KR102017229B1 - A text sentence automatic generating system based deep learning for improving infinity of speech pattern - Google Patents

Abstract

The present invention relates to a system to automatically generate a text sentence based on deep learning for increasing the infinity of speech patterns. More specifically, common intention information of texts and speech patterns of the texts are inputted through a deep learning modeling part (100) to conduct deep learning training, and then, conducted deep learning model information by intention is provided to an automatic test sentence generation part. The automatic test sentence generation part (300) is used to obtain the deep learning model information by intention from the deep learning modeling part, and then, an intention type provided through an output intention input part (200) is inputted into the corresponding deep learning model information by intention to output an automatically generated natural language speech pattern text sentence, and then, a credibility value for the outputted text sentence is outputted.

Description

A text sentence automatic generating system based deep learning for improving infinity of speech pattern}

The present invention relates to a deep learning-based automatic text sentence generation system for improving the infinity of the speech pattern, and more particularly, by using the speech pattern of the texts and the intention information of the texts, which are deep learning materials collected from native speakers. Performing deep learning training, deriving deep learning model information for each intention as a result, and generating and outputting a plurality of text sentences having a specific intention by applying specific intention information to the derived deep learning model information for each intention The present invention relates to a deep learning based text sentence generation system for improving the infinity of speech patterns.

With the development of speech recognition technology, the demand for natural language recognition is increasing.

In the prior art, natural language recognition is performed using various models such as rules, statistical models, and deep learning model information for each intention to identify intentions and extract objects. Data collection is not easy.

In other words, a large number of utterance data must be collected before this model can be created. Since most of the collected data has already been used to generate the model, it is not suitable for reuse for verification purposes. Accordingly, new natural sentence data for verification must be collected, which is a very expensive and time-consuming task.

Therefore, there is a need for a technology for generating a new test sentence based on data collected using deep learning technology and a technology for obtaining pure verification data not used for model generation.

In addition, in the prior art, the format of the recognition result was changed due to the application of the Dictation and NLU technologies. Therefore, automatic evaluation was impossible with the existing Grammar Matching method, and there is a need for automatic correction of errors occurring in the existing evaluation system.

In addition, in the conventional technology, as the natural language processing technology is applied to the speech recognition equipment, speech patterns in all cases can be input as speech recognition commands, but manual creation and testing of the potential speech patterns of the user are physically time-consuming. There is a cost and a limit to the patterns that can be generated.

Accordingly, there is a need for the development of an automated system that automatically generates a user's potential speech pattern that improves the infinity of the speech pattern.

In the present invention, in order to improve the problems of the prior art as described above, deep learning training is performed using the speech patterns of texts, which are the deep learning learning materials collected from native speakers, and common intention information of the texts, and the deep learning model for each intention as a result. We propose a deep learning-based natural language utterance pattern text sentence generation system for automatically generating and outputting a large number of text sentences with specific intent by deriving information and applying specific intent information to the derived deep learning model information. Was done.

(Previous Document 1) Republic of Korea Patent No. 10-0733469

Accordingly, the present invention has been proposed in view of the above-described problems of the prior art, and a first object of the present invention is to perform deep learning training by using utterance patterns of native speakers' texts, which are collected deep learning materials, and intention information of the texts. The purpose is to generate deep learning model information for each intent.

A second object of the present invention is to automatically generate text sentences having a specific intention by reflecting specific intention information in the generated deep learning model information for each intention.

A third object of the present invention is to utilize the generated text sentence as an evaluation text of Natural Language Understanding (NLU) of the automatic speech recognition system.

In order to achieve the problem to be solved by the present invention, deep learning based text sentence automatic generation system for improving the infinity of the speech pattern,

Deep learning training is performed by receiving various texts, which are deep learning learning materials having speech pattern information and intention information, to generate deep learning model information for each intention including a plurality of deep learning models, and to generate deep learning model information for each intent generated. And the deep learning modeling unit 100 to provide a test sentence automatic generation unit 300,

An output intention input unit 200 for providing the test sentence automatic generation unit 200 with intent type information necessary for generating the text sentence of the test sentence automatic generation unit 200;

By using the deep learning model information for each intent provided by the deep learning modeling unit 100 and the intention type information provided by the output intention input unit 200, a text sentence corresponding to the intention type information is generated and output. And a test sentence automatic generation unit 300 for outputting a confidence value for the text sentence.

