KR20200047272A - Indexing system and method using variational recurrent autoencoding - Google Patents

Abstract

An objective of the present invention is to provide an automatic indexing system using variational recurrent autoencoding and a method thereof which can easily find an answer in a database. According to embodiments of the present invention, the indexing system using variational recurrent autoencoding comprises: a first variational recurrent autoencoder to learn potential variable values through a training process of encoding and decoding a preset prescribed keyword set and encode a keyword set corresponding to a character input or a voice of a user based on the learning to calculate potential variable values; a second variational recurrent autoencoder to learn potential variable values through a training process of encoding and decoding a preset prescribed image and encode an image input of a user based on the learning to calculate potential variable values; and an indexing unit to index an input from a user based on a result of comparing the potential variable values learned by the first variational recurrent autoencoder and the second variational recurrent autoencoder and the potential variable values calculated by the first variational recurrent autoencoder or the second variational recurrent autoencoder. The potential variable values learned by the first variational recurrent autoencoder and the second variational recurrent autoencoder are shared.

Description

INDEXING SYSTEM AND METHOD USING VARIATIONAL RECURRENT AUTOENCODING}

The present invention relates to an automatic indexing system and method using a variable circulation autoencoding method.

Recently, intelligent information search systems and chatbot services that provide various services and information are increasing. The existing intelligent information retrieval engine uses a database in which answers to questions are stored, and calculates the relationship between the term frequency and the inverse document frequency to find the most similar question to the entered question and Use the method to find the corresponding answer.

In this regard, Korean Patent Publication No. -2018-0042763 includes a communication module that receives a user query in the form of a complete or incomplete sentence input through an access channel and transmits a response to the user query; A natural language processing engine that performs syntax analysis and morpheme analysis on user queries received through the communication module; A database including a language dictionary database, a financial knowledge database, and a conversation scenario database for providing financial services; A dialogue that interprets the dictionary meaning and contextual meaning of the user query based on the syntax and morpheme analyzed by the natural language processing engine and generates a response to the user query based on the financial service provision scenario stored in the dialogue scenario database Provided is a chat-type financial robot including a processing engine.

However, this is about finding the answer (response) directly, not indexing the question to find the answer.

In the present invention, when the input question is a natural language or an image, a variable recurrent autoencoding method that facilitates finding an answer in a database by indexing (classifying) using the value of a latent variable of the variable recurrent encoder. It is intended to provide an automatic indexing system and method.

The index system of the variable circulating auto-encoding method according to an embodiment of the present invention learns the value of a potential variable through a training process of encoding and decoding a predetermined set of keywords, and based on this, inputs the user's voice or text. A first variable cyclic autoencoder that calculates a value of a latent variable by encoding a corresponding keyword set; A second variable cyclic autoencoder that learns the value of the latent variable through a training process of encoding and decoding a predetermined image, and calculates the latent variable value by encoding a user's image input based on the training process; The result of comparing the value of the potential variable learned by the first variable circulating autoencoder and the second variable circulating autoencoder and the value of the potential variable calculated by the first variable cyclic autoencoder or the second variable cyclic autoencoder And an index unit that indexes input from a user based on the values of the potential variables learned from the first variable cyclic autoencoder and the second variable cyclic autoencoder.

The index system of the variable circulating auto-encoding method may further include a keyword generator configured to generate a keyword set corresponding to the user's text or voice input by analyzing a morpheme of the user's text or voice input.

The first incremental cyclic autoencoder may include: a first encoding unit encoding the preset predetermined keyword set or a keyword set corresponding to the user's voice or character input; And a first decoding unit decoding the potential variable generated by encoding the predetermined keyword set by the first encoding unit.

The first encoding unit includes the same number of Long Short Term Memory (LSTM) as the number of keywords constituting the input keyword set, and the keywords are respectively input to the LSTM in the input order, and the LSTM is a chain format. Can be connected to.

The second variable circulating autoencoder may include: a second encoding unit encoding the preset predetermined image or the user's image input; And a second decoding unit decoding a value of a latent variable generated by encoding the predetermined image by the second encoding unit, and the second encoding unit and the second decoding unit include a convolutional neural network (CNN). can do.

The update unit may further include an update unit that updates the index of the index unit by using the value of the latent variable calculated by the first variable cyclic autoencoder or the second variable cyclic autoencoder.

