

1 is a diagram schematically illustrating a process in which an electronic device provides an answer to a user's query, according to an exemplary embodiment.
2 is a flowchart of a method for providing an answer to a user's query by an electronic device according to an exemplary embodiment.
3 is a diagram for describing a process in which an electronic device provides an answer to a user's query by using a plurality of neural network models, according to another embodiment.
4 is a block diagram of an electronic device according to an embodiment.

Terms used in this specification will be briefly described, and the present disclosure will be described in detail.

The terms used in the present disclosure have been selected as currently widely used general terms as possible while considering the functions in the present disclosure, 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 terms used in the present disclosure should be defined based on the meaning of the term and the contents of the present disclosure, rather than the simple name of the 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, with reference to the accompanying drawings, embodiments of the present disclosure will be described in detail so that those of ordinary skill in the art to which the present disclosure pertains can easily implement them. However, the present disclosure may be implemented in several different forms and is not limited to the embodiments described herein. And in order to clearly explain the present disclosure in the drawings, parts irrelevant to the description are omitted, and similar reference numerals are attached to similar parts throughout the specification.

1 is a diagram schematically illustrating a process in which an electronic device provides an answer to a user's query, according to an exemplary embodiment.

According to an embodiment, the electronic device 1000 may obtain the query 102 and provide the answer 104 as a response to the obtained query 102 . According to an embodiment, the electronic device 1000 may provide an answer to the query 104 by interworking with the server 2000 or another electronic device 1002 to generate an answer to the query. According to an embodiment, the electronic device 1000 may provide an answer 104 to the query 102 using at least one neural network model. According to an embodiment, the electronic device 1000 obtains some of the output values of the first neural network model from the electronic device 1002, and uses the obtained output values of the first neural network model to answer ( 104) can be provided.

However, according to another embodiment, the electronic device 1000 obtains the first neural network model 162 or some of the output values of the first neural network model 162 from the electronic device 1002, and the obtained first neural network model 162 or the output value of the first neural network model 162 and the second neural network model 134 may be used to provide an answer 104 to the query 102 . According to another embodiment, the electronic device 1000 obtains at least one of the first neural network model 162 or the output value of the first neural network model 162 from the server 2000, and obtains the The answer 104 to the query 102 may be provided by using the information on the first neural network model.

According to an embodiment, the neural network model used by the electronic device 1000 may be a neural network model having a sequence-to-sequence model structure. According to an embodiment, the neural network model used by the electronic device 1000 may include a deep neural network (DNN). For example, a neural network model used by the electronic device 1000 to provide an answer to a query may include a Convolutional Neural Network (CNN), a Deep Neural Network (DNN), a Recurrent Neural Network (RNN), a Restricted Boltzmann Machine (RBM), DBN (Deep Belief Network), BRDNN (Bidirectional Recurrent Deep Neural Network), or deep Q-networks, etc., but are also not limited thereto.

In general, the conventional sequence-to-sequence model has a limitation in providing a general answer that can be answered to any question (answer that can be a common answer to a plurality of questions). This is because, in the case of a general answer generator, information about the entire correct answer is not reflected at every moment through token unit answer token generation during decoding. Technology development is required to effectively generate answers to questions by reflecting the contents of

The electronic device 1000 according to the present disclosure obtains the importance of each token for the query 102 from the electronic device 1002 and provides an answer to the query by using the obtained importance value for each token of the query, thereby responding to the query. An answer 104 can be effectively provided. For example, the electronic device 1002 may include a network interface 140 , a processor 150 , and a memory 160 . The processor 150 in the electronic device 1002 may extract a relationship between the query 102 and the expected answer by executing one or more instructions in the memory 160 .

In more detail, the processor 150 may determine the importance of each query token by using the first neural network model 162 stored in the memory 160 . According to an embodiment, the first neural network model may be a neural network model trained to output a similarity of a query and all answers to the query based on the importance of each query token. The electronic device 1002 may train the first neural network model in advance by using a database including predetermined conversation pair data. As a query is input to the learned first neural network model, the electronic device 1000 may acquire the importance of each token of the query from the first neural network model. The processor 150 in the electronic device 1002 may transmit information on the importance of each token of the query determined from the first neural network model to the electronic device 1000 through the network interface 140 .