The deep learning based automatic text sentence generation system for improving the infinity of the utterance pattern according to the present invention having the above-described configuration and operation utilizes various utterance patterns having specific intentions collected in order to be used as the existing data (deep learning material). And automatically generate and output a new text sentence (text sentence having the same intention as the collected speech pattern of the collected text but having a different speech pattern) based on the collected data based on the collected text). It will provide the effect of improving the infinity of the potential speech pattern.

In addition, it is possible to automatically obtain text sentences (pure verification data not used for model generation) by intention type without collecting corpus (corpus) directly for the user, which is the NLU of the automatic speech recognition system. By using it as an evaluation text of (Natural Language Understanding), it is possible to reduce a long time and a huge cost by manually creating and testing a user's potential speech pattern.

1 is an overall configuration diagram schematically showing a deep learning based automatic text sentence generation system for improving the infinity of the speech pattern according to a first embodiment of the present invention.
2 is a block diagram of a deep learning modeling unit 100 of a deep learning based automatic text sentence generation system for improving the infinity of the utterance pattern according to the first embodiment of the present invention.
3 is a basic diagram illustrating a basic structure of an RNN of a deep learning based text sentence automatic generation system for improving infinity of a speech pattern according to a first embodiment of the present invention.
Figure 4 is a block diagram of the automatic test sentence generation unit 300 of the deep text-based automatic text sentence generation system for improving the infinity of the utterance pattern according to the first embodiment of the present invention.
5 is an automatic text sentence generation structure diagram generated by the test sentence automatic generation unit 300 of the deep text-based automatic text sentence generation system for improving the infinity of the utterance pattern according to the first embodiment of the present invention.
6 is an exemplary structure and training parameters of an RNN algorithm of an automatic text sentence generation system based on deep learning for improving infinity of a speech pattern according to a first embodiment of the present invention.
7 is an exemplary RNN model training diagram of a deep learning based automatic text sentence generation system for improving the infinity of the speech pattern according to the first embodiment of the present invention.
8 is an exemplary diagram illustrating an automatic generation result test generated by a deep learning based text sentence automatic generation system for improving the infinity of the utterance pattern according to the first embodiment of the present invention.
9 is an exemplary view of outputting a text sentence for each intent type generated by an automatic system for automatically generating text sentences based on deep learning to improve infinity of the speech pattern according to the first embodiment of the present invention.

The following merely illustrates the principles of the invention. Therefore, those skilled in the art, although not explicitly described or illustrated herein, can embody the principles of the present invention and invent various devices that fall within the spirit and scope of the present invention.

In addition, all conditional terms and embodiments listed herein are in principle clearly intended to be understood only for the purpose of understanding the concept of the invention and are not to be limited to the specifically listed embodiments and states. do.

In describing the present invention, terms such as first and second may be used to describe various components, but the components may not be limited by the terms.

For example, without departing from the scope of the present invention, the first component may be referred to as the second component, and similarly, the second component may also be referred to as the first component.

When a component is referred to as being connected or connected to another component, it may be understood that the component may be directly connected to or connected to the other component, but there may be other components in between. .

The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention, and singular forms may include plural forms unless the context clearly indicates otherwise.

In this specification, the terms including or including are intended to designate that there exists a feature, a number, a step, an operation, a component, a part, or a combination thereof described in the specification, and one or more other features or numbers, It can be understood that it does not exclude in advance the possibility of the presence or addition of steps, actions, components, parts or combinations thereof.

