KR102376570B1 - A device for classifying the semantic relationship of sentences using artificial intelligence technology - Google Patents

Abstract

The present invention provides a sentence semantic relationship classification device for providing a conversation service, comprising: an input unit that receives learning document data that is a learning target; a word-vector conversion unit that performs word vector conversion processing to map words of the learning document onto a vector space; a sentence-vector conversion unit that performs sentence vector conversion processing corresponding to sentences of the learning document using the word vector; And according to the degree of relationship between the sentence vectors, perform semantic relationship classification learning between sentences, generate a sentence-based classifier whose semantic relationship has been learned according to the semantic relationship classification learning, and classify the user input document based on the generated classifier. A sentence semantic relationship classification device for providing a conversation service including a semantic relationship classification unit that performs intent classification processing is provided.

Description

A device for classifying the semantic relationship of sentences using artificial intelligence technology}

The present invention relates to a device that classifies the semantic relationships of sentences using artificial intelligence technology. More specifically, the present invention relates to a sentence semantic relationship classification device for a sentence semantic relationship learning process for providing conversation services.

Artificial Intelligence is revolutionizing business, organizational operations, lifestyles, and communication methods. Various informatization projects are underway to provide optimal services to the lifestyles of modern culture that change rapidly every day and the constantly changing needs of customers.

Among them, big data and deep learning-related technologies have recently developed at a rapid pace, and artificial intelligence technology that is being applied to real life in certain fields has been implemented, analyzing specific data, and providing and utilizing information in various fields specialized for each individual. It is also being applied to intelligent personal services.

In order to provide various user services such as artificial intelligence and data analysis services, data mining of very large amounts of text and other language-based information is required, and machine learning methods are mainly used for this purpose.

For machine learning of these documents, a classification method using document vectors, which represent only sentences with vector values embedding words in the document, is mainly used.

However, the existing classification method has a problem in that it is difficult to determine the semantic relationship between sentences. Problems like these ultimately limit the performance of the classifier.

For example, if there are classes called 'Europe' and 'Soccer' with different topics, and there is content related to 'Soccer in Europe' in the document data called 'Europe', sentences related to the class 'Europe' and 'Soccer' and 'Soccer' classes can be said to have a semantic relationship, but the current classifier is not processing any information related to this.

In other words, because existing methods learn and classify data for each given sentence class, even if some semantic relationship appears between sentences of different classes, that is, sentences with different topics even within a document, the current method does not Semantic relationships cannot be understood.

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

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

(001) Korean Patent Publication No. 10-2006-0043333 (2006.05.15.) (Name: System and method for determining the intent of data and responding to data based on intent)

The present invention was created to solve the problems described above, and provides a conversation service that infers the relationship between sentences, determines the meaning of a new sentence, and determines the class, ultimately improving classifier performance and facilitating the interpretation of semantic relationships. The purpose is to provide a sentence semantic relationship classification device for

The sentence semantic relationship classification device according to an embodiment of the present invention to solve the problems described above is,

A sentence semantic relationship classification device for providing a data processing-based user conversation service using a semantic relationship model between sentences,

a word-vector conversion unit that receives a learning document from a learning database that stores learning model information generated by performing a learning model creation process, and performs word vector conversion processing to map words in the learning document onto a vector space; a sentence-vector conversion unit that performs sentence vector conversion processing corresponding to sentences of the learning document using the word vector; And according to the degree of relationship between the sentence vectors, perform semantic relationship classification learning between sentences, generate a sentence-based classifier whose semantic relationship has been learned according to the semantic relationship classification learning, and classify the user input document based on the generated classifier. It includes a semantic relationship classification unit that performs intent classification processing.

In addition, the classification method according to an embodiment of the present invention to solve the problems described above is,

A method of operating a sentence semantic relationship classification device for providing a data processing-based user conversation service using a semantic relationship model between sentences, comprising: receiving learning document data that is a learning target; performing word vector conversion processing to map words of the learning document into a vector space; performing sentence vector conversion processing corresponding to sentences of the learning document using the word vector; performing semantic relationship classification learning between sentences according to the degree of relationship between the sentence vectors; generating a classifier based on sentences in which semantic relationships have been learned according to the semantic relationship classification learning; and performing intent classification processing of the user input document based on the generated classifier.