According to another embodiment, the electronic device 1002 may transmit all information on the first neural network model 162 to the electronic device 1000 . In this case, the electronic device 1000 obtains the query 102 , inputs the obtained query 102 to the first neural network model 162 , and obtains the importance value for each token of the query from the first neural network model 162 . It is possible to effectively provide an answer to a query by acquiring and using the importance value for each token of the acquired query.

According to an embodiment, the electronic device 1000 may include a configuration corresponding to the electronic device 1002 . For example, the electronic device 1000 may include a network interface 110 , a processor 120 , and a memory 130 . The processor 120 performs a method for providing an answer to a query by executing one or more instructions in the memory 130 . For example, by controlling the network interface 110 , the processor 120 provides information on the first neural network model 162 from the electronic device 1002 or the importance of each token of a query from the first neural network model 162 . information can be obtained. The processor 120 may provide an answer to the query using the first neural network model 162 obtained from the electronic device 1002 or the importance value for each query token obtainable from the first neural network model 162 .

In more detail, the electronic device 1000 obtains information on the importance of each token of the query from the electronic device 1000 , and inputs the query 102 and the importance value of each token of the query into the second neural network model 134 . By doing so, it is possible to generate an answer 104 to the query. However, according to another embodiment, the electronic device 1000 does not acquire the importance value for each query token from the electronic device 1002 , but acquires information about the first neural network model learned by the electronic device 1000 . Thereafter, by using the first neural network model stored in the electronic device 1000 , the importance value for each token for the query may be directly obtained.

According to another embodiment, the electronic device 1000 obtains only information on the query token importance determiner 132 among the configurations of the first neural network model 162 , thereby Only the model may be stored in the electronic device 1000 . The electronic device 1000 may effectively provide the answer 104 to the query using the query token importance determiner 132 and the second neural network model 134 previously stored in the memory 130 .

According to an embodiment, the electronic device 1000 may be connected to the server 2000 through a predetermined network. According to an embodiment, the electronic device 1000 is connected to the electronic device 1002 or the server 2000 through a predetermined network, thereby providing a neural network model, query token information, and query information necessary to provide an answer to a user query. It is also possible to transmit/receive importance information, etc. for each token.

That is, the electronic device 1002 trains the first neural network model to extract the relationship between the query token and the entire answer, and uses the first neural network model to determine the relationship between the query and the overall answer (eg, the importance of each query token). can be extracted. Although the electronic device 1002 may transmit the importance value for each query token to the electronic device 1000, information on the first neural network model itself or information on some models such as the query token importance determiner in the first neural network model is directly transmitted. It may also be transmitted to the electronic device 1000 . The electronic device 1000 determines by itself by directly operating the importance value for each query token obtained from the electronic device 1002 or the query token importance determining unit 134 or the first neural network model 162 obtained from the electronic device 1002 . An answer to a query can be provided by using the importance value for each query token.

2 is a flowchart of a method for providing an answer to a user's query by an electronic device according to an exemplary embodiment.

In S210 , the electronic device 1000 may obtain a query from the user. According to an embodiment, the electronic device 1000 may acquire a query in a text form (eg, a word or a sentence) based on a user input. However, according to another embodiment, of course, the electronic device 1000 may acquire the query in the form of an image or audio.

In S220 , when a query is input, the electronic device 1000 determines the importance of each token of the query, and outputs the similarity of the query and all answers to the query based on the determined importance of each token of the query. Importance of each query token can be obtained from For example, the electronic device 1000 may determine the importance of each token of the query by using the first neural network model. A token may mean a language element that cannot be further grammatically divided into a query. The electronic device 1000 or the electronic device 1002 may acquire a token for each query by tokenizing the query using the first neural network model.

According to an embodiment, the first neural network model obtains the query from a database in which conversation pairs including a predetermined query and an answer to the query are stored, and generates an expression vector for the query by encoding the query. and a full answer expression vector encoder for obtaining an answer to the query and encoding the entire obtained answer to generate an expression vector for the entire answer. That is, the first neural network model according to the present disclosure may include two or more encoders.