The predetermined keyword set may be related to taxation, and the predetermined predetermined image may be an image capable of distinguishing a document type related to taxation.

In the indexing method of the variable cyclic autoencoding method according to an embodiment of the present invention, a potential variable is learned through a training process of encoding and decoding a predetermined set of keywords using a first variable cyclic autoencoder, and in advance Learning a value of a latent variable shared with a latent variable learned in the first variable cyclic autoencoder through a training process of encoding and decoding a predetermined image using a second variable cyclic autoencoder; Calculating a value of a potential variable corresponding to a user input using the first variable cyclic autoencoder or the second variable cyclic autoencoder; And indexing input from a user based on a result of comparing the value of the learned latent variable with the calculated latent variable.

The index method of the cyclic cyclic auto-encoding method further includes generating a keyword set corresponding to the user's text or voice input by analyzing a morpheme of the user's text or voice input, and in the calculating step, the user When the input of is text or voice, the value of the potential variable corresponding to the generated keyword set may be calculated using the first variable cyclic autoencoder.

In the indexing method of the variable cyclic autoencoding method, updating the index used in the indexing step using the value of the latent variable calculated by the first variable cyclic autoencoder or the second variable cyclic autoencoder. It may further include.

The predetermined keyword set may be related to taxation, and the predetermined predetermined image may be an image capable of distinguishing a document type related to taxation.

A keyword generator configured to generate a set of keywords corresponding to the user's natural language input by analyzing a morpheme of the user's natural language input;

The index system of the variable cyclic autoencoding method according to an embodiment of the present invention learns the value of a potential variable through a training process of encoding and decoding a predetermined set of keywords, and based on this, generates the keyword generation unit A variable cyclic autoencoder that calculates the value of a latent variable by encoding a keyword set; An indexing unit that indexes input from a user based on a result of comparing a value of a potential variable learned by the variable cyclic autoencoder with a value of a potential variable calculated by the variable cyclic autoencoder; And an updating unit updating the index of the index unit by using the value of the potential variable calculated by the variable circulating autoencoder.

According to the exemplary embodiment of the present invention, since the inputted question is indexed using the incremental cycle auto-encoding method, indexing with high accuracy is possible.

1 is a view showing the configuration of the index system of the variable circulating auto-encoding method according to an embodiment of the present invention.
FIG. 2 is a view showing the operation of the morpheme analyzer of FIG. 1.
3 is a view for explaining the operation of the variable-circulation auto-encoder of FIG.
4 is a view showing an example of the structure of the LSTM of FIG. 3.
5 is a view showing the configuration of the index system of the variable circulating auto-encoding method according to another embodiment of the present invention.
FIG. 6 is a view for explaining the operation of some components of FIG. 5.
7 is a view showing the configuration of a specialized field response service system of a variable circulating auto-encoding method according to an embodiment of the present invention.
8 is a view showing a specialized field response service method using a variable rotation auto-encoding method using a chatbot according to an embodiment of the present invention.
9 is a diagram showing experimental results of a specialized field response service system and method of a variable circulating auto-encoding method according to an embodiment of the present invention.

Based on the principle that the inventor can appropriately define the concept of terms in order to explain his or her invention in the best way, terms or words used in the specification and claims conform to the technical spirit of the present invention. It should be interpreted as meaning and concept.

Throughout the specification, when a part “includes” a certain component, this means that other components may be further included rather than excluding other components unless specifically stated to the contrary. In addition, when one component is said to be "connected", "send", "send", "receive" or "forward" to another component, this is not only the case where it is directly connected, transmitted, sent, received or forwarded, It also includes the case of indirect connection, transmission, transmission, reception, or transmission via a component. In addition, terms such as “… unit”, “… group”, “module”, and “device” described in the specification mean a unit that processes at least one function or operation, which is hardware or software or a combination of hardware and software. Can be implemented as

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

First, Example 1 of the present invention will be described with reference to FIGS. 1 to 4.

1 is a diagram showing the configuration of an indexing system 1 of a variable recurrent autoencoding method according to an embodiment of the present invention.

Referring to FIG. 1, the index system 1 is a system that receives a natural language and generates an index value for searching a corresponding response in a database. The indexing system 1 includes a morpheme analyzer 100, a variance circulation autoencoder 200, an indexing unit 300, and an updating unit 400.