In the deep learning based automatic text sentence generation system for improving the infinity of the speech pattern according to the first embodiment of the present invention,

Deep learning training is performed by receiving various texts, which are deep learning learning materials having speech pattern information and intention information, to generate deep learning model information for each intention including a plurality of deep learning models, and to generate deep learning model information for each intent generated. And the deep learning modeling unit 100 to provide a test sentence automatic generation unit 300,

An output intention input unit 200 for providing the test sentence automatic generation unit 200 with intent type information necessary for generating the text sentence of the test sentence automatic generation unit 200;

By using the deep learning model information for each intent provided by the deep learning modeling unit 100 and the intention type information provided by the output intention input unit 200, a text sentence corresponding to the intention type information is generated and output. And a test sentence automatic generation unit 300 for outputting a confidence value for the text sentence.

In addition, the deep learning modeling unit 100,

A model text corpus unit 110 for receiving a plurality of texts having a specific speech pattern as a deep learning material and providing the input texts to the training modeling unit 130;

To receive the intents of the texts input to the model text corpus unit 110, tag the corresponding intentions to the corresponding texts, and provide the training modeling unit 130 with intent tagging information for each text tagged with the intents for each text. Model intention input unit 120;

Deep learning model for each intention by performing deep learning training using a plurality of texts having a specific speech pattern provided by the model text corpus 110 and intention tagging information for each text provided by the model intention input unit 120 And a training modeling unit 130 for generating the information and providing the generated deep learning model information for each test to the test sentence automatic generation unit 300.

At this time, the training modeling unit 130,

In the deep learning modeling is characterized in that it comprises a modeling parameter providing module 131 for adjusting the parameters related to the modeling.

In addition, the test sentence automatic generation unit 300,

By using the deep learning model information for each intent provided by the deep learning modeling unit 100 and the intention type information provided by the output intention input unit 200, various speech patterns having intentions corresponding to the intention type information are included. A calculation model unit 310 for automatically generating text sentences;

A corpus output unit 320 for outputting text sentences generated by the calculation model unit 310;

A confidence value output unit 330 for outputting a confidence value indicating how similar the text sentence output through the corpus output unit 320 to the intention of the intention type information input through the output intention input unit 200; Characterized in that it comprises a.

At this time, the calculation model unit 310,

And an output parameter providing module 311 for adjusting a parameter related to an output option when outputting the generated text sentence.

In addition, the text sentence generated and output by the test sentence automatic generation unit 300,

Characterized in that it is a text sentence having the same intention type but different speech patterns than the input text to be used as a deep learning learning material.

Further, the text sentence generated and output by the test sentence automatic generation unit 300,

Characterized in that it is used as the text for evaluation of the Natural Language Understanding (NLU) of the speech recognition automatic evaluation system.

Hereinafter, a deep learning based text sentence automatic generation system for improving the infinity of the speech pattern according to the present invention will be described in detail.

1 is an overall configuration diagram schematically showing a deep learning based automatic text sentence generation system for improving the infinity of the utterance pattern according to the first embodiment of the present invention.

As shown in FIG. 1, the deep learning-based text sentence automatic generation system 1000 for improving the infinity of the utterance pattern of the present invention includes a deep learning modeling unit 100, an output intention input unit 200, and a test sentence automatic system. It is configured to include a generation unit 300.

The deep learning modeling unit 100 receives various texts, which are deep learning learning materials having speech pattern information and intention information, and performs deep learning training to generate deep learning model information for each intention including a plurality of deep learning models. It is a means for providing the generated test results deep learning model information to the test sentence automatic generation unit 300.

For example, when the deep learning material input to the deep learning modeling unit 100 is '1.Find a panda restaurant', '2.I want to know who Sherlock Holmes' is, the pattern of speech of each text inputted The information is "Panda_Restaurant_Find", "I_Sherlock Holmes_I want to know who", and the intention information of each text entered is 'Restaurant Search' and 'People Search'.

As such, the deep learning modeling unit 100 performs deep learning training by receiving texts of a plurality of deep learning materials having speech patterns and intention information, and performs deep learning model information for each intention (a plurality of deep learning models). Including) to the test sentence automatic generation unit 300.

