KR102375874B1 - A chat service system based on semantic relationships inferred sentence using artificial intelligence technology - Google Patents

Abstract

The present invention comprises a user terminal, a service providing device, and a learning device, the learning device comprising: an input unit for receiving learning document data to be learned; a word-vector conversion unit that performs a word vector conversion process for mapping the words of the learning document onto a vector space; a sentence-vector converting unit that performs sentence vector conversion processing corresponding to the sentences of the learning document by using the word vector; And according to the degree of the relationship between the sentence vectors, perform semantic relationship classification learning between sentences, generate a sentence-based classifier in which the semantic relationship is learned according to the semantic relationship classification learning, and the user input document based on the generated classifier. and a semantic relation classifying unit for performing intention classification processing, wherein the service providing apparatus can provide a conversation service by inferring a semantic relation of a user input sentence classified by the semantic relation classifying unit.

Description

A chat service system based on semantic relationships inferred sentence using artificial intelligence technology

The present invention relates to a conversation service system for inferring a semantic relationship of a sentence using artificial intelligence technology.

Artificial intelligence is revolutionizing business, how organizations operate, how they live, and how they communicate. Various informatization projects are underway to provide optimal services to the rapidly changing lifestyles of modern culture and diverse and constantly changing customer requirements.

Among them, as technologies related to big data and deep learning have developed at a rapid pace in recent years, artificial intelligence technology that is being applied to real life has been implemented in certain fields. It is also being applied to intelligent personal services.

In order to provide various user services such as artificial intelligence and data analysis services as described above, data mining for a very large amount of text and other language-based information is required, and for this, a machine learning method is mainly used.

In machine learning of such a document, a classification method using a document vector representing only a sentence as a vector value in which words of the current document are embedded is mainly used.

However, the existing classification method has a problem in that it is difficult to grasp the semantic relationship between sentences. As a result, such problems show limitations in the performance of the classifier.

For example, if there is a class called 'Europe' and a class called 'Soccer' that have different topics, and there is content related to 'European soccer' in the document data 'Europe', sentences related to 'Europe' and 'Soccer' It can be said that the sentences of the 'soccer' class and 'soccer' have a semantic relationship, but the current classifier does not process any related information in relation to it.

In other words, since the existing methods learn and classify data for each given sentence class, even if any semantic relationship appears between sentences of different classes, that is, sentences with different subjects, even within the document, in the current method, the relevant I can't figure out the semantic relationship.

This problem leads to a decrease in class distinction for new sentences that are not pre-classified in conversation services such as machine learning-based chatbots and counseling services, and as a result, it does not lead to a normal conversation and leads to a decrease in service quality.

In addition, since the sentence vector and the classification process are currently integrated, it is difficult to apply the classification model variably even if the domain of the document is changed, so there is a problem in that classification efficiency also decreases.

(001) Korean Patent Application Laid-Open No. 10-2006-0043333 (May 15, 2006) (Title: System and method for determining intent of data and responding to data based on intent)

The present invention has been devised to solve the above-described problems, and by inferring the relationship between sentences to determine the meaning of a new sentence to determine the class, finally improving the classifier performance and facilitating the interpretation of the semantic relation. An object of the present invention is to provide a dialogue service system.

Sentence semantic relation inference dialogue service system according to an embodiment of the present invention for solving the above problems,

a user terminal capable of providing a conversation service by voice input and output with a user; a service providing device for providing a data processing-based user conversation service using a semantic relation model between sentences to a user terminal based on sentence semantic relation classification for input sentences input through the user terminal; On the basis of the learning document of the learning database, the learning model information generated by performing the learning model generation process of the user input document is stored in the learning database, and semantic relation classification learning between sentences is performed based on the learning document, and , a sentence-based classifier is generated according to the semantic relation classification learning, and sentence semantic relation classification processing is performed on the user input sentence input through the service providing device based on the generated classifier and output to the service providing device learning device; includes

The learning device,

A learning apparatus for providing a conversation service, comprising: an input unit for receiving learning document data to be learned; a word-vector conversion unit that performs a word vector conversion process for mapping the words of the learning document onto a vector space; a sentence-vector converting unit that performs sentence vector conversion processing corresponding to the sentences of the learning document by using the word vector; And according to the degree of the relationship between the sentence vectors, perform semantic relationship classification learning between sentences, generate a sentence-based classifier in which the semantic relationship is learned according to the semantic relationship classification learning, and the user input document based on the generated classifier. and a semantic relation classification unit that performs intention classification processing.

