KR101834564B1 - Domain matching device and method for multi-domain natural language processing - Google Patents

Abstract

The present invention relates to a domain matching apparatus using specific scores of a chatbot module, a search engine module or a speech recognition module and a method thereof. To this end, the domain matching apparatus comprises: a domain score calculation module calculating a domain score which is information about which a domain pipe line is suitable to match phrase information; and a domain matching module matching the phrase information to the domain pipe line related to a domain having a highest domain score.

Description

FIELD OF THE INVENTION [0001] The present invention relates to a domain matching device and method for domain matching using domain-

Field of the Invention [0002] The present invention relates to an apparatus and method for domain matching using a domain-specific score for multi-domain natural language processing of a chatbod module, a search engine module or a speech recognition module.

Human languages are rich and complex, and contain many vocabularies with complex grammatical and contextual meanings, but hardware or software applications generally require data to be entered according to a particular format or rule. Here, natural language input can be used in almost all software applications for interacting with people. A typical natural language processing process separates the natural language into tokens, maps to one or more operations provided by the software application, and each software application is set to have a set of unique operations. That is, software developers write code that interprets natural language input and maps the input to the appropriate behavior for each application. Such a natural language processing apparatus is used in a speech recognition engine, a search engine, a sentence recognition engine, and the like.

A typical example of such a natural language processing apparatus is a chatbot. Recently, there have been a lot of companies trying to solve the problem by improving the quality of the customer support service by introducing the chatbot. For example, AT & T calls the CS to let the bot respond.

1 is a schematic diagram showing a basic module configuration of a chatbot. As shown in FIG. 1, the chatbot receives a lexical input composed of a text or voice of a chat in a form that can be processed on a computer by a NLP (Natural Language Processor) module, In response to the Context Analyzer, which analyzes the Context of the natural language, the Decision Maker decides the answer contents according to the Context, classifies the answers and confirms them, and responds to the user according to the determined answers. output.

FIG. 2 is a schematic diagram specifically showing a general chatbot module configuration. As shown in FIG. 2, a general chatbot analyzes a morpheme of a linguistic input inputted with a Morphemic Analyzer attached to a NLP, attaches a Taxonomy Analyzer to a semantically meaningful word, and processes the sentence on a computer It is processed in a form that can be done. The Context Analyzer will match the context in the Context Database through the search engine. The Decision Maker will find the answer to the matching context. Usually, the answer to a specific context is determined by the customer support protocol data template of the particular company that is trying to introduce the chatbot.

FIG. 3 is a schematic diagram specifically showing the configuration of a chatbot module using machine learning. As shown in FIG. 3, a chatbot using machine learning has replaced a lot of elements of a conventional chatbot. Deep-learning model is applied to NLP in the chatbot module configuration using machine learning. Deep learning models applied to NLP can be understood as an early version of NLU (Natural Language Understanding), which is available from Google's SyntexNet published in April 2016. Through SyntexNet, certain verbal input changes to Vector format. We then integrate the Context Analyzer, Decision Maker, and Response Generator to apply the Deep Learning model and continue learning the training set. At least one model such as RNN (Recurrent Neural Network) is used as the deep learning model used in the integration module of Context Analyzer, Decision Maker and Response Generator.

In addition to chatbots, various services based on natural language processing and response generation are being provided. In addition to chatbots, services based on natural language processing and response generation include speech recognition and search engines.

Korean Patent Publication No. 10-2017-0021902,

In the past, there have been the following problems in designing natural language processing devices such as chatbots, search engines, and voice recognition modules.

First, since it is very difficult to obtain a refined data set for natural language learning, a preprocessing process for refining an uncleaned data set is required. In order to facilitate data supply and preprocessing, it is the easiest and most accurate way to design a chatbot after collecting the data of a specific domain intensively. This is because the narrower the range of questions and answers, the better the performance of natural language processing, as the meaning of the terms used for each domain is different.

When a natural language processing device such as a chatbot that intensively learns data of a specific domain is created, it becomes a natural language processing device capable of only talking about the domain. A chatbot, a search engine, a speech recognition module, and the like, which are made of such a natural language processing device, can work well in a single domain but have a disadvantage that they are not versatile.