The morpheme analyzer 100 analyzes the morpheme of the natural language and generates a keyword set of the natural language. The natural language input to the morpheme analyzer 100 may be a completed sentence or an incomplete sentence composed of one or more words. The natural language may be, for example, Korean. The keyword set may be general nouns, proper nouns, verbs, etc., except for searches that do not have a special meaning in natural language input. The keyword set includes at least one keyword.

The variable circulating autoencoder 200 processes a set of keywords in a variable cyclic autoencoding method to generate latent variables.

The indexing unit 300 indexes the keyword set according to the value of the latent variable generated by the variable rotation auto-encoder 200.

The update unit 400 updates the index by using the value of the latent variable generated by the variable rotation auto-encoder 200.

2 is a view showing the operation of the morpheme analyzer 100 of FIG. 1. The morpheme analyzer 100 generates a set of keywords by tokenizing natural language.

Referring to FIG. 2, when the natural language of “this year's corporate tax” is input, it indicates that {this year, corporate tax} is generated as a keyword set. In this embodiment, the keyword set {this year, corporate tax} is composed of two keywords: "this year" and "corporate tax". For example, the morpheme analyzer 100 may be used in the lucene (Lucene) based search engine used in search engines.

Next, the operation of the variable circulation autoencoder 200 will be described with reference to FIG. 3.

The variable-circulation autoencoder is a kind of unsupervised learning, and is often used for dimensional reduction and generation models. The essence of the variable cyclic autoencoder is learning that the latent variable (Z) follows the normal distribution of the mean (μ) and variance (σ). Because the posterior p (z | x) is difficult to calculate (Intractable), q (z | x) and p (z | x) are approximated using variable inference. Using Kullback-Leibler Divergence, we can derive as follows to minimize the difference between q (z | x) and p (z | x).

는 항상 0보다 크거나 같다.

를 최대화하면

가 최대가 되도록 하는 하한(Evidence Lower Bound)이며 목적함수(Objective Function)가 된다.In the above equation

Is always greater than or equal to 0.

Maximizing

Evidence Lower Bound to be the maximum and it becomes the objective function.

FIG. 3 is a view for explaining the operation of the variable circulating autoencoder 200 of FIG. 1.

Referring to FIG. 3, the variable cyclic autoencoder 200 includes encoding long short term memory (encoding units) 210 and 220 into which one or more keywords constituting a keyword set are input, and encoding LSTMs 210 and 220. It may include a decoding LSTM (decoding unit; 230, 240, 250) is the potential variable (Z) generated in the decoding.

First, the encoding LSTMs 210 and 220 and the decoding LSTMs 230, 240 and 250 of the variable cyclic autoencoder 200 may perform training operations. The encoding LSTMs 210 and 220 calculate potential variables by encoding each of a plurality of preset keyword sets, and perform training in a manner of decoding the calculated potential variables. Accordingly, a value of a potential variable suitable for a plurality of preset keyword sets may be calculated. Although FIG. 3 shows that two encoding LSTMs 210 and 220 and three decoding LSTMs 230 and 240 and 250 are included, the number of encoding LSTM and decoding LSTM varies according to the number of keywords in the input keyword set You can.

As illustrated in FIG. 3, the output of the LSTM 210 in the front stage may be a chain structure that is the input of the LSTM 220 in the rear stage. Since the variable circulating auto-encoder 200 of this embodiment sequentially inputs and encodes each keyword in LSTM, linkage between keywords may be considered.

Next, when the natural language is input from the user and the keyword set is generated by the morpheme analysis unit 100, the encoding LSTMs 210 and 220 calculate the latent variable for the keyword set corresponding to the natural language. The calculated latent variables can serve as a basis for future indexing.

As shown in FIG. 3, when the keyword set {this year, corporate tax} is input, “this year” is input to the LSTM 210 and the output of the LSTM 210 and the “corporate tax” are input to the LSTM 220. . Accordingly, a latent variable Z having an average (μ) and a variance (σ) is generated in the LSTM 220. Although FIG. 3 illustrates that two encoding LSTMs 210 and LSTMs corresponding to two keywords are included, the number of encoding LSTMs may vary depending on the number of keywords constituting the keyword set.

4 is a view showing an example of the structure of the LSTM of FIG. 3.