In addition, the learning method according to an embodiment of the present invention for solving the above problems,

A method of operating a learning apparatus for providing a conversation service, the method comprising: receiving learning document data as a learning object; performing word vector conversion processing for mapping the words of the learning document onto a vector space; performing sentence vector conversion processing corresponding to the sentences of the learning document by using the word vector; performing semantic relation classification learning between sentences according to the degree of relation between the sentence vectors; generating a sentence-based classifier in which a semantic relationship is learned according to the semantic relationship classification learning; and performing intention classification processing of the user input document based on the generated classifier.

On the other hand, the method according to an embodiment of the present invention for solving the above problems may be implemented as a program for executing the method in a computer and a recording medium in which the program is recorded.

According to an embodiment of the present invention, the words of the learning document are mapped on a vector space, and sentence vector conversion processing corresponding to the sentences of the learning document is performed using the word vector, and the degree of relationship between the sentence vectors is determined. Accordingly, by performing semantic relationship classification learning between sentences, a sentence-based classifier in which the semantic relationship is learned according to the semantic relationship classification learning may be generated.

Accordingly, according to an embodiment of the present invention, it is possible to provide a sentence-based classifier in which a semantic relation is learned that can determine the class by inferring the relation between sentences to determine the meaning of a new sentence, and accordingly, Through semantic relation-based sentence classification processing, it is possible to provide a sentence semantic relation learning apparatus and an operating method thereof for providing a dialogue service that improves classifier performance and facilitates semantic relation interpretation as it enables a richer response.

In addition, the present invention provides individual classification processes for each of the word vectorization, sentence vector, and classification process, so that the classification model can be variably applied even if the domain of the document is changed, thereby increasing system applicability and classification efficiency. has the advantage of being

1 is a conceptual diagram schematically illustrating an entire system according to an embodiment of the present invention.
2 is a block diagram illustrating a learning apparatus according to an embodiment of the present invention in more detail.
3 is a flowchart for explaining in more detail a method of operating a learning apparatus according to an embodiment of the present invention.
4 is a diagram for explaining a vector conversion process according to an embodiment of the present invention.
5 and 6 are diagrams illustrating a classification process and results according to an embodiment of the present invention.
7 is a flowchart illustrating a service providing process according to an embodiment of the present invention.
8 and 9 are diagrams illustrating result data according to convolutional neural network-based sampling and application of a triplet loss function according to an embodiment of the present invention.

The following is merely illustrative of the principles of the invention. Therefore, those skilled in the art will be able to devise various devices that, although not explicitly described or shown herein, embody the principles of the present invention and are included within the spirit and scope of the present invention. Further, it is to be understood that all conditional terms and examples listed herein are, in principle, expressly intended solely for the purpose of enabling the concept of the present invention to be understood, and not limited to the specifically enumerated embodiments and states as such. should be

Moreover, it is to be understood that all detailed description reciting the principles, aspects, and embodiments of the invention, as well as specific embodiments, are intended to cover structural and functional equivalents of such matters. It should also be understood that such equivalents include not only currently known equivalents, but also equivalents developed in the future, i.e., all devices invented to perform the same function, regardless of structure.

Thus, for example, the block diagrams herein are to be understood as representing conceptual views of illustrative circuitry embodying the principles of the present invention. Similarly, all flowcharts, state transition diagrams, pseudo code, etc. may be tangibly embodied on computer-readable media and be understood to represent various processes performed by a computer or processor, whether or not a computer or processor is explicitly shown. should be

In addition, the clear use of terms presented as processor, control, or similar concepts should not be construed as exclusively referring to hardware having the ability to execute software, and without limitation, digital signal processor (DSP) hardware, ROM for storing software. It should be understood to implicitly include (ROM), RAM (RAM) and non-volatile memory. Other common hardware may also be included.