In order to solve the above disadvantage, that is, to solve the versatility, if the data sets spanning several domains are collected and machine learning is performed on the natural language processing device, the problem that the output of the natural language processing device for input is derived from the undesired domain, There arises a problem that it is difficult to process a dependent task.

SUMMARY OF THE INVENTION Accordingly, it is an object of the present invention to design a plurality of response modules specific to a specific domain in order to solve the above problem, and to provide input data (for example, input speech data, input sentence data, query data, A domain matching apparatus and method for multi-domain natural language processing related to a method of classifying input data into one of a plurality of domains and assigning input data to a domain-specific response module suited to the classified domain, .

It is another object of the present invention to provide a multi-domain natural language processing method for selecting output data of an appropriate domain for output data such as answer data generated in a plurality of response modules specific to a specific domain in order to solve the above- And to provide a domain matching apparatus and method for the same.

Another object of the present invention is to provide a method and apparatus for multi-domain natural language processing for a method of assigning weights to selection of domains of a current dialogue or query according to which domain is classified in a previous dialogue or query in order to solve the above- Domain matching apparatus and method.

Another object of the present invention is to solve the above problems by using a transition probability which is a probability of transition from a specific domain to another domain, And a domain matching apparatus and method for multi-domain natural language processing on a method of assigning weights to selection.

It is another object of the present invention to provide a method and apparatus for classifying a domain for a specific conversation or query by a specific classification method, outputting output data, and analyzing user response data according to output data And a domain matching apparatus and method using a domain-specific score for a method of performing reinforcement learning on a classification model of a corresponding domain.

Hereinafter, specific means for achieving the object of the present invention will be described.

An object of the present invention is to provide a classifying module that receives input data in a general string form and generates phrase information composed of a word or word combination by dividing the input data into entities or intents. And a plurality of domain pipelines for a plurality of domains for generating output data for a specific domain on the basis of the phrase information and the set of question data. The domain matching method of the multi- A domain score calculation module for calculating a domain score which is information on which domain pipeline is suitable to match the phrase information; And a domain matching module for matching the phrase information to the domain pipeline associated with the domain having the highest domain score, wherein the calculation of the domain score is based on a degree of similarity between the phrase information and the query data set And a domain pipeline of one of the plurality of domain pipelines matched with the phrase information is configured to generate the output data, wherein a domain matching apparatus for multi-domain natural language processing is provided .

Further comprising a weight module for generating a weight and providing the weight to the domain score calculation of the domain score calculation module, wherein the calculation of the domain score is performed based on the similarity and the weight, May be characterized by being information on a previous domain score.

Further comprising a weight module for generating a weight and providing the weight to the domain score calculation of the domain score calculation module, wherein the calculation of the domain score is performed based on the similarity and the weight, Is the same domain section information immediately before.

Further comprising a weight module for generating a weight and providing the weight to the domain score calculation of the domain score calculation module, wherein the calculation of the domain score is performed based on the similarity and the weight, Is a transition probability to another domain.

In addition, the calculation of the domain score may be performed by using the word information as a state, the domain matching apparatus as an agent, the matching of the domain pipeline as an action, And the information is updated by reinforcement learning that makes the information reward.

And a weight module for generating weights that are information about previous domain scores and providing the weights to the domain score calculation of the domain score calculation module, wherein the calculation of the domain scores comprises: Wherein the generation of the weights of the weighting module is performed by using the word information as a state, the weighting module as an agent, the generation of the weighting as an action, new input data according to the output data Is rewritten by the reinforcement learning that rewrites the information on the information to be rewritten.

It is another object of the present invention to provide a classifying module for inputting input data in a general string form and generating phrase information composed of words or word combinations by dividing the input data into entities or intents; And a plurality of domain pipelines for a plurality of domains for generating output data for a specific domain on the basis of the phrase information and the set of question data. The domain matching method of the multi- Wherein a domain score calculation module that is a component of the domain matching device calculates a domain score which is information on which domain pipeline is suitable to match the phrase information to the domain matching device, A domain score calculation step; And a domain matching step in which the domain matching module, which is a component of the domain matching apparatus, matches the phrase information to the domain pipeline related to the domain having the highest domain score, Wherein the plurality of domain pipelines are constructed on the basis of similarity between the query information and the query data set, and one of the plurality of domain pipelines matched with the query information is configured to generate the output data. And a domain matching method for natural language processing of the domain name.