Meanwhile, the method according to an embodiment of the present invention for solving the above-described problem can be implemented with a program for executing the method on a computer and a recording medium on which the program is recorded.

According to an embodiment of the present invention, words of a learning document are mapped onto a vector space, 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 classifier based on sentences in which semantic relationships are learned according to the semantic relationship classification learning can be generated.

Accordingly, according to an embodiment of the present invention, it is possible to provide a sentence-based classifier with learned semantic relationships that can determine the class by inferring the relationship between sentences to determine the meaning of a new sentence, and the resulting conversation service Through semantic relationship-based sentence classification processing, richer responses are possible, and it is possible to provide a sentence semantic relationship classification device and its operation method for providing a conversation service that improves classifier performance and facilitates semantic relationship interpretation.

In addition, the present invention provides individual classification processes for word vectorization, sentence vector, and classification processes, allowing the classification model to be variably applied even if the domain of the document changes, thereby improving system applicability and classification efficiency. There is an advantage to this.

1 is a conceptual diagram schematically showing the entire system according to an embodiment of the present invention.
Figure 2 is a block diagram for explaining in more detail a classification device according to an embodiment of the present invention.
Figure 3 is a flowchart to explain in more detail the operation method of the classification device according to an embodiment of the present invention.
Figure 4 is a diagram for explaining vector conversion processing according to an embodiment of the present invention.
Figures 5 and 6 are illustrations of a classification process and results according to an embodiment of the present invention.
Figure 7 is a flowchart for explaining the service provision process according to an embodiment of the present invention.
Figures 8 and 9 are diagrams showing result data resulting from application of convolutional neural network-based sampling and triplet loss function according to an embodiment of the present invention.

The following merely illustrates the principles of the invention. Therefore, those skilled in the art will be able to invent various devices that embody the principles of the present invention and are included in the spirit and scope of the present invention, although not explicitly described or shown herein. In addition, it is understood that all conditional terms and embodiments listed herein are, in principle, expressly intended solely for the purpose of ensuring that the concept of the invention is understood, and are not limited to the embodiments and conditions specifically listed as such. It has to be.

Additionally, it is to be understood that any detailed description reciting principles, aspects, and embodiments of the invention, as well as specific embodiments, is intended to encompass structural and functional equivalents thereof. In addition, these equivalents should be understood to include not only currently known equivalents but also equivalents developed in the future, that is, all elements invented to perform the same function regardless of structure.

Accordingly, for example, the block diagrams herein should be understood as representing a conceptual view of an example circuit embodying the principles of the invention. Similarly, all flow diagrams, state transition diagrams, pseudo-code, etc. are understood to represent various processes that can be substantially represented on a computer-readable medium and are performed by a computer or processor, whether or not the computer or processor is explicitly shown. It has to be.

Additionally, the clear use of terms such as processor, control, or similar concepts should not be construed as exclusively referring to hardware capable of executing software, and should not be construed as referring exclusively to hardware capable of executing software, including, without limitation, digital signal processor (DSP) hardware, and ROM for storing software. It should be understood as implicitly including ROM, RAM, and non-volatile memory. Other hardware for public administration may also be included.

The above-described purpose, features and advantages will become clearer through the following detailed description in conjunction with the accompanying drawings, and accordingly, those skilled in the art will be able to easily implement the technical idea of the present invention. There will be. Additionally, in describing the present invention, if it is determined that a detailed description of known technologies related to the present invention may unnecessarily obscure the gist of the present invention, the detailed description will be omitted.

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

1 is a diagram schematically showing the 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 device 100, a service providing device 200, and a user terminal 300, and the learning device 100 includes a learning database 180. ) or may be connected to the learning database 180.