The output intention input unit 200 is a means for providing the test sentence automatic generation unit 300 with the intent type information necessary for generating the text sentence of the test sentence automatic generation unit 300.

For example, by providing a test sentence automatic generation unit 300 with an intention type of 'finding restaurant', the test sentence automatic generation unit 300 generates text sentences of various utterance patterns having an intention of 'finding restaurant'. To do that.

The test sentence automatic generation unit 300 corresponds to intention type information by using the deep learning model information for each intention provided by the deep learning modeling unit 100 and the intention type information provided by the output intention input unit 200. It is configured to generate and output a text sentence, and to output a confidence value for the output text sentence.

For example, if the intention type information provided by the output intention input unit 200 is 'find restaurant', among the deep learning model information for each intention provided by the deep learning modeling unit 100, 'find restaurant' is a deep intention Finding a restaurant with various speech patterns using the running model information generates and outputs text sentences intended.

2 is a block diagram of a deep learning modeling unit 100 of an automatic text sentence generation system with improved deep learning based speech pattern infinity according to a first embodiment of the present invention.

As shown in FIG. 2, the deep learning modeling unit 100 includes a model text corpus unit 110, a model intention input unit 120, and a training modeling unit 130.

In detail, the model text corpus unit 110 receives a plurality of texts having a specific speech pattern which is a deep learning material, and provides the input texts to the training modeling unit 130.

In particular, the plurality of texts having the specific speech pattern inputted have their respective intentions.

For example, a plurality of texts having a specific speech pattern inputted are 1.Where is the <Panda Restaurant> ?, 2.I want to look for <Panda Restaurant>, 3.Could you please find <Panda Restaurant>? Ramen, the texts are texts intended to find a restaurant called 'Panda'. That is, the texts are texts intended to be 'Find Restaurant'.

Similarly, as another example, texts having various speech patterns corresponding to various intentions such as music search, address search, weather search, stock search, etc. may be input through the model text corpus unit 110.

In particular, the input of the plurality of texts having a specific speech pattern, which is a deep learning material, into the model text corpus unit 110 is a manual input of collected texts by a person and an automatic input of texts collected by a text collecting robot. Characterized in that it comprises a.

In addition, the model intention input unit 120 receives the intentions of the texts inputted into the model text corpus unit 110, tags the corresponding intentions into the corresponding texts, and displays the intention tagging information for each text in which the intentions are tagged for each text. Characterized in that provided to the training modeling unit 130.

For example, when the text 1 inputted to the model text corpus unit 110 is "Where is the <Panda Restaurant>?" And the information of the text 1 is to find a restaurant, the information is input to the model input unit 120. The model intention input unit 120 tags the intention of "finding a restaurant" to the inputted text "Where is the <Panda Restaurant>?", And the intention tagging for each text tagged with the intention information of "finding a restaurant." Information (eg, “Where is the <Panda Restaurant>? _ Find a restaurant”) is provided to the training modeling unit 300.

Thereafter, the training modeling unit 130 uses the plurality of texts having a specific speech pattern provided by the model text corpus 110 and the intention tagging information for each text provided by the model intention input unit 120. By performing the running training to generate the deep learning model information for each intention, and performs the function of providing the generated deep learning model information for each intent to the test sentence automatic generation unit 300.

The training modeling unit 130 may utilize the conventional Recurrent Neural Network (RNN) algorithm illustrated in FIG. 3 to generate deep learning model information for each intention.

The Recurrent Neural Network (RNN) algorithm has a characteristic that the chain of the input texts can vary in sequence according to the time, so that not only one word is understood, but the relationship between the previous word and the current word is studied. Since it is an algorithm for predicting a word afterwards, a conventional description thereof will be omitted.

On the other hand, according to an additional aspect, the training modeling unit 130 is characterized in that it further comprises a modeling parameter providing module 131 for adjusting the parameters related to modeling during deep learning modeling.