It is another object of the present invention to provide a classifying module for inputting input data in a general string form and generating phrase information composed of words or word combinations by dividing the input data into entities or intents; A plurality of domain pipelines for a plurality of domains for generating output data for a specific domain based on the set of query information and query data; A domain score calculation module for calculating a domain score, which is information on which domain pipeline is suitable to match the phrase information; And a domain matching module for matching the phrase information to the domain pipeline associated with the domain having the highest domain score, wherein the calculation of the domain score is based on a degree of similarity between the phrase information and the query data set And a domain pipeline of one of the plurality of domain pipelines matched with the phrase information is configured to generate the output data, wherein the natural language processing apparatus for performing multi-domain natural language processing is provided Can be achieved.

Another object of the present invention is to provide a data processing apparatus, comprising: a first segmenting module for receiving input data and generating the phrase information of the input data; A second classification module for receiving a query data set of a specific domain pipeline that generates output data for a specific domain and generating phrase information of the query data set; A domain score calculation module for calculating a domain score for each domain pipeline using the phrase information of the input data and the phrase information of the query data set of the specific domain pipeline; And a domain matching module for matching the input information of the input data with the domain pipeline having the highest domain score, and inputting the matching information. The domain matching device for multi-domain natural language processing can be achieved.

It is another object of the present invention to provide a data processing apparatus, comprising: a first segmenting step in which a first segmenting module receives input data and generates a phrase information of the input data; A second segmentation module receiving a query data set of a specific domain pipeline that generates output data for a specific domain and generating the query information of the query data set; A domain score calculation step of calculating a domain score for each domain pipeline using the phrase information of the input data and the phrase information of the query data set of the specific domain pipeline; And a domain matching step in which the domain matching module matches the input information of the input data to the domain pipeline having the highest domain score and inputs the same to the domain pipeline. Can be achieved.

It is another object of the present invention to provide a data processing apparatus, comprising: a first segmenting step in which a first segmenting module receives input data and generates a phrase information of the input data; A second segmentation module receiving a query data set of a specific domain pipeline that generates output data for a specific domain and generating the query information of the query data set; A domain score calculation step of calculating a domain score for each domain pipeline using the phrase information of the input data and the phrase information of the query data set of the specific domain pipeline; And a domain matching step in which the domain matching module matches the word information of the input data to the domain pipeline having the highest domain score and inputs the word information to the domain pipeline. And a program stored in a recording medium to be executed.

As described above, the present invention has the following effects.

First, according to one embodiment of the present invention, the effect of enabling generation of a response specific to each specific domain occurs.

Second, according to an embodiment of the present invention, input data specialized for each specific domain is accumulated, and intensive learning of a response module is enabled for each domain.

Thirdly, according to the domain matching apparatus according to an embodiment of the present invention, the calculation amount of the task of receiving input data and assigning it to each specific domain pipeline is very small and the processing speed is improved.

Fourthly, according to the weight module according to an embodiment of the present invention, the accuracy is gradually improved even though the calculation amount of the task of receiving input data and allocating it to each specific domain pipeline is very small.