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

In particular, the learning device 100 according to an embodiment of the present invention can be used to provide a conversation service of the service providing device 200, receives learning document data that is a learning target, and converts the words of the learning document into a vector space. Perform word vector conversion processing to map to, use the word vector to perform sentence vector conversion processing corresponding to the sentences of the learning document, and learn to classify semantic relationships between sentences according to the degree of relationship between the sentence vectors. , generate a classifier based on a sentence whose semantic relationship has been learned according to the semantic relationship classification learning, and perform intent classification processing of the user input document based on the generated classifier.

Furthermore, the learning device 100 performs learning processing based on a convolutional neural network between the sentence vectors to generate a sentence vector learning model, and a triplet loss between the sentence vectors corresponding to the sentence vector learning model. (Triplet-loss) function is processed, and according to the result of the triplet loss function processing, a semantic relationship learning model between sentences based on the degree of relationship between sentence vectors can be output, which infers the relationship between sentences to classify the same By learning to make the sentences closer to each other. The output layer of the learning device 100 may include a sentence vector expressing the relationship between sentences after all relationship inference learning between sentences is completed.

In addition, the service providing device 200 can provide a data processing-based user conversation service using the semantic relationship model between sentences processed in this way to the user terminal 300. The user conversation service uses the learning database 180 and the resulting sentence semantic relationship model generated by the operation of the learning device 100 according to an embodiment of the present invention to provide a conversation in response to the input sentence of the user terminal 300. May include services.

For example, the service providing device 200 may provide chatbot services, consultation services, etc. to the user terminal 300.

For example, because the user's expression method always changes, if an unclassified sentence is input, there may be limits 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, the data is not classified to utilize existing consultation data, and it takes a long time for humans to classify it. There is a problem with its use.

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

Meanwhile, the user terminal 300 and the service providing device 200 may be connected wired or wirelessly through a network, and each user terminal 100 and the service providing device 300 may be connected to an Internet network or LAN for mutual communication between networks. , data can be transmitted and received through WAN, PSTN (Public Switched Telephone Network), PSDN (Public Switched Data Network), cable TV network, WIFI, mobile communication network, and other wireless communication networks. Additionally, each user terminal 300 and the service providing device 200 may include respective communication modules for communicating using a protocol corresponding to each communication network.

In addition, the user terminal 100 described in this specification includes mobile phones, smart phones, laptop computers, digital broadcasting terminals, PDAs (Personal Digital Assistants), PMPs (Portable Multimedia Players), navigation, etc. However, the present invention is not limited to this and may be a variety of devices capable of user input and information display.

In such a system, the user terminal 300 is connected to the service providing device 200 and can receive the above-described data-based conversation service.

Figure 2 is a block diagram for explaining in more detail a learning device according to an embodiment of the present invention.

Referring to Figure 2, the learning device 100 according to an embodiment of the present invention may include an input unit 110, a learning unit 120, a classification device 150, and a learning database 180, and the classification device (150) includes a word-vector conversion unit 151, a sentence-vector conversion unit 152, a semantic relationship classification unit 153, and a result output unit 154.

First, the input unit 110 receives learning documents for learning from the learning database 180 and delivers them to the learning unit 120 to be processed for use in learning, or receives user input documents delivered from the service providing device 200. The resulting data that has been classified and processed by being transmitted to the classification device 150 can be processed to be output from the result output unit 154. The result data from the result output unit 154 may be transmitted back to the service providing device 200.

Meanwhile, the learning unit 120 may perform a learning model creation process based on the operation of the classification device 150 and the learning data input from the input unit 110. The learning model information generated in this way may be stored in the learning database 180.

In particular, the classification device 150 according to an embodiment of the present invention enables classification of semantic relationships of sentences by processing a learning model of semantic relationships between sentences in generating a learning model.

For this purpose, the classification device 150 includes a word-vector conversion unit 151 that performs word vector conversion processing to map words of the learning document onto a vector space, and sentences of the learning document using the word vector. a sentence-vector conversion unit 152 that performs sentence vector conversion processing corresponding to the sentence vectors, and performs semantic relationship classification learning between sentences according to the degree of relationship between the sentence vectors, and semantic relationships are determined according to the semantic relationship classification learning. It may include a semantic relationship classification unit 153 that generates a classifier based on the learned sentence and performs intent classification processing of the user input document based on the generated classifier.