The above objects, features and advantages will become more apparent through the following detailed description in relation to the accompanying drawings, and accordingly, those of ordinary skill in the art to which the present invention pertains can easily implement the technical idea of the present invention. There will be. In addition, in the description of the present invention, if it is determined that a detailed description of a known technology related to the present invention may unnecessarily obscure the gist of the present invention, the detailed description thereof will be omitted.

Hereinafter, a preferred embodiment according to the present invention will be described in detail with reference to the accompanying drawings.

1 is a diagram schematically illustrating an entire system according to an embodiment of the present invention.

Referring to FIG. 1 , a system according to an embodiment of the present invention may include a learning apparatus 100 , a service providing apparatus 200 , and a user terminal 300 , and the learning apparatus 100 includes a learning database 180 . ) or may be connected to the learning database 180 .

More specifically, the learning apparatus 100 may generate a document learning model by performing machine learning based on the learning document data stored in the learning database 180, and using the generated learning model , may output classification information corresponding to a user input sentence input from the service providing device 200 , or provide a response sentence corresponding thereto to the service providing device 200 .

In particular, the learning apparatus 100 according to an embodiment of the present invention can be used to provide a conversation service of the service providing apparatus 200, receives learning document data as a learning object, and converts the words of the learning document into a vector space. A word vector conversion process is performed to map to , and a sentence vector conversion process corresponding to the sentences of the learning document is performed using the word vector, and semantic relation classification learning between sentences is performed according to the degree of relationship between the sentence vectors. , generate a sentence-based classifier in which a semantic relation is learned according to the semantic relation classification learning, and perform intention classification processing of a user input document based on the generated classifier.

Furthermore, the learning apparatus 100 generates a sentence vector learning model by performing a convolutional neural network-based learning process between the sentence vectors, and a triplet loss between sentence vectors corresponding to the sentence vector learning model (Triple-loss) function is processed, and according to the processing result of the triplet loss function, it is possible to output a semantic relationship learning model between sentences based on the degree of the relationship between the sentence vectors, which infers the relationship between sentences to classify the same By learning the sentences to be close to each other. The output layer of the learning apparatus 100 may include a sentence vector expressing the relation between sentences after all relational inference learning between sentences is finished.

In addition, the service providing apparatus 200 may provide a data processing-based user conversation service using the semantic relationship model between the processed sentences to the user terminal 300 . The user conversation service uses the learning database 180 generated by the operation of the learning apparatus 100 according to an embodiment of the present invention and the sentence semantic relation model according to the conversation in response to the input sentence of the user terminal 300 . service may be included.

For example, the service providing apparatus 200 may provide a chatbot service, a consultation service, and the like to the user terminal 300 .

For example, since the user's expression method always changes, if an unclassified sentence is input, there may be a limit to the response. For this purpose, deep learning technology is being used in chatbots, but there is a problem that there is significantly less learning data to process responses to all inquiries. There is a problem with its use.

Therefore, the learning apparatus 100 according to an embodiment of the present invention has the advantage of being able to effectively classify similar sentences according to the semantic relation model between sentences even when unclassified sentences are input.

Meanwhile, the user terminal 300 and the service providing apparatus 200 may be connected by wire or wirelessly through a network, and each user terminal 100 and the service providing apparatus 300 are connected to an Internet network, a LAN for mutual communication between networks. , WAN, PSTN (Public Switched Telephone Network), PSDN (Public Switched Data Network), cable TV network, WIFI, mobile communication network and other wireless communication networks can transmit and receive data. In addition, each user terminal 300 and the service providing apparatus 200 may include respective communication modules for communicating with a protocol corresponding to each communication network.

In addition, the user terminal 100 described in this specification includes a mobile phone, a smart phone, a laptop computer, a digital broadcasting terminal, Personal Digital Assistants (PDA), Portable Multimedia Player (PMP), navigation, etc. may be included, but the present invention is not limited thereto, and may be various devices capable of user input and information display other than that.

In such a system, the user terminal 300 may be connected to the service providing apparatus 200 to receive the above-described data-based conversation service.