BRIEF DESCRIPTION OF THE DRAWINGS The accompanying drawings, which are included to provide a further understanding of the invention and are incorporated in and constitute a part of this application, illustrate preferred embodiments of the invention and, together with the description, And shall not be interpreted.
1 is a schematic diagram showing a basic module configuration of a chatbox,
FIG. 2 is a schematic diagram specifically showing a general chatbot module configuration,
FIG. 3 is a schematic diagram specifically showing the configuration of a chatbot module using machine learning,
4 is a schematic diagram illustrating a relationship between a domain matching apparatus and a response module according to an embodiment of the present invention;
5 is a schematic diagram showing a signal flow of the natural language processing apparatus according to the first extended example of the present invention;
6 is a schematic diagram showing a signal flow of the natural language processing apparatus according to the fourth extended example of the present invention;
FIG. 7 is a schematic diagram showing reinforcement learning of a natural language processing apparatus according to a fourth extended example of the present invention;
8 is a schematic diagram showing a state flow of reinforcement learning of a natural language processing apparatus according to a fourth extended example of the present invention;
9 is a schematic diagram showing a signal flow of reinforcement learning in the natural language processing apparatus according to the fourth extended example of the present invention,
10 is a graph showing a Q function relating to the predicted value of reinforcement learning of the natural language processing apparatus according to the fifth extended example of the present invention,
11 and 12 are diagrams showing reinforcement learning using a neural network according to a seventh extended example of the present invention,
13 shows a network structure of reinforcement learning using a neural network of a natural language processing apparatus according to an eighth extended example of the present invention,
FIG. 14 shows a signal flow of a natural language processing apparatus according to a modification of the present invention,
15 and 16 are schematic diagrams showing a signal flow of reinforcement learning of a natural language processing apparatus according to a modification of the present invention;
FIG. 17 is a flowchart illustrating a domain matching method of a natural language processing apparatus for multi-domain natural language processing according to an embodiment of the present invention;
FIG. 18 illustrates an example of results of each step of the domain matching method according to an embodiment of the present invention,
FIG. 19 is a flowchart illustrating a domain matching method according to a first extended example of the present invention;
20 is a flowchart illustrating a domain matching method according to a second extended example of the present invention,
FIG. 21 is a flowchart illustrating a domain matching method according to a third extended example of the present invention;
FIG. 22 is a flowchart illustrating a domain matching method according to a fourth extended example of the present invention;
23 is a flowchart showing another example of a domain matching method according to a fourth extended example of the present invention;
24 is a schematic diagram showing a signal flow of a natural language processing apparatus according to a modification of the present invention.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings. In the following detailed description of the operation principle of the preferred embodiment of the present invention, a detailed description of known functions and configurations incorporated herein will be omitted when it may unnecessarily obscure the subject matter of the present invention.

The same reference numerals are used for portions having similar functions and functions throughout the drawings. In the specification, when a specific portion is connected to another portion, it includes not only a direct connection but also a case where the other portion is indirectly connected with another element in between. In addition, the inclusion of a specific constituent element does not exclude other constituent elements unless specifically stated otherwise, but may include other constituent elements.

Domain matching device for multi-domain natural language processing

4 is a schematic diagram illustrating a relationship between a domain matching apparatus and a response module according to an embodiment of the present invention. As shown in FIG. 4, input data 200 composed of a question or a search query is received by the classification module 2, which is one component of the multi-domain natural language processing device 1, Device 3 to match the particular domain pipeline 40 of the response module 4 to produce output data.

The classification module 2 receives the input data 200 in a normal text form and divides the general character string type into an entity and an intent by the NLP module 21 . The input data may be divided into words or phrases that are combinations of words by the classification module (2). The NLP module 21 may specifically include morphological analysis, stem extraction, abstraction extraction, TF, TFIDF, and the like. Thereafter, the objects and semantic phrases separated by the vectorization module 22 (Sentence2vec, Word2vec, SyntexNet) are processed as vector values. Or Word2vec may be used, and specifically, an n-gram, a CBOW model for predicting a word from a context, or a Skip-gram model for predicting a context from a word may be used. That is, the classification module 2 may generate the phrase information including the entity and the semantic phrase, and the phrase information by the vectorization module 22 may be expressed by the vector value. According to an embodiment of the present invention, the vectorization module 22 can be selectively configured only when performing the similarity determination in the domain matching device 3. [

The domain matching apparatus 3 generates a domain score by calculating the degree of similarity with the query data set of each specific domain pipeline 40 based on the vectorized phrase information received from the classification module 2, The input data is matched to the domain having the highest domain score. At this time, similarity calculation using General Corpus can be used. This is called Corpus-used Similarity. For example, in the case of Korean, the National Corpus of General Corpus can be used, and English can use Brown Corpus. Alternatively, a similarity measure using a semantic network can be used. This is called Knowledge-based Similarity. For example, wordNet. The similarity measure between word and word using wordNet is based on the similarity of path, Leacock & Chodorow Similarity, which utilizes minimum distance and maximum depth of node, Wu & Palmer Similarity Or the like.

Also, the similarity calculation method of the domain matching apparatus 3 according to another embodiment of the present invention may use a k-nearest neighbors algorithm, a Naive Bayes classifier, or a support vector machine. Also, for example, according to an embodiment of the present invention, the degree of similarity can be calculated by the following equation.