For example, during deep learning modeling, it is necessary to adjust the modeling order, depth, network type, and the like by adjusting modeling options required for modeling, such as modeling depth, modeling depth, modeling network type, and the like. 131 provides modeling parameters related to these modeling options to the training modeling unit 130 so that modeling may be performed using the provided parameters.

FIG. 4 is a block diagram of an automatic test sentence generation unit 300 of a system for automatically generating text sentences based on deep learning for improving infinity of a speech pattern according to a first embodiment of the present invention.

5 is an automatic text sentence generation structure diagram generated by the test sentence automatic generation unit 300 of the deep text-based automatic text sentence generation system for improving the infinity of the utterance pattern according to the first embodiment of the present invention.

As shown in FIGS. 4 to 5, the test sentence automatic generation unit 300 includes a calculation model unit 310, a corpus output unit 320, and a confidence value output unit 330.

The computational model unit 310 corresponds to the intention type information using the intention type deep learning model information 50 provided from the deep learning modeling unit 100 and the intention type information provided through the output intention input unit 200. Automatically generate text sentences with various intention patterns.

For example, 50,000 native speakers are input to the deep learning modeling unit 100 and the input texts are classified into 150 intention types, and 50,000 native texts have a specific speech pattern for each of 150 intention types. As described above, the training modeling unit 130 of the deep learning modeling unit 100 generates deep learning model information 50 for each intention by using a Recurrent Neural Network (RNN) algorithm and provides the computational model unit 310 to the computational model unit 310. will be.

The computational model unit 310 that receives the deep learning model information 50 for each intention from the deep learning modeling unit 100 and the intention type information from the output intention input unit 200 has an intention corresponding to the intention type information. Automatically generate text sentences with various speech patterns

Referring to FIG. 6, the operation model unit 310 provided with the deep learning model information 50 for each intention from the deep learning modeling unit 100 and the intention type information from the output intention input unit 200 is an example. For example, using the representative RNN algorithm called GRU Cell (Gated Recurrent Unit) automatically generates text sentences of various speech patterns as output values.

The automatic generation of text sentences of various speech patterns using an RNN algorithm called a GRU Cell (Gated Recurrent Unit) will be described with reference to FIG. 7.

For example, the input data <STR> corresponds to the intention type information provided from the output intention input unit 200. For example, when the intention type information is "find restaurant", the input data <STR> is S1. If it passes, the output is Find, and if the input data, Find, is passed to S2, the output is me, and if you pass me to S3, the output is POI (meaning destination) and POI (meaning destination) is S4. If you pass in, a period is terminated and a sentence is printed, which creates a text sentence with a specific speech pattern with the intention of "Find me POI."

The corpus output unit 320 outputs a text sentence automatically generated by the calculation model unit 310. In this case, the text sentences automatically generated by the calculation model unit 310 may be text sentences having the same intention as the text input to the deep learning modeling unit 100 as the deep learning learning material but having different speech patterns.

For example, the text entered into the deep learning modeling unit 100 as a deep learning material is 1. Where is the <Restaurant> ?, 2. I want to look for <Restaurant>., 3. Could you please find < Restaurant>? If the text sentence automatically generated by the calculation model unit 310 is 4. Find <Restaurant>., 5. I want to find <Restaurant>., 6. Could you please looking for <Restaurant>., 7. Where is the <Restaurant>. As such, the text sentence is the same as "find restaurant" but different speech patterns are generated, and the generated text sentences are output through the corpus output unit 320.

Therefore, when the intention type information of 'finding a restaurant' is input through the output intention input unit 200, 'finding a restaurant' has the same intention, but text sentences having different utterance patterns, such as 4 to 7, are calculated. The text sentences generated by the model unit 310 are generated through the corpus output unit 320.

That is, the text sentence generated and output by the test sentence automatic generation unit 300 has the same intent type as the texts input to the deep learning modeling unit 100 to be used as a deep learning material, but has a different speech pattern. Characterized in that the sentence.