2 is a block diagram illustrating a learning apparatus according to an embodiment of the present invention in more detail.

Referring to FIG. 2 , the learning apparatus 100 according to an embodiment of the present invention may include an input unit 110 , a learning unit 120 , an operating unit 140 , and a learning database 180 , and the operating unit 140 . ) includes a word-vector conversion unit 141 , a sentence-vector conversion unit 142 , a semantic relation classification unit 143 , and a result output unit 144 .

First, the input unit 110 receives a learning document for learning from the learning database 180 and transmits it to the learning unit 120 to be used for learning, or receives a user input document transmitted from the service providing device 200 . It may be processed so that the result data that has been classified and processed is transmitted to the operation unit 140 and outputted from the result output unit 144 . The result data of the result output unit 144 may be transferred back to the service providing apparatus 200 .

Meanwhile, the learning unit 120 may perform a learning model creation process based on the operation of the operation unit 140 and the learning data input from the input unit 110 . The generated learning model information may be stored in the learning database 180 .

In particular, in generating the learning model, the operation unit 140 according to an embodiment of the present invention makes it possible to classify the semantic relationship of sentences by processing to generate a semantic relationship learning model between sentences.

To this end, the operation unit 140 includes a word-vector conversion unit 141 that performs a word vector conversion process for mapping the words of the learning document onto a vector space, and uses the word vector to obtain sentences of the learning document. The sentence-vector transforming unit 142 that performs sentence vector conversion processing corresponding to The generated sentence-based classifier may include a semantic relation classifier 143 that performs an intention classification process of a user input document based on the generated classifier.

In addition, the result output unit 144 may be composed of one or more output modules for outputting the classification processing result.

More specifically, the semantic relationship classification unit 143 generates a sentence vector learning model by performing a learning process based on a convolutional neural network (CNN) between the sentence vectors, and corresponds to the sentence vector learning model. A triplet-loss function between one sentence vector is processed, and according to the processing result of the triplet loss function, a semantic relationship learning model between sentences is generated based on the degree of relationship between the sentence vectors.

Here, the sentence-vector converting unit 142, when the word-vector converting unit 141 calculates word vector conversion values corresponding to the words of the learning document, the word according to the sentence vector conversion process using the word vector It is possible to generate sentence vectors expressed as vectors. As such, sentence vectors expressed as word vectors may be used as input data of the convolutional neural network.

Here, in order to increase the word recognition capacity, the learning document may be predetermined as a document collected from a predetermined information site, for example, a Wikipedia site. Various vocabularies can be accommodated from these documents, and semantic relation information between sentences expressed by word vectors of various vocabularies can be effectively classified.

On the other hand, the semantic relation classifier 143 can obtain a desired range within a sentence by adjusting the size of the filter size of the convolutional neural network, and thus the sentence is stacked using filters of different sizes. It can also handle progressive sentence classification from multiple ranges of words to phrases and sentences.

In addition, the semantic relation classification unit 143 applies a triplet-loss function, based on anchor data according to a classification class classified primarily according to the convolutional neural network learning, sentence vectors are It is classified into positive data, which is the same class, and negative data, which is a different class.

And, the semantic relation classification unit 143 adjusts the positive data of the same class as the anchor data so that the Euclidean distance from the anchor data is located closer and the distance from the negative data, which is another class, to be located farther, so that the cluster Balanced classification processing between data is possible. Such a triplet loss function may be a loss function that facilitates classification between data clusters.

Accordingly, the semantic relation classification unit 143 according to an embodiment of the present invention samples anchors and positive and negative data for each classification class according to the subject classification corresponding to the sentence vectors, and between the respective sampled data. It is possible to vary the loss value between the sampling data until the Euclidean distance value is within a certain range. can

In addition, when the value of the Euclidean distance between the sampling data is within a predetermined range, the semantic relationship classifying unit 143 may output a sentence vector learning model according to the currently adjusted weight value as the semantic relationship learning model between sentences.