In the above equation (1), Score is a domain score as a probability score that a vector x belongs to a D _k domain, D _k is a kth domain, x is a vector composed of n elements, x _i is an i th element of x, P _k ) and P (x) may denote a scaling factor. According to the method of Equation (1) according to an embodiment of the present invention, the input vector x is classified into the domain D _k having the highest degree of similarity P (D _k | x). According to the method using the probability score that the vector x belongs to the D _k domain as described above, it is possible to quickly match the domain even with a small amount of calculation.

In addition, the domain matching apparatus 3 according to another embodiment of the present invention may further include a query data set of a specific domain pipeline 40 to determine whether the phrases classified in the classification module 2 are characteristic to a particular domain (Or probability of appearance of a word) of the input data 200 of the specific domain pipeline 40 and the query appearance frequency of the input data 200 of the specific domain pipeline 40 are compared with each other, In this way, a domain-specific score, which is a score of how characteristic phrases of the input data are to the domain, is determined and reflected in the domain score, and input data 200 to a specific domain having the highest domain score among the respective domains Can be matched.

In addition, the domain matching apparatus 3 according to another embodiment of the present invention may further include a query data set of a specific domain pipeline 40 to determine whether the phrases classified in the classification module 2 are characteristic to a particular domain (Or probability of occurrence of a word) of the input data 200 of the general corpus is compared with the frequency of occurrence of a word (or probability of occurrence of a word of the input data 200 of the general corpus) A domain specific score, which is a score of how much the input word of the input data is in the domain, is reflected in the domain score, and the input data 200 is assigned to the specific domain having the highest domain score among the respective domains Can be matched. (Or probabilities of occurrence of words) in the general corpus by checking the left and right words of each word in the general corpus, and comparing the frequency of occurrence of the words in the question data set The probability of appearance of the word) is higher, the corresponding phrase of the input data 200 can be determined to be characteristic to the corresponding domain. Depending on the amount of data in the general corpus, we can increase or decrease the distance of the correlation, which will be the basis for determining the frequency of occurrence of the word, to i + n, i-n. Here, i is the index of the word in the sentence. The amount of data and the constant n can be proportional.

delete

According to the domain matching apparatus according to the embodiment of the present invention, the amount of calculation of the task of receiving input data and assigning it to each specific domain pipeline is very small and the processing speed is improved.

According to another embodiment of the present invention, the classifying module is composed of a first classifying module and a second classifying module, and the first classifying module can be configured to receive input data and generate the phrase information of the input data. The second segmentation module may be configured to receive a query data set of a particular domain pipeline that generates output data for a particular domain, and to generate the query information of the query data set. At this time, the domain score calculation module may be configured to calculate a domain score for each domain pipeline using the input information of the input data and the query information of the query data set of the specific domain pipeline.

The response module 4 has a plurality of specific domain pipelines 40 and assigns input data to the specific domain pipeline 40 for the domain matched by the domain matching device 3 to generate a response .

The specific domain pipeline 40 includes a Context Analyzer, a Decision Maker, a Response Generator, and the like, which are modules other than the classification module in the existing natural language response module as shown in FIGS. 1 to 4 Lt; RTI ID = 0.0 > domain < / RTI >

[ First extended example ]

With respect to the first extension example of the domain matching apparatus for multi-domain natural language processing according to an embodiment of the present invention, the domain matching apparatus can apply the weight for matching of the domain to the calculation of the domain score, And can be linked to a weight module that updates the time thermally.

5 is a schematic diagram showing a signal flow of the natural language processing apparatus according to the first extended example of the present invention. As shown in FIG. 5, the natural language processing apparatus according to the first extended example of the present invention may further include a weight module 5 between the natural language processing apparatus at time t and the natural language processing apparatus at time t + 1.

The weighting module 5 calculates information on the domain score or the specific domain calculated in the domain matching apparatus 3 at time t as a weight for the matching of the domain score or the specific domain in the domain matching apparatus 3 at time t + It is a module to apply.