According to an embodiment, the first neural network model may be trained so that a query expression vector that is an expression vector for a query and an entire answer expression vector that is an expression vector for all answers are similar to each other. According to an embodiment, if the query expression vector and the full answer expression vector are similar (eg, the similarity is close to 1), the query and answer corresponding to the full answer expression vector corresponding to the query expression vector are linguistically meaningful query and It can represent answer pairs (eg, meaningful real conversation pairs that real people have).

However, if the query expression vector and the full answer expression vector are not similar, the answer corresponding to the query expression vector and the full answer expression vector corresponding to the query expression vector is a linguistically meaningless query and answer pair (e.g., conversations between real people It can represent a linguistically meaningless conversational pair).

Also, according to an embodiment, the query expression vector encoder in the first neural network model generates at least one token for a query by tokenizing the query, and generates a query token expression vector for each generated token expression vector. It may include a generating unit and a query token importance determining unit that determines the importance of the generated expression vector for each token. For example, the query token expression vector generator may generate the expression vector by tokenizing the query and encoding tokens of the query. According to an embodiment, the query token importance determining unit may determine the importance of the expression vector for each token.

In more detail, the query expression vector encoder may generate the expression vector for the query by weight summing the expression vectors for each token based on the determined query token importance. The first neural network model may be learned by modifying and updating the weights of the layers in the neural network model and the connection strength between the layers so that the query expression vector generated according to the above-described method is similar to the entire answer expression vector. When learning of the first neural network model is completed so that the query expression vector and the entire answer expression vector are similar to each other, the electronic device 1000 makes a query based on the importance value for each query token output from the query token importance determiner in the first neural network model. It can be used to generate answers to

When the query expression vector encoder and the full answer expression vector encoder in the above-described first neural network model are pre-trained so that the expression vector for the query and the expression vector for the entire answer are similar to each other, the query token importance determining unit, It is possible to determine the importance of each token that sufficiently reflects the contents of the entire answer. According to an embodiment, the importance determined by the query token importance determining unit in the first neural network model may mean a degree of similarity to all answers for each query token. The above-described learning process of the first neural network model may be performed by the electronic device 1000 . However, according to another embodiment, it goes without saying that the learning of the first neural network model may also be performed by the server 2000 connected to the electronic device 1000 or another electronic device 1002 connected to the electronic device 1000 .

In S230, the electronic device 1000 inputs the importance of each token of the query and the obtained query to the second neural network model that outputs an answer to the query when the query and the obtained query are of importance for each token are input, A response to the above query may be generated. According to an embodiment, the second neural network model is a neural network model having a sequence-to-sequence model structure and may include an encoder and a decoder. The second neural network model used by the electronic device 1000 includes the importance value for each query token obtained from a partial neural network structure (eg, the query token importance determiner) in the first neural network model and, when a query is input, a query token for the entire answer. You can create an answer that reflects the importance of each.

According to an embodiment, when a query is input from the user, the second neural network model is an encoder that encodes the query, an output value of the encoder, and an encoder determined by reflecting the importance of each token of the query in an output value of the encoder and a decoder to obtain an output transform value, and to decode the obtained output value of the encoder and the encoder output transform value. According to an embodiment, the decoder in the second neural network model may acquire the encoder output value and the encoder output transform value sequentially or together.

For example, the encoder may generate an encoder output value by encoding the query when the query is input. The first neural network model may determine the encoder output transformation value multiplied by the importance reflecting the relation with the overall answer by scalar-multiplying the importance value of each query token to the output value of the encoder for each token. The first neural network model may decode the above-described encoder output value and encoder output transform value by inputting the decoder, and generate an answer to the query based on the decoding result.

Although not shown in FIG. 2 , the electronic device 1000 may provide the generated response as an answer to the user's query. According to an embodiment, the electronic device 1000 may output the generated answer in the form of an image, video, or audio by using a speaker or a display included in the electronic device.