In addition, the semantic relation classifier 143 may generate a sentence intention classifier using the semantic relation learning model between sentences, and when the user input document is included in a preset classification criterion, a classification corresponding to the classification criterion processing is performed, and when the user input document is not included in the preset classification criteria, similar sentences having similar intent to the user input document are indexed based on a sentence intent classifier using the semantic relationship learning model between sentences can do.

The result output unit 144 may output classification information corresponding to the indexed similar sentence, and may process a similar sentence to be provided to the user even if it is a sentence that is not in the classification criteria.

3 is a flowchart illustrating in more detail a method of operating a learning apparatus according to an embodiment of the present invention.

Referring to FIG. 3 , when a learning document is first inputted through the input unit 110 ( S101 ), the learning apparatus 100 according to an embodiment of the present invention performs word vector conversion processing corresponding to words in the learning document. do (S109).

Then, the learning apparatus 100 processes a sentence vector conversion using a word vector in the learning document (S111).

Thereafter, the learning apparatus 100 performs a semantic relation classification learning process between sentences according to sentence vector embedding ( S113 ), and generates a classifier using the sentences from which the semantic relation is learned ( S115 ).

Thereafter, the learning apparatus 100 performs an intention classification process of the user input document using a classifier (S117).

4 is a diagram for explaining a vector conversion process according to an embodiment of the present invention.

Referring to FIG. 4 , data used in the existing classification criterion-based chatbot may be input as learning document data, and the learning document may include 30 sentences written according to an expected scenario that a user can input. As shown in FIG. 4 , the learning document is stored as a label and content, and may be used as an input value of the word vectorization process based on the semantic relation classifying unit 143 of the learning unit 120 .

According to the vectorization process, the word-vector conversion unit 141 vectorizes the words of each sentence and maps them on a vector space, and the sentence-vector conversion unit 142 obtains sentence vector data using the word vector data. can do.

Thereafter, the semantic relation classifying unit 143 may input the sentence vector expressed as the word vector as an input value of the sampling module, and use the sampled vector data to generate the above-described convolutional neural network and apply the triplet loss function. It can be processed to indicate the semantic relationship between sentences through

For example, as shown in FIG. 4 , embedding vector data clustered by sampling may be generated, and the data on the vector space clustered as described above is an embedding vector for function processing of the semantic relation classifier 143 . It may be used as a data input value, and the semantic relationship classifying unit 143 may generate a semantic relationship learning model between sentences by learning the relationship between sentences using the clustered sentences.

5 and 6 are diagrams illustrating a classification process and results according to an embodiment of the present invention.

As shown in FIG. 5 , when the classifier 143 according to the semantic relationship learning model between sentences is generated, the operation unit 140 operates the word-vector converting unit 141 and the sentence-vector converting unit ( 142 ) and a conversation service based on the classifier 143 may be provided.

In particular, the word-vector conversion unit 141, the sentence-vector conversion unit 142, and the classifier 143 according to an embodiment of the present invention can each be independently generated, so future technological developments and domain changes are made. Even if it loses, it can be easily configured to change the classification model, and can be processed so that it can be independently connected to other processes. Accordingly, the classification process according to an embodiment of the present invention may be configured by easily combining with separate service processes.

The sample data used as input data may be data input by the user on the chatbot, or it may be a sentence that is not actually learned and is not defined in the existing classification. In general, as disclosed in the upper part of FIG. 6 , in the case of the existing classification method, the intention of the sentence may not be grasped.

On the other hand, referring to FIG. 5 , even if unlearned sample data is input, the operation unit 140 can check which sentence and semantic relationship it has. As a result, the classification result corresponding to each word and sentence is the predicted value. It can represent the closestness to the sentence with the highest label 10 in vector space.

Accordingly, the operation unit 140 may output the sentences having the label 10 with the highest predicted value through the result output unit 144 , which, as shown in the lower part of FIG. 6 , has a related meaning. Make it possible to provide diverse and rich answers.

7 is a flowchart illustrating a service providing process according to an embodiment of the present invention.

Referring to FIG. 7 , the learning apparatus 100 according to an embodiment of the present invention performs sentence learning processing based on an existing rule (existing rule) (S201), but when a user-defined sentence is input through the input unit 110 ( S203), it is determined whether the rule-based sentence is (S205).