Additionally, the result output unit 154 may be composed of one or more output modules for outputting the classification processing results.

More specifically, the semantic relationship classification unit 153 performs learning processing based on a convolutional neural network (CNN) between the sentence vectors to generate a sentence vector learning model and corresponds to the sentence vector learning model. A triplet-loss function between one sentence vector is processed, and according to the result of the triplet-loss function processing, a semantic relationship learning model between sentences is generated based on the degree of relationship between the sentence vectors.

Here, when the word-vector conversion unit 151 calculates word vector conversion values corresponding to words in the learning document, the sentence-vector conversion unit 152 performs sentence vector conversion using the word vector. You can create sentence vectors expressed as vectors. Sentence vectors expressed as word vectors in this way can be used as input data for the convolutional neural network.

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

Meanwhile, the semantic relationship classification unit 153 can obtain a desired range within a sentence by adjusting the size of the filter size of the convolutional neural network, and thus can stack layers using filters of different sizes to create a sentence It can also process progressive sentence classification from a wide range of words to phrases and sentences.

In order to apply the triplet-loss function, the semantic relationship classification unit 153 generates sentence vectors based on anchor data according to the classification class first classified according to the convolutional neural network learning. It is classified into positive data, which is the same class, and negative data, which is a different class.

In addition, the semantic relationship classification unit 153 adjusts the Euclidean distance of the positive data, which is the same class as the anchor data, to be located closer to the anchor data, and the distance to the negative data, which is a different class, to be located further away, resulting in clustering. Balanced classification of data can be processed. This triplet loss function may be a loss function that facilitates classification between data clusters.

Accordingly, the semantic relationship classification unit 153 according to an embodiment of the present invention samples anchor, positive and negative data for each classification class according to the subject classification corresponding to the sentence vectors, and samples the anchor, positive and negative data for each classification class, and The loss value between sampling data can be varied until the Euclidean distance value is within a certain range, and in order to vary the loss value, the neural network weight value corresponding to the sentence vector learning model can be adjusted. You can.

In addition, if the Euclidean distance value between the sampling data is within a certain range, the semantic relationship classification unit 153 may output a sentence vector learning model according to the currently adjusted weight value as a semantic relationship learning model between sentences.

In addition, the semantic relationship classification unit 153 may generate a sentence intention classifier using the semantic relationship learning model between sentences, and if the user input document is included in a preset classification standard, classification corresponding to the classification standard. Perform processing, and if the user input document is not included in the preset classification criteria, similar sentences with similar intent as the user input document are indexed based on a sentence intent classifier using a semantic relationship learning model between sentences. can do.

The result output unit 154 can output classification information corresponding to the indexed similar sentences, and can process the similar sentences so that they are provided to the user even if they are sentences that are not in the classification criteria.

Figure 3 is a flowchart to explain in more detail the operation method of the learning device according to an embodiment of the present invention.

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

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

Afterwards, the learning device 100 performs semantic relationship classification learning processing between sentences according to sentence vector embedding (S113) and generates a classifier using sentences for which semantic relationships have been learned (S115).

Afterwards, the learning device 100 performs intent classification processing of the user input document using a classifier (S117).

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

Referring to FIG. 4, data used in an existing classification criteria-based chatbot can be input as learning document data, and the learning document can be an example of 30 sentences written to fit the expected scenario that the user can input. As shown in FIG. 4, the learning document is stored separately into label and content and can be used as an input value for the word vectorization process based on the semantic relationship classification unit 153 of the learning unit 120.

According to the vectorization process, the word-vector conversion unit 151 converts the words of each sentence into word vectors and maps them on the vector space, and the sentence-vector conversion unit 152 acquires sentence vector data using word vector data. can do.

Afterwards, the semantic relationship classification unit 153 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 semantic relationships between sentences.