Referring to FIG. 4, Xt-1 and Xt are sequentially input with keywords of the keyword set generated by the morpheme analyzer 100. In the embodiment of FIG. 4, "this year" may be input to Xt-1, and "corporate tax" may be input to Xt. ht-1 and ht are encoded vector values, and finally output ht may correspond to the latent variable Z in FIG. 3.

The LSTM 210 and the LSTM 220 may have the same structure. Therefore, hereinafter, the LSTM 220 will be mainly described.

The LSTM 220 changes the cell state from Ct-1 to Ct when viewed as a whole. In Fig. 4, the square marks 41, 42, 43, and 47 indicate neural network layers, and the circle marks 44, 45, 46, and 49 represent point operations.

First, the sigmoid layer 41 receives ht-1 and xt and outputs ft having a value between 0 and 1. And ft is multiplied by Ct-1 representing the previous cell state (see 44) to determine how much to reflect the previous cell state. ft can be expressed by the following equation.

The sigmoid layer 42 and tanh layer 43 receive ht-1 and xt and determine how much to reflect in the current cell state (Ct) (see 46).

The sigmoid layer 42 receives ht-1 and xt similarly to the sigmoid layer 41 and outputs it having a value between 0 and 1, which can be expressed by the following equation.

를 생성한다. tanh layer(213)의 출력값

는 다음의 수식으로 표현될 수 있다.The tanh layer 43 receives ht-1 and xt, and is a new candidate value (vector).

Produces Output value of tanh layer (213)

Can be expressed by the following equation.

Accordingly, the current cell state Ct can be expressed by the following equation (see 44, 45 and 46).

Next, the sigmoid layer 47 receives ht-1 and xt and outputs a value between 0 and 1, and the output ot of the sigmoid layer 47 can be expressed by the following equation.

ht is output through point operations 48 and 49, and ht can be expressed as follows.

The LSTM having the structure of FIG. 4 has been described as an example, but the scope of the present invention is not limited thereto. For example, a portion of the neural network layers 41, 42, 43, 47 may be omitted, and the connection of the neural network layers 41, 42, 43, 47 and point operations 44, 45, 46, 48, and 49 Relationships may be different.

Next, Embodiment 2 of the present invention will be described with reference to FIGS. 5 and 6.

5 is a view showing the configuration of the index system 1000 of the variable circulating auto-encoding method according to an embodiment of the present invention.

Referring to FIG. 5, the indexing system 1000 is a system that receives a natural language in the form of text or voice or receives an image, and outputs an index value to search for a response corresponding thereto. The indexing system 2 includes a morphological analysis unit 1100, a variable circulating autoencoder 1200, an indexing unit 1300, and an updating unit 1400.

Referring to FIG. 5, the morpheme analysis unit 1100 receives a user's text or voice type natural language and generates keyword sets. Since the function of the morpheme analysis unit 1100 is the same as the morpheme analysis unit 100 described with reference to FIGS. 1 and 2, a detailed description thereof will be omitted.

The variable circulating autoencoder 1200 includes a first variable circulating autoencoder VAE1 and a second variable circulating autoencoder VAE2.

The first variable cyclic autoencoder (VAE1) learns the value of the latent variable (Z) through a training process of encoding and decoding a predetermined predetermined keyword set, and based on this, sets the keyword corresponding to the user's voice or character input The value of the latent variable Z is calculated by encoding.

The second variable cyclic autoencoder (VAE2) learns the value of the latent variable (Z) through a training process of encoding and decoding a predetermined predetermined image, and based on this, encodes the user's image input to determine the latent variable (Z) Calculate the value of

The indexing unit 1300 includes the value of the latent variable Z learned from the first variable circulating autoencoder VAE1 or the second variable cyclic autoencoder VAE2, and the first variable cyclic autoencoder VAE1 or the first The input from the user is indexed based on the result of comparing the value of the latent variable Z calculated by the two-variable cyclic autoencoder (VAE2).

The updater 1400 updates the index using the value of the latent variable Z calculated by the first variable cyclic autoencoder VAE1 or the second variable cyclic autoencoder VAE2.

FIG. 6 is a view for explaining the operation of the first variable circulating autoencoder (VAE1), the second variable circulating autoencoder (VAE2), and the updating unit 1400 of FIG. 5.