First, in the case of a rule-based sentence, the learning apparatus 100 may respond to the service providing apparatus 200 to output the rule-based response content (S207).

However, if it is not a rule-based sentence, the learning apparatus 100 may transmit it to the operation unit 140 and process a similar sentence to be inferred through sentence semantic relation inference (S209), and a response corresponding to the inferred similar sentence It may respond to the service providing apparatus 200 so that the contents are output (S211).

As such, the existing rule-based chatbot service can only respond to user inquiries according to the written scenario, so there is a problem that the chatbot service cannot understand when an unsaved sentence comes in.

In the case of using the learning apparatus 100 capable of classifying a sentence semantic relationship according to an embodiment of the present invention, even if a sentence that has not been learned or classified or stored comes in, it is determined which sentence and semantic relation are similar through the learned sentence semantic relation classifier This will allow you to respond appropriately to the content.

8 and 9 are diagrams illustrating result data according to convolutional neural network-based sampling and application of a triplet loss function according to an embodiment of the present invention.

First, referring to FIG. 8, FIG. 8 is a diagram for explaining an example of a process of sampling each anchor, negative, and positive data in order to learn a convolutional neural network.

First, referring to Equation 1 below, a convolutional neural network variable for applying the triplet loss function may be set as follows.

Here, when C is a set of all classes existing in data, P representing the number of positive data may represent the data set size of any class included in C. Also, since Anchor is the same class as P, it can represent one of the classes of C like P. N, which expresses the number of negative data, can represent the number of data of a class different from P among the classes belonging to C, but with a smaller number of data sets than P.

And, the triplet loss function according to an embodiment of the present invention is to implement a classifier by adjusting the number of positive and anchor data samples to N, that is, equal to the negative.

For example, referring to FIG. 8 , in the drawing, class may indicate a classification of sentences having the same subject. Each S represents a sentence as a vector of word embeddings, and the number indicates the identification number of the sentence. As shown in FIG. 8 , a total of three classes exist in the data, class A may have a total of 100 sentences, class B may have a total of 120 sentences, and class C may have a total of 30 sentences.

Here, if the positive sentence set corresponding to a specific sentence classification is a sentence included in class A, the negative sentence set is a sentence that has fewer data than the number of data in class A among other classes other than class A, class B or class C. It can contain statements of Class C, which is a class.

Accordingly, a difference exists in the number of data in the negative and positive sets, the distance to the anchor can be calculated, and a balance according to the distance can be balanced for accurate classification. The criterion for determining the number of sample data for distance calculation can be based on the class with the smallest number of sentences among the classes, and therefore 30, the number of data of class C, can be the criterion for the number of sample data.

Accordingly, the semantic relation classification unit 143 may apply the triplet loss function to sample data for sentence embedding for learning the relation between the sentences, and learn the semantic relation of each sentence according to distance calculation and adjustment. It can be processed so that a learning model that infers is created. Equation 2 below shows the triplet loss function application formula.

는 anchor 의 임베딩 값,

는 positive 의 임베딩 값,

는 negative 의 임베딩 값이라고 할 수 있다. In Equation 2, f is an embedding function, a is acho, p is positive, and n is negative. therefore

is the embedding value of anchor ,

is the embedding value of positive ,

can be said to be an embedding value of negative .

이 만족하도록 anchor 와 positive 사이의 거리는 작아지게 하고 negative 사이의 거리는 커지게 조절할 수 있다.

는

와

가 0 이 되지 않게 하기 위한 즉 두 개의 값이 너무 가깝지 않게 하는 margin 을 조절하는 하이퍼파라미터이다.And, the semantic relation classification unit 143, using the Euclidean distance, the distance between the anchor and the positive, the difference between the anchor and the negative distance

To satisfy this, the distance between the anchor and the positive can be made small and the distance between the negatives can be adjusted to be large.

Is

Wow

It is a hyperparameter that adjusts the margin so that the value does not become 0, that is, the two values are not too close.

9 shows changes in the size of input data and result data from the sampling process for applying the triplet loss function to the process of adjusting the loss within a certain range according to a convolutional neural network (CNN) and the triplet loss function.