For example, as shown in FIG. 4, clustered embedding vector data can be generated by sampling, and the data in the vector space clustered as above is an embedding vector for function processing of the semantic relationship classification unit 153. It can be used as a data input value, and the semantic relationship classification unit 153 can create a semantic relationship learning model between sentences by learning the relationship between each sentence using the clustered sentences.

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

As shown in FIG. 5, when the classifier 153 is created according to the semantic relationship learning model between sentences, the classification device 150 includes a word-vector conversion unit 151 and a sentence-vector conversion unit, as described above. A conversation service based on (152) and classifier (153) can be provided.

In particular, the word-vector conversion unit 151, the sentence-vector conversion unit 152, and the classifier 153 according to an embodiment of the present invention can be created independently, so future technological developments and domain changes are possible. Even if it is, the classification model can be easily configured to change, and can be processed independently so that it can be connected to other processes. Therefore, the classification process according to an embodiment of the present invention can be easily combined with separate service processes.

Sample data used as input data may be data that the user inputs on the chatbot, or may be sentences that have not been actually learned and thus are not defined in the existing classification. Typically, in the case of existing classification methods, as shown at the top of Figure 6, the intent of the sentence may not be identified.

On the other hand, referring to FIG. 5, even if unlearned sample data is input, the classification device 150 can determine which sentence it has a semantic relationship with, and the classification result corresponding to each word and sentence is a predicted value. This can indicate that it is closest in vector space to the sentence with the highest label number 10.

Accordingly, the classification device 150 can output sentences with label 10 with the highest prediction value through the result output unit 154, which means that, as shown at the bottom of FIG. 6, Enable diverse and rich related answers.

Figure 7 is a flowchart for explaining the service provision process according to an embodiment of the present invention.

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

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

However, if the sentence is not a rule-based sentence, the learning device 100 can transmit it to the classification device 150 and process it so that a similar sentence is inferred through sentence semantic relationship inference (S209), and the learning device 100 can process the sentence corresponding to the inferred similar sentence. A response may be sent to the service providing device 200 so that the response content is output (S211).

As such, the existing rule-based chatbot service can only respond to user inquiries based on written scenarios, so there is a problem that the chatbot service cannot understand when unsaved sentences are received.

When using the learning device 100 capable of classifying sentence semantic relationships according to an embodiment of the present invention, even if a sentence that has not been learned or classified is entered, it is determined which sentences and semantic relationships are similar through the learned sentence semantic relationship classifier. This allows you to respond appropriately to the relevant content.

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

First, referring to FIG. 8, FIG. 8 is a diagram illustrating an example of the process of sampling anchor, negative, and positive data to learn a convolutional neural network.

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

Here, when C is a set of all classes existing in the data, P, which represents the number of positive data, can represent the data set size of any class included in C. Additionally, since Anchor is also the same class as P, it can represent one of the classes that C has like P. N, which represents the number of negative data, can represent the number of data of a class belonging to C that is different from P and has a smaller number of data sets than P.

In addition, the triplet loss function according to an embodiment of the present invention seeks to implement a classifier by setting the number of data samples of Positive and Anchor to N, that is, equal to Negative.

For example, referring to Figure 8, in the figure, class may represent a classification of sentences having the same topic. Each S represents a sentence as a vector of word embedding, and the number represents the identification number of the sentence. As shown in Figure 8, there are a total of three classes 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 one of the remaining classes other than class A, class B or class C, which has less data than class A. It can contain sentences of Class C, a class.

Accordingly, there is a difference in the number of data in the negative and positive sets, the distance from the anchor can be calculated, and for accurate classification, the balance according to the distance can be adjusted. The standard 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 for class C, can be the standard for the number of sample data.

Accordingly, the semantic relationship classification unit 153 can apply the triplet loss function to sample data for sentence embedding to learn the relationship between the sentences, and learn the semantic relationship of each sentence according to distance calculation and adjustment. It can be processed to create an inference learning model. Equation 2 below represents the triplet loss function application equation.