3 is a diagram for describing a process in which an electronic device provides an answer to a user's query by using a plurality of neural network models, according to another embodiment.

A process in which the electronic device 1000 provides an answer to a query using a plurality of neural network models will be described with reference to FIG. 3 . According to an embodiment, the electronic device 1000 may use the first neural network model 310 and the second neural network model 330 to provide an answer to a query based on the importance of each query token. According to an embodiment, the first neural network model 310 is a relationship extractor that extracts a relationship between a query and an answer, and the second neural network model answers the query based on the relationship between the query and the answer extracted from the relationship extractor. may correspond to an answer generator that provides

According to an embodiment, the first neural network model 310 and the second neural network model 330 used by the electronic device 1000 may be stored in a memory of the electronic device. However, according to another embodiment, information about the query token importance determiner in the first neural network model 310 and the second neural network model may be stored in the memory of the electronic device. However, according to another embodiment, only information about the second neural network model 330 may be stored in the memory of the electronic device 1000 .

According to an embodiment, the first neural network model 310 includes a database 312 in which predetermined conversation pairs are stored, a query token expression vector generator 314 , a query token importance calculator 316 , and an answer full expression vector generator 318 . ) and a similarity checker 320 . According to an embodiment, the database 312 is not included in the first neural network model 310, but may be provided externally for providing conversation pair learning data to the first neural network model.

According to an embodiment, the query token expression vector generator 314 may correspond to the above-described query token expression vector generator as an encoder, and the query token importance calculator may correspond to the above-described query token importance determiner. Also, according to an embodiment, the full answer expression vector generator 318 may correspond to the above-described full answer expression vector encoder. The first neural network model 310 used by the electronic device 1000 may be learned by the electronic device 1000 or the electronic device 1002 connected to the electronic device. According to an embodiment, the query token expression vector generator 314 in the first neural network model obtains query learning data from the conversation pair database 312 , and the answer full expression vector generator 318 generates the conversation pair database 312 . ) to obtain the learning answer data.

The query token expression vector generator 314 generates a query expression vector by encoding the query learning data, and the answer full expression vector generator 318 generates an expression vector for the entire answer by encoding the learning answer data. The query token importance calculator 316 determines the importance of the expression vector for each query token, and reflects the determined importance as a weight in the expression vector for each query token. The query token expression vector generator 314 may generate an expression vector for the entire query by reflecting the importance of each query token as a weight in the query token expression vector.

The similarity checker 320 may compare the similarity between the query expression vector and the entire answer expression vector. According to an embodiment, when the similarity value of the query expression vector and the full answer expression vector is 1, the similarity check unit 320 determines that the query expression vector and the full answer expression vector are similar, and the currently input query and answer are actually You can decide to form a conversation pair. However, according to another embodiment, when the similarity check unit 320 identifies the similarity value of the query expression vector and the entire answer expression vector as 0, the query and answer currently input to the first neural network model constitute a false conversation pair. may be decided.

As the first neural network model 310 is trained so that the query expression vector and the full answer expression vector become similar, when calculating the importance of each token of each query, an expression vector similar to the actual correct answer can be finally generated, A query expression vector and a similarity checker may be trained. When the learning process of the first neural network model 310 is completed, even if no answer is inputted thereafter and only information about the query is inputted, the importance of each token for each query in which the answer is considered may be output.

The importance of each query token determined in the first neural network model 310 may be input to the second neural network model 330 . For example, the second neural network model 330 may acquire the importance 334 for each query token output from the first neural network model 310 and the user input sentence 332 corresponding to the user query. The second neural network model 330 inputs the user input sentence 332 and the query token importance 334 to the answer generator 336 .

According to an embodiment, the answer generator 336 may correspond to the first neural network model. According to an embodiment, the answer generator 336 may include an encoder and a decoder. For example, the encoder in answer generator 336 may determine the first encoded output value by encoding the query. Also, the answer generator 336 may determine the encoder output conversion value (eg, the second encoded output value) by scalar-multiplying the first encoded output value by the importance of each token of the query for each token. The answer generator 336 may use the first encoded output value as a first input of the decoder and the second encoded output value as a second input of the decoder.