In this way, the semantic relationship classification unit 143 uses a loss function using a triplet loss to make the loss of data of the same class, Anchor, and Positive, smaller, to make the distance closer, and to increase the loss between Anchor and Negative, which is another class, to be smaller. We can adjust the weights of the convolutional neural network to create Accordingly, the semantic relation classifier 143 may generate a vector model classifier in which sentence semantic relations for various classes are learned.

Meanwhile, the above-described method according to various embodiments of the present invention may be implemented as a program and provided to each server or device while being stored in various non-transitory computer readable media. Accordingly, the user terminal 100 may access the server or device to download the program.

The non-transitory readable medium refers to a medium that stores data semi-permanently, rather than a medium that stores data for a short moment, such as a register, cache, memory, and the like, and can be read by a device. Specifically, the above-described various applications or programs may be provided by being stored in a non-transitory readable medium such as a CD, DVD, hard disk, Blu-ray disk, USB, memory card, ROM, and the like.

In addition, although preferred embodiments of the present invention have been illustrated and described above, the present invention is not limited to the specific embodiments described above, and the technical field to which the present invention belongs without departing from the gist of the present invention as claimed in the claims In addition, various modifications are possible by those of ordinary skill in the art, and these modifications should not be individually understood from the technical spirit or perspective of the present invention.

Claims (5)

In the conversation service system,
a user terminal capable of providing a conversation service by voice input and output with a user;
a service providing apparatus for providing a data processing-based user conversation service using a semantic relation model between sentences to a user terminal based on sentence semantic relation classification for input sentences input through the user terminal;
Based on the learning document of the learning database, the learning model information generated by performing the learning model generation process of the input sentence is stored in the learning database, and semantic relation classification learning between sentences is performed based on the learning document, generating a sentence-based classifier according to the semantic relation classification learning, performing sentence semantic relation classification processing on the input sentence input through the service providing device based on the generated classifier, and outputting it to the service providing device learning device; including,
The learning device,
an input unit for receiving learning document data as a learning target;
a word-vector conversion unit that performs word-vector conversion for mapping the words of the learning document onto a vector space;
a sentence-vector converting unit that performs sentence vector conversion processing corresponding to the sentences of the learning document by using the word vector; and
According to the degree of the relationship between the sentence vectors, a semantic relationship classification learning between sentences is performed, a sentence-based classifier in which the semantic relationship is learned according to the semantic relationship classification learning is generated, and the intent of the input sentence based on the generated classifier. A semantic relation classification unit that performs classification processing,
The semantic relation classification unit,
According to the subject classification corresponding to the sentence vectors, the anchor, positive and negative data for each classification class are sampled, and the loss between the sampling data until the value of the Euclidean distance between each sampling data is within a certain range. Adjust the neural network weight value corresponding to the sentence vector learning model to vary the value,
When the value of the Euclidean distance between the sampling data is within a certain range, the sentence vector learning model according to the currently adjusted weight value is output as a semantic relationship learning model between sentences,
When the input sentence is included in the preset classification criterion, a classification process corresponding to the classification criterion is performed, and when the input sentence is not included in the preset classification criterion, a semantic relationship learning model between sentences is used. Based on the sentence intent classifier, index similar sentences having similar intent to the input sentence,
The semantic relation classification unit,
determining a second class from among the remaining classes except for the first class to which the anchor data and the positive data belong from among a plurality of classification classes according to the subject classification corresponding to the sentence vectors;
The second class is selected as a class having the smallest number of sampling data among the remaining classes except for the first class,
The number of sampling data for loss calculation is set based on a class having a smaller number of sampling data among the first class and the second class so that the first class and the second class are the same
Conversation service system. The method according to claim 1,
The learning device,
a learning unit configured to store learning model information generated by performing a learning model generation process of the input sentence based on the learning document of the learning database input through the input unit in the learning database;
Receive the learning document, perform semantic relation classification learning between sentences, generate a sentence-based classifier in which semantic relation is learned, and perform intention classification processing of the input sentence based on the generated classifier to output a classification processing result operating department that includes;
Conversation service system. delete delete delete

Images