The first variable circulating autoencoder VAE1 is the same as the variable circulating autoencoder 200 except that the latent variable Z is shared with the second variable circulating autoencoder VAE2. Specifically, the first incremental cyclic autoencoder (VAE1) includes: a first encoding unit (1210, 1220) for encoding a predetermined set of keywords or a keyword set corresponding to the user's voice or character input; And first decoding units 1230, 1240, and 1250 that decode the value of the latent variable Z generated by the first encoding units 1210 and 1220 encoding predetermined predetermined keyword sets.

The first encoding units 1210 and 1220 include the same number of LSTMs 1210 and 1220 as the number of keywords constituting the input keyword set, and the keywords are input to the LSTMs 1210 and 1220, respectively, in the input order. The LSTMs 1210 and 1220 may be connected in a chain form.

The second incremental cyclic autoencoder (VAE2) includes: a second encoding unit 1211, 1221 for encoding a predetermined predetermined image or the image input of the user; And a second decoding unit 1231, 1241 for decoding the value of the latent variable Z generated by the second encoding units 1211, 1221 encoding a predetermined predetermined image, and the second encoding unit 1211 , 1221) and the second decoding units 1231 and 1241 may include a convolutional neural network (CNN).

An image from a user is input to the second incremental circulation autoencoder (VAE2). For example, tax-related documents may include different images depending on the type. Depending on the type of corporate tax delay, income tax refund, VAT, etc., the image or color of the form included in the document may be different. The second variable circulation autoencoder (VAE2) calculates the value of the latent variable (Z) according to this image.

As shown in FIG. 6, the second incremental circulation autoencoder (VAE2) includes a convolutional neural network (CNN). CNN refers to the structure of coping with the existing deep neural etwork (DNN) full-connected layer (FC) as a convolution layer. Since FC can receive only one-dimensional data, spatial information is lost (inappropriate) when processing three-dimensional image data (Width, Height, Channel (3 Channel for RGB, 1 Channel for Black and White)). CNN generally consists of Convolutional Layer-Pooling-Convolutional Layer-Pooling-Fully Connected Layer, and is suitable for image processing. The Convolutional Layer creates a Feature Map by summing the filter (kernel) input image and element-wise multiplication.

At this time, the value of the potential variable Z learned from the first variable circulating autoencoder VAE1 and the second variable circulating autoencoder VAE2 is shared. Accordingly, the first variable circulating autoencoder VAE1 and the second variable circulating autoencoder VAE2 learn and calculate the value of the potential variable Z under the influence of each other. In other words, the first variable cyclic autoencoder (VAE1) is based on the value of the potential variable (Z) learned and calculated by the second variable cyclic autoencoder (VAE2) as well as the first variable cyclic autoencoder (VAE1). The value of the latent variable Z corresponding to the input is calculated, and the second variable circulating autoencoder VAE2 is learned and calculated by the first variable cyclic autoencoder VAE2 as well as the second variable cyclic autoencoder VAE2. Based on the value of the latent variable Z, a value of the latent variable Z corresponding to a user's input is calculated.

The updating unit 1400 updates the index of the indexing unit 1300 by using the value of the latent variable Z calculated by the first variable circulating autoencoder VAE1 or the second variable circulating autoencoder VAE2.

For example, as illustrated in FIG. 6, the user inputs the text “This year's corporate tax”, and the value of the potential variable Z from the first variable circulation autoencoder (VAE1) is [0.5, 2.0, 0.4. ] Is output, the update unit 1400 updates the index unit so that the value of the potential variable of [0.5, 2.0, 0.4] and the index value of the corporate tax item correspond.

Alternatively, if the user inputs the image shown in FIG. 6 and the value of the potential variable Z is output from the second variable circulating autoencoder (VAE2) [1.5, 0.7, 0.9], the updater 1400 displays [ 1.5, 0.7, 0.9] The index is updated so that the value of the latent variable corresponds to the index value of the year-end settlement item.

Next, with reference to FIG. 7, a specialized field response service system 2 of a variable circulating auto-encoding method according to an embodiment of the present invention will be described.

Referring to FIG. 7, the specialized field response service system 2 includes an input / output unit 2100, an indexing system 2200, a search unit 2300, and a database 2400.