The confidence value output unit 330 outputs a confidence value indicating how similar the text sentence output through the corpus output unit 320 is to the intention of the intention type information input through the output intention input unit 200. do.

That is, a function of outputting a confidence value, that is, a probability value, which can confirm how similar the generated text sentence is to the intention of the intention type information input through the output intention input unit 200 is performed.

The calculation of the confidence value (probability value) between the output text sentence and the intent type can generally use a probabilistic distance value, and the detailed description thereof includes 'voice recognition', which is registered and filed by the applicant of the Republic of Korea Patent No. 10-1890704. Simple message output device and method using language modeling and 'Korean language registration No. 10-1913191', 'Language extraction based language understanding performance improving device and performance improvement method' are described in detail.

The technique of calculating the confidence value is a technique commonly used in statistics or speech recognition technology, and it is a general technique used in topic model, opinion mining, text summarization, data analysis, and public opinion survey. It is obvious that those skilled in the art can fully understand the above meanings.

Meanwhile, according to an additional aspect, the operation model unit 310 may further include an output parameter providing module 311 for adjusting a parameter related to an output option when outputting the generated text sentence.

For example, the parameter related to the output option that the output parameter providing module 311 can adjust may be an output text sentence quantity for each intent type and intention type.

For example, if the intention type is specified through the output parameter providing module 311 (eg, directions, radio search, address search, etc.), and the quantity of output text sentences for each designated intention type is adjusted to five, as shown in FIG. 9. In this case, five text sentences are automatically generated for each intent type.

In conclusion, by using the computation model unit 310, the corpus output unit 320, and the confidence value output unit 330 of the test sentence automatic generation unit 300, the pure natural language for verification that is not used for generating the deep learning model. It is possible to generate a pattern text sentence, and to use it as a test text sentence for a speech recognition device. In particular, the text sentence generated and output by the test sentence automatic generation unit 300 may be utilized as the text for evaluation of the Natural Language Understanding (NLU) of the automatic speech recognition system.

Next, an example of testing the adequacy of the text sentences output by intention, that is, reliability and redundancy of input sentences, by the deep learning-based automatic text sentence generation system for improving the infiniteness of the speech pattern according to the present invention will be described with reference to FIG. 8. Do it.

As shown in FIG. 8, when Navigate, Search, FM, Search Nearest, Create Calendar, Find Address, Search Menu, DisableGPS, StockMarketTrend, and Search Weather are input as the intent content of each intent type, the reliability, that is, reliability is 100%, respectively. , 100%, 100%, 100%, 100%, 80%, 100%, 100%, 60%, 100%, the average confidence value of 94% was confirmed that there is considerable reliability.

Also, the redundancy of input sentences is 40%, 40%, 80%, 40%, 0%, 60%, 20%, 40%, 20%, 20%, respectively, and the average input sentence redundancy reaches 36%. It can be seen that it has the advantage that new acquisition of pure verification data that is not used for generation is possible.

In addition, as shown in FIG. 9, a text sentence of a natural language utterance pattern newly automatically generated may be generated and output as described below with respect to Navigate, which is the intent type 1, and may be used as an NLU evaluation text.

In other words, Generated Pattern 1: navigate to a <POI> on <STREET> and <STREET>,

Generated Pattern 2: navigate me over to a <POI> on <STREET> and <STREET>,

Generated Pattern 3: navigate me over to <POI>,

Generated Pattern 4: navigate to the <STREET> and <STREET> intersection,

Generated Pattern 5: navigate to <POI> on the corner of <STREET> and <STREET>.

As a result, the probability of generating a non-legal sentence for each type of intention is different, but it shows that it shows a level of matching that can be fully utilized, and thus the newly printed sentences are refined (removing original duplicate sentences, removing non-legal sentences, After removing domain information shortage, etc., it can be used as NLU evaluation text.