A decoder in the answer generator 336 may generate an answer sentence 338 to the query by decoding the first encoded output value and the second encoded output value. According to an embodiment, the decoder in the answer generator 336 may sequentially obtain the first encoded output value and the second encoded output value. However, according to another embodiment, the decoder in the answer generator 336 may acquire the first encoded output value and the second encoded output value together.

4 is a block diagram of an electronic device according to an embodiment.

According to an embodiment, the electronic device 1000 may include a processor 1300 , a network interface 1500 , and a memory 1700 . However, it is not limited to the above-described configuration, and the electronic device 1000 may further include a speaker, a display, and the like for outputting a response or an answer. According to an embodiment, the processor 1300 performs a process of providing an answer to a query by executing one or more instructions stored in the memory 1700 .

According to an embodiment, the network interface 1500 may include one or more components that allow the electronic device 1000 to communicate with another device (not shown) and a server (not shown). Another device (not shown) is a device that provides a response to a query, such as the electronic device 1000 , and may be a device for training the first neural network model. The network interface 1500 may obtain information on the first neural network model, information on the query token importance determiner of the first neural network model, and importance values for each query token from the other electronic device 1002 .

The memory 1700 may store a program for processing and control of the processor 1300 , and may store response or correct answer information as data output from a query inputted to the electronic device 1000 and from the electronic device 1000 . . Also, the memory 1700 may further store the above-described first neural network model, the second neural network model, information on the query token importance determiner in the first neural network model, information on the query token importance value, and the like. In addition, the memory 1700 may further store information about a database for learning each neural network model and training data to be stored in the database.

The memory 1700 may store layers in a plurality of neural network models and weight values related to connection strengths of the layers. Weight values stored in the memory 1700 may be modified and updated according to a preset period. That is, in the memory 1700, the query token importance determining unit newly added based on data obtained by the electronic device 1000 from other electronic devices 1002 (eg, the importance of each query token), and the updated second neural network model are stored in the memory 1700 . Information or information about the updated first neural network model may be added and updated.

The memory 1700 may include a flash memory type, a hard disk type, a multimedia card micro type, a card type memory (eg, SD or XD memory), and a RAM. (RAM, Random Access Memory) SRAM (Static Random Access Memory), ROM (Read-Only Memory), EEPROM (Electrically Erasable Programmable Read-Only Memory), PROM (Programmable Read-Only Memory), magnetic memory, magnetic disk , may include at least one type of storage medium among optical disks.

The processor 1300 obtains a query by executing one or more instructions stored in the memory, and determines an answer to the query using at least one of the first neural network model and the second neural network model. For example, the processor determines the importance of each token of the query, and based on the determined importance of each token of the query, the query from the first neural network model that outputs the similarity of all answers to the query and the query. By obtaining the importance of each token, and inputting the importance of each token of the query and the obtained query to a second neural network model that outputs an answer to the query when the importance of each token of the query and the obtained query is input, You can create responses to queries.

Also, the at least one processor 1300 may provide the generated response as an answer to the user's query. In addition, the processor 1300 controls overall operations of the network interface 1500 and the memory 1700 in order to perform the method of providing answers to the queries described above in FIGS. 1 to 3 .

According to an embodiment, the processor 1300 obtains the query from a database in which the first neural network model stores conversation pairs including a predetermined query and an answer to the query, and encodes the query to respond to the query. It can be controlled to generate an expression vector for Also, the processor 1300 may control the first neural network model to generate an expression vector for the entire answer by obtaining an answer to the query and encoding the entire obtained answer.

In addition, the processor 1300 causes the first neural network model to generate at least one token for the query by tokenizing the query, to generate the generated expression vector for each token, and the importance of the generated expression vector for each token. and controlling the first neural network model to generate an expression vector for the query by weight summing the expression vectors for each token based on the determined importance of the query token.