When the user inputs a question into a chat window displayed on a screen of an input terminal such as a mobile phone or a computer, the input / output unit 2100 transmits the input question to the indexing system 2200. In addition, when the answer corresponding to the input question is searched in the database 2400, the input / output unit 2100 outputs the searched answer in the chat window. The input / output unit 2100 may perform a function of converting text into text when a user inputs it by voice.

Additionally, the input / output unit 2100 may block access to indiscriminate data by issuing an authentication key for each user. Specifically, the input / output unit 210 may generate an authentication key for each user, store it in a user database (not shown), and provide the generated authentication key to the user.

The indexing system 2200 may be the indexing system 1 according to Embodiment 1 of the present invention or the indexing system 1000 according to Embodiment 2 of the present invention. The indexing system 2200 calculates the value of the latent variable according to a user's input, and outputs an index value corresponding to the calculated value of the latent variable.

The search unit 2300 searches for an answer in the database 2400 according to the index value received from the index system 2200. Then, the searched answer is provided to the input / output unit 2100 to be output to the user.

A plurality of questions and corresponding answers are stored in the database 2400, and the plurality of questions are classified according to index values. The questions stored in the database 2400 and answers corresponding thereto may be related to a specific specialized field, for example, a tax field. However, the scope of the present invention is not limited to this, and may be related to other specialized fields such as law and finance.

8 is a view showing a specialized field response service method using a variable rotation auto-encoding method using a chatbot according to an embodiment of the present invention.

Referring to FIG. 8, in a method for responding to a specialized field, first, a database 2400 for a specialized field is generated (S100). As described above, in the database 2400, answers according to questions in a specific field of expertise are stored, and the questions may be indexed (classified). In addition, with the creation of the database 2400, the index system 2200 learns the value of the latent variable using the variable cyclic autoencoder. For example, in the case of the first embodiment, the value of the latent variable may be learned through a training process of encoding and decoding a predetermined set of predetermined keywords using a variable cyclic autoencoder including LSTM. In the case of the second embodiment, a potential variable is learned through a training process of encoding and decoding a predetermined set of keywords using a first variable cyclic autoencoder including LSTM, and a predetermined image is CNN. Through the training process of encoding and decoding using the second variable cyclic autoencoder including, the value of the latent variable shared with the latent variable learned in the first variable cyclic autoencoder may be learned.

Next, the input is received from the user (S110). The user may input a question about a specialized field in the form of text or voice through the user terminal or in the form of an image. The input / output unit 2100 of FIG. 7 transmits the input question to the indexing system 2200.

Next, the morpheme analysis unit of the indexing system 2200 analyzes the received morpheme of the natural language to generate a set of keywords corresponding to the natural language (S120). The morpheme analysis unit may generate a keyword set composed of one or more keywords except for a survey having no special meaning. This step is performed only when the user's input is in the form of text or voice, and is omitted in the form of an image.

Next, the variable cycle autoencoder of the indexing system 2200 calculates a potential variable by encoding a keyword set generated by the morpheme analysis unit or an image input from a user (S130).

Next, the updating unit of the indexing system 2200 updates the index value according to the value of the potential variable generated by the variable rotation auto-encoder (S140).

Next, the search unit 2300 searches for an answer in the database 2400 according to the index value (S150).

Depending on the embodiment, the order of the index updating step S140 and the searching step S150 may be changed.

9 is a diagram showing experimental results of a specialized field response service system and method of a variable circulating auto-encoding method according to Embodiment 1 of the present invention.

The data used in the experiments used more than 1000 sets of questions and expert answers, such as corporate tax, value-added tax, and year-end settlement in the Knowledge IN Naver tax field. The learning rate was set to 0.001 and the batch size to 32, and the latent space was studied in 2D and 3D, respectively.

In FIGS. 9A and 9B, blue represents corporate tax, red represents VAT, and green represents year-end settlement. As shown in FIG. 9, it can be seen that corporate tax, VAT, and year-end settlement are clearly classified according to the value of the potential variable.

As described above, the present invention has been described in detail through preferred embodiments, but the present invention is not limited thereto, and various modifications and applications can be made without departing from the spirit of the present invention. It is obvious to the technician. Therefore, the true scope of protection of the present invention should be interpreted by the following claims, and all technical spirits within the equivalent scope should be interpreted as being included in the scope of the present invention.