는 anchor 의 임베딩 값,

는 positive 의 임베딩 값,

는 negative 의 임베딩 값이라고 할 수 있다. In Equation 2 above, f is the embedding function, a is achor, p is positive, and n is negative. thus

is the embedding value of anchor,

is the embedding value of positive,

can be said to be the 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 relationship classification unit 153 uses the Euclidean distance to determine the difference between the distance between the anchor and the positive and the distance between the anchor and the negative.

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

Is

and

It is a hyperparameter that controls the margin to prevent 0 from becoming 0, that is, to prevent two values from being too close.

Figure 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 loss within a certain range according to the CNN (Convolutional Neural Network) and triplet loss function.

In this way, the semantic relationship classification unit 153 uses a loss function using triplet loss to reduce the loss of data of the same class, Anchor and Positive, to make the distance closer, and to increase the loss between Anchor and Negative, which are different classes. You can adjust the weight of the convolutional neural network to create Accordingly, the semantic relationship classification unit 153 can generate a vector model classifier in which sentence semantic relationships for several classes are learned.

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

A non-transitory readable medium refers to a medium that stores data semi-permanently and can be read by a device, rather than a medium that stores data for a short period of time, such as registers, caches, and memories. Specifically, the various applications or programs described above may be stored and provided on non-transitory readable media such as CD, DVD, hard disk, Blu-ray disk, USB, memory card, ROM, etc.

In addition, although preferred embodiments of the present invention have been shown and described above, the present invention is not limited to the specific embodiments described above, and the technical field to which the invention pertains without departing from the gist of the present invention as claimed in the claims. Of course, various modifications can be made by those skilled in the art, and these modifications should not be understood individually from the technical idea or perspective of the present invention.

Claims (7)

A sentence semantic relationship classification device for receiving user voice as a user input document through an input unit and providing a user conversation service based on data processing using a semantic relationship model between sentences,
A learning database that stores learning model information created by performing a learning model creation process;
a word-vector conversion unit that receives a learning document from the learning database and performs a word vector conversion process to map words of the learning document into a vector space;
a sentence-vector conversion unit that performs sentence vector conversion processing corresponding to sentences of the learning document using the word vector; and
According to the degree of relationship between the sentence vectors, semantic relationship classification learning between sentences is performed, a classifier based on sentences with learned semantic relationships is generated according to the semantic relationship classification learning, and the intention of the user input document based on the generated classifier. It includes a semantic relationship classification unit that performs classification processing,
The semantic relationship classification unit,
According to the topic classification corresponding to the sentence vectors, anchor, positive and negative data for each classification class are sampled, and loss between sampled data is applied until the Euclidean distance value between each sampled data is within a certain range. To change the value, adjust the neural network weight value corresponding to the sentence vector learning model,
If the Euclidean distance value between the sampling data is within a certain range, output a sentence vector learning model according to the currently adjusted weight value as a semantic relationship learning model between sentences,
If the user input document is included in the preset classification criteria, classification processing corresponding to the classification criteria is performed, and if the user input document is not included in the preset classification criteria, a semantic relationship learning model between the sentences is performed. Based on a sentence intent classifier using, similar sentences with similar intent as the user input document are indexed,
The semantic relationship classification unit,
Determining a second class among the remaining classes excluding the first class to which the anchor data and the positive data belong among a plurality of classification classes according to subject classification corresponding to the sentence vectors,
The second class is selected as the class with the smallest number of sampling data among the remaining classes excluding the first class,
The number of sampling data for loss calculation is set based on the class with 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.
Sentence semantic relationship classification device for providing conversation services. According to paragraph 1,
The semantic relationship classification unit,
Create a sentence vector learning model by performing learning processing based on a convolutional neural network between the sentence vectors, and process a triplet-loss function between the sentence vectors corresponding to the sentence vector learning model. , according to the results of the triplet loss function processing, to generate a learning model of semantic relationships between sentences based on the degree of relationship between the sentence vectors.
Sentence semantic relationship classification device for providing conversation services. delete delete delete delete According to paragraph 1,
The semantic relationship classification unit
Further comprising a result output unit that outputs classification information corresponding to the indexed similar sentences.
Sentence semantic relationship classification device for providing conversation services.

Images