Also, the processor 1300 may train the query expression vector encoder and the full answer expression vector encoder so that the expression vector for the query and the expression vector for the entire answer are similar to each other. In addition, as the query expression vector encoder is trained so that the expression vector for the query and the expression vector for the entire answer are similar to each other, the processor 1300 determines that the query token importance determining unit in the query expression vector encoder is The first neural network model may be trained to determine the importance of each token of the query based on the contents of .

In addition, the processor 1300 controls the encoder in the second neural network model to encode the query when a query is input from the user, and determines by reflecting the importance of each token of the query in the output value of the encoder and the output value of the encoder A decoder in the second neural network model may be controlled to obtain an encoder output transform value that is a given value, and to decode the obtained encoder output value and the encoder output transform value.

The method in which the electronic device provides an answer to a user's query according to an embodiment may be implemented in the form of a program command that can be executed through various computer means and recorded in a computer-readable medium. The computer-readable medium may include program instructions, data files, data structures, etc. alone or in combination. The program instructions recorded on the medium may be specially designed and configured for the present invention, or may be known and available to those skilled in the art of computer software.

Also, there may be provided a computer program device including a recording medium storing a program for enabling the electronic device to perform a method of providing an answer to a user's query.

Examples of the computer-readable recording medium include magnetic media such as hard disks, floppy disks and magnetic tapes, optical media such as CD-ROMs and DVDs, and magnetic such as floppy disks. - includes magneto-optical media, and hardware devices specially configured to store and execute program instructions, such as ROM, RAM, flash memory, and the like. Examples of program instructions include not only machine language codes such as those generated by a compiler, but also high-level language codes that can be executed by a computer using an interpreter or the like.

Some embodiments may also be implemented in the form of a recording medium containing instructions executable by a computer, such as program modules executed by a computer. Computer-readable media can be any available media that can be accessed by a computer and includes both volatile and nonvolatile media, removable and non-removable media. In addition, computer-readable media may include both computer storage media and communication media. Computer storage media includes both volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data. Communication media typically includes computer readable instructions, data structures, program modules, or other data in a modulated data signal such as a carrier wave, or other transport mechanism, and includes any information delivery media. Also, some embodiments may be implemented as a computer program or computer program product comprising instructions executable by a computer, such as a computer program executed by a computer. Although the embodiments of the present invention have been described in detail above, the scope of the present invention is not limited thereto. belong to the scope of the right.

Claims (17)