Claims (12)

The first variable for calculating the value of the potential variable by learning the value of the potential variable through the training process of encoding and decoding the predetermined keyword set, and encoding the keyword set corresponding to the user's voice or character input based on this training process Cyclic autoencoder (Variational Recurrent Autoencoder);
A second variable circulating autoencoder that learns the value of the latent variable through a training process of encoding and decoding a predetermined image and calculates the latent variable value by encoding a user's image input based on the training process;
The result of comparing the value of the potential variable learned by the first variable circulating autoencoder and the second variable circulating autoencoder and the value of the potential variable calculated by the first variable cyclic autoencoder or the second variable cyclic autoencoder An index unit that indexes input from a user based on the;
It includes,
The index system of the variable circulating autoencoding method, characterized in that the values of the potential variables learned from the first variable cyclic autoencoder and the second variable cyclic autoencoder are shared. According to claim 1,
A keyword generator that generates a set of keywords corresponding to the user's text or voice input by analyzing the morphemes of the user's text or voice input
Index system of the variable circulating auto-encoding method further comprising a. According to claim 1,
The first variable circulation autoencoder,
A first encoding unit encoding the predetermined keyword set or a keyword set corresponding to the user's voice or character input; And
A first decoding unit for decoding the value of the latent variable generated by the first encoding unit encoding the predetermined keyword set
Index system of the variable circulating auto-encoding method comprising a. According to claim 3,
The first encoding unit,
Includes the same number of long term term memory (LSTM) as the number of keywords constituting the input keyword set,
The keywords are respectively input to the LSTM in the input order,
The LSTM is an index system of a variable circulating autoencoding method, characterized in that it is connected in a chain form. According to claim 1,
The second variable circulation autoencoder,
A second encoding unit encoding the preset predetermined image or the user's image input; And
A second decoding unit decoding the potential variable generated by encoding the predetermined image by the second encoding unit
Including,
The second encoding unit and the second decoding unit comprises a convolutional neural network (CNN) index system of the variable cyclic auto-encoding method. According to claim 1,
An update unit that updates the index of the index unit by using the value of the latent variable calculated by the first variable cyclic autoencoder or the second variable cyclic autoencoder.
Index system of the variable cyclic auto-encoding method further comprising a. According to claim 1,
The predetermined keyword set is related to taxation,
The preset predetermined image is an index system of the variable circulating auto-encoding method, characterized in that it is an image capable of distinguishing a type of a document related to taxation. A potential variable is learned through a training process of encoding and decoding a predetermined set of predetermined keywords using a first variable cyclic autoencoder, and encoding and decoding a predetermined image using a second variable cyclic autoencoder. Learning a value of a latent variable shared with a latent variable learned in the first variable circulation autoencoder through a training process;
Calculating a value of a potential variable corresponding to a user input using the first variable cyclic autoencoder or the second variable cyclic autoencoder; And
Indexing input from a user based on a result of comparing the value of the learned latent variable with the calculated latent variable value;
Index method of the cyclic cyclic autoencoding method comprising a. The method of claim 8,
Generating a keyword set corresponding to the user's text or voice input by analyzing a morpheme of the user's text or voice input
Further comprising,
In the calculating step,
When the user input is text or voice, the index system of the variable circulating autoencoding method, characterized in that the potential variable corresponding to the generated keyword set is calculated using the first variable cyclic autoencoder. The method of claim 8,
Updating the index used in the indexing step by using the value of the latent variable calculated by the first variable cyclic autoencoder or the second variable cyclic autoencoder.
Indexing method of the circulating variable auto-encoding method further comprising a. The method of claim 8,
The predetermined keyword set is related to taxation,
The preset predetermined image is an index method of a variable circulating auto-encoding method, characterized in that it is an image capable of distinguishing a type of a document related to taxation. A keyword generator configured to generate a set of keywords corresponding to the user's natural language input by analyzing a morpheme of the user's natural language input;
A variable cyclic autoencoder that learns the value of a potential variable through a training process of encoding and decoding a predetermined keyword set and calculates the value of the potential variable by encoding the keyword set generated by the keyword generator based on this;
An index unit that indexes input from a user based on a result of comparing a value of a potential variable learned by the variable cyclic autoencoder with a value of a potential variable calculated by the variable cyclic autoencoder; And
An update unit that updates the index of the index unit by using the value of a latent variable calculated by the variable cyclic autoencoder.
Index system of the variable circulating auto-encoding method comprising a.

Images