As a result, according to the present invention, the deep learning modeling unit 100 performs a deep learning training by receiving native speaker texts having a speech pattern and intention information, and the test sentence automatic generation unit generates the deep learning model information for each intention which is a result of the execution. When provided to the 300, the test sentence automatic generation unit 300 has different speech patterns corresponding to the intention type using the intention type provided through the deep intention model information and the intention type provided through the output intention input unit 200. By generating and outputting text sentences, it is possible to improve the infinity of the user's potential speech pattern.

In addition, although the preferred embodiment of the present invention has been shown and described above, the present invention is not limited to the above-described specific embodiment, the technical field to which the invention belongs without departing from the spirit of the invention claimed in the claims. Of course, various modifications can be made by those skilled in the art, and these modifications should not be individually understood from the technical spirit or the prospect of the present invention.

100: deep learning modeling unit
200: output intention input unit
300: test sentence automatic generation unit

Claims (7)

In the deep learning based automatic text sentence generation system for improving the infinity of the speech pattern,
Deep learning training is performed by receiving various texts, which are deep learning learning materials having speech pattern information and intention information, to generate deep learning model information for each intention including a plurality of deep learning models, and to generate deep learning model information for each intent generated. And the deep learning modeling unit 100 to provide a test sentence automatic generation unit 300,

An output intention input unit 200 for providing the test sentence automatic generation unit 300 with the intent type information necessary for generating the text sentence of the test sentence automatic generation unit 300;

By using the deep learning model information for each intent provided by the deep learning modeling unit 100 and the intention type information provided by the output intention input unit 200, a text sentence corresponding to the intention type information is generated and output. It includes a test sentence automatic generation unit 300 for outputting a confidence value for the text sentence to be,

The deep learning modeling unit 100,
A model text corpus unit 110 for receiving a plurality of texts having a specific speech pattern as a deep learning material and providing the input texts to the training modeling unit 130;
To receive the intents of the texts input to the model text corpus unit 110, tag the corresponding intentions to the corresponding texts, and provide the training modeling unit 130 with intent tagging information for each text tagged with the intents for each text. Model intention input unit 120;
Deep learning model for each intention by performing deep learning training using a plurality of texts having a specific speech pattern provided by the model text corpus 110 and intention tagging information for each text provided by the model intention input unit 120 And a training modeling unit 130 for generating the information and providing the generated deep learning model information for each test to the test sentence automatic generation unit 300.

The training modeling unit 130,
The deep learning modeling is configured to further include a modeling parameter providing module for adjusting the parameters related to the modeling, wherein the parameters relating to the modeling is characterized in that the degree of modeling, modeling depth, modeling network type Deep learning based text sentence generation system to improve the infiniteness of patterns.

delete delete The method of claim 1,
The test sentence automatic generation unit 300,
By using the deep learning model information for each intent provided by the deep learning modeling unit 100 and the intention type information provided by the output intention input unit 200, various speech patterns having intentions corresponding to the intention type information are included. A calculation model unit 310 for automatically generating text sentences;
A corpus output unit 320 for outputting text sentences generated by the calculation model unit 310;
A confidence value output unit 330 for outputting a confidence value indicating how similar the text sentence output through the corpus output unit 320 to the intention of the intention type information input through the output intention input unit 200; Deep learning based automatic text sentence generation system for improving the infinity of the speech pattern, characterized in that comprises.
The method of claim 4, wherein
The calculation model unit 310,
It is configured to further include; output parameter providing module (311) for adjusting the parameters related to the output option when outputting the generated text sentence,
And a parameter related to the output option includes an intention type and an output text sentence quantity for each intention type.
The method of claim 1,
The text sentence generated and output by the test sentence automatic generation unit 300,
Deep learning-based text sentence automatic generation system for improving the infinity of the utterance pattern, characterized in that the text sentence having the same intention type and different utterance pattern than the input text to be used as deep learning learning materials.
The method of claim 1,
The text sentence generated and output by the test sentence automatic generation unit 300,
Deep learning based text sentence generation system for improving the infiniteness of speech patterns, characterized in that it is used as the text for evaluation of NLU (Natural Language Understanding) of the speech recognition automatic evaluation system.

Images