A method for an electronic device to provide an answer to a user's query, the method comprising:
obtaining a query from the user;
When a query is input from the user, the importance of each token of the query obtained from the user is determined, and based on the determined importance of each token of the query obtained from the user, the query obtained from the user and obtained from the user obtaining a level of importance for each token of a query obtained from the user from a first neural network model that outputs a degree of similarity of all answers to a query; and
When the query obtained from the user and the importance of each token of the query obtained from the user are input, the query obtained from the user and the query obtained from the user are input to a second neural network model that outputs an answer to the query obtained from the user. generating a response to a query obtained from the user by inputting the level of importance for each token of ; including,
The second neural network model is
an encoder that encodes a query obtained from the user when a query is input from the user; and
Obtain an encoder output transform value determined by reflecting the importance of each token of the query obtained from the user in the output value of the encoder and the output value of the encoder, and decode the obtained output value of the encoder and the encoder output transform value decoder; A method comprising: The method of claim 1, wherein the method
providing the generated response as an answer to the user's query; A method further comprising: The method of claim 1, wherein the first neural network model is
In the query obtained from the database by obtaining a query from a database in which conversation pairs including a query obtained from a predetermined user and an answer to the query obtained from the user are stored, and encoding the query obtained from the database a query expression vector encoder that generates an expression vector for and
an answer full expression vector encoder that obtains an answer to the query obtained from the database and generates an expression vector for the entire answer by encoding the entire obtained answer; A method comprising: 4. The method of claim 3, wherein the query expression vector encoder comprises:
a query token expression vector generator generating at least one token for the query by tokenizing the query, and generating an expression vector for each generated token; and
a query token importance determining unit that determines the importance of the generated expression vector for each token; including,
and the query expression vector encoder generates the expression vector for the query by weight summing the token-specific expression vector based on the determined query token importance. 5. The method of claim 4,
The query expression vector encoder and the full answer expression vector encoder are trained in advance so that the expression vector for the query and the expression vector for the entire answer are similar to each other,
As the query expression vector encoder is trained so that the expression vector for the query and the expression vector for the entire answer are similar to each other, the query token importance determining unit in the query expression vector encoder is based on the contents of the entire answer. A method, characterized in that the importance of each token of the query is determined. delete The method of claim 1, wherein the decoder
A method, characterized in that the output value of the encoder and the encoder output transform value are acquired sequentially or together. The method of claim 1, wherein the second neural network model is
A method, characterized in that it is a neural network model of a sequence-to-sequence structure. An electronic device that provides an answer to a user's query, the electronic device comprising:
network interface;
a memory storing one or more instructions; and
at least one processor executing the one or more instructions; including,
the at least one processor
obtain a query from the user;
When a query is input from the user, the importance of each token of the query obtained from the user is determined, and based on the determined importance of each token of the query obtained from the user, the query obtained from the user and obtained from the user Obtaining the importance of each token of the query obtained from the user from the first neural network model that outputs the similarity of all answers to the query,
When the query obtained from the user and the importance of each token of the query obtained from the user are input, the query obtained from the user and the query obtained from the user are input to a second neural network model that outputs an answer to the query obtained from the user. By entering the level of importance for each token of
The second neural network model is
an encoder that encodes a query obtained from the user when a query is input from the user; and
Obtain an encoder output transform value determined by reflecting the importance of each token of the query obtained from the user in the output value of the encoder and the output value of the encoder, and decode the obtained output value of the encoder and the encoder output transform value decoder; An electronic device comprising a. 10. The method of claim 9, wherein the at least one processor is
An electronic device that provides the generated response as an answer to the user's query. The method of claim 9, wherein the first neural network model is
In the query obtained from the database by obtaining a query from a database in which conversation pairs including a query obtained from a predetermined user and an answer to the query obtained from the user are stored, and encoding the query obtained from the database a query expression vector encoder that generates an expression vector for and
an answer full expression vector encoder that obtains an answer to the query obtained from the database and generates an expression vector for the entire answer by encoding the entire obtained answer; The electronic device comprising a. 12. The method of claim 11, wherein the query expression vector encoder comprises:
a query token expression vector generator generating at least one token for the query by tokenizing the query, and generating an expression vector for each generated token; and
a query token importance determining unit that determines the importance of the generated expression vector for each token; including,
and the query expression vector encoder generates the expression vector for the query by weight summing the expression vector for each token based on the determined query token importance. 13. The method of claim 12,
The query expression vector encoder and the full answer expression vector encoder are trained in advance so that the expression vector for the query and the expression vector for the entire answer are similar to each other,
As the query expression vector encoder is trained so that the expression vector for the query and the expression vector for the entire answer are similar to each other, the query token importance determining unit in the query expression vector encoder is based on the contents of the entire answer. An electronic device, characterized in that the importance of each token of the query is determined. delete The method of claim 9, wherein the decoder comprises:
The electronic device, characterized in that the output value of the encoder and the encoder output conversion value are acquired sequentially or together. The method of claim 9, wherein the second neural network model is
An electronic device, characterized in that it is a neural network model of a sequence-to-sequence structure. obtaining a query from a user;
When the query is input, the first neural network model determines the importance of each token of the query, and outputs the similarity of the query and all answers to the query based on the determined importance of each token of the query. acquiring importance for each token; and
When the importance of each token of the query and the obtained query is input, the response to the query is generated by inputting the importance of each token of the query and the obtained query to a second neural network model that outputs an answer to the query step; including,
The second neural network model is
an encoder that encodes a query obtained from the user when a query is input from the user; and
Obtain an encoder output transform value determined by reflecting the importance of each token of the query obtained from the user in the output value of the encoder and the output value of the encoder, and decode the obtained output value of the encoder and the encoder output transform value decoder; A computer-readable recording medium storing a program for causing an electronic device to perform a method of providing an answer to a user's query, comprising:

Images