As the method of calculating the weight, the domain score of the domain matched with the input data inputted at the time t affects the domain score of the input data at the time t + 1, and as the input data is continuously input in the near time, The weighting of these quantities is reduced as time goes by and the domain scores of the other specific domains are higher and they are constructed in such a way as to give negative weights when they are to be matched to different domains . That is, the weight may mean the probability that the domain will be maintained according to the time difference in which the input data is input.

Alternatively, with another form of weighting calculation method, it may mean how the domain score was previously obtained for certain phrase information, or which domain was previously matched with which phrase information.

According to a weight module that updates these weights in a time-wise manner, for example, "How much?" It is possible to finely perform domain matching on ambiguous input data.

An extended example of the weight calculation method according to an embodiment of the present invention is as follows.

[ Second Extended Example - Memory type]

In a situation where a series of input data is input, the weight module 5 according to the second extended example of the present invention may be configured such that the domain matching device 3 assigns a weight to the domain matched with the input data inputted in advance have. X (i-1), x (i-2), x (i-3) ... when the current input data, that is, the i- If all of the domains to which x (iM) are assigned are the same as D _k , it can be seen that M input data is continuous in one domain. In this case, i-1, i-2, i-3, ..., iM may be defined as 'immediately before the same domain period'. Also, D _{k, which} is a matched domain in the immediately preceding same domain period, can be defined as a 'matching domain'.

Accordingly, the domain matching apparatus 3 _{multiplies the} weight W (D _k , x (i)) by P (D _k | x) calculated by the similarity method to calculate the similarity of the domain P Matching can be maintained. In the second extension example of the present invention, the weight module 5 may generate a weight to be applied to the domain score at time t based on the domain score of the immediately preceding same domain period.

More specifically, it is expressed by the following equation.

Equation (3) describes a weight W (D _k , x) expressed as a function of M when M input data are included in D _k consecutively or not.

In Equations (2) and (3), Score is a domain score as a probability score that a vector x belongs to a D _k domain, D _k is a k th domain, x is a vector composed of n elements, x _i is an i th element of x, P _k ) and P (x) may mean a scaling factor or a probability distribution. (I-1), x (i-2), x (i-3) ... If all the domains to which x (iM) are assigned are the same as D _k , it can be considered that M input data is continuous in one domain. Thus, weights W (D _k , x The domain matching can be maintained by multiplying the calculated P (D _k | x) by calculating the domain score.

Alternatively, the discontinuous differential value of the similarity P (D _k | x) may be calculated, and the weight W (D _k , x (i)) for the recent M input data may be obtained.

In the above equation, N may mean the number of times the input data is finally input. The W (D _k ) aspect of each domain can be changed as follows.

The time-wise aspect of W (D _k ), which is the weight of each domain according to the second expansion example

In another example of the second extended example of the present invention, the weight is a measure of the probability of belonging to the domain of the immediately preceding same domain period. This can be learned based on time series sequence data in the same way as Recurrent Neural Network. For example, the weight in the second extended example of the present invention may mean the probability that input data B is sequentially input after input data A is input in a specific domain.

A schematic diagram of the RNN scheme according to another example of the second extended example

In Table 2 above, wrec can mean the feedback weights in the RNN. In Table 2 above, the S function can refer to a specific domain pipeline. In this case, it is possible to predict which word information will be input as input data at time t by the output data y (t) of the RNN. This input data prediction information can be used as the weight in the second extended example of the present invention.

When applying this method, for example, when input data classified as a chicken delivery domain is processed (eg, how much is chicken / golden olive coupon applies), input data (eg, what time is it / how much?) is input, it is possible to classify it into a domain that has been processed previously and converge it into a conversation of consistent subject.

Further, in matching the domain pipeline through the input data, the effect of reducing the amount of computation is generated. When the previous input vector (vectorized input data) x (i-1) belongs to D _k , the product of P (D _k | x (i)) and W (D _k , x examples the effects that can be configured to immediately assign the x (i-1) to D _k, without having to calculate the degree of similarity with other domains except for D _k greater than 50%) is generated.

[ Third Expansion Example - transition probability method]

In the case where a series of input data comes in, and a situation is to be transferred from the currently matched domain pipeline to another domain pipeline, the weight module 5 according to the third extended example of the present invention is a module And to assign a weight to the domain matching apparatus 3 based on the probability that the data will be transferred from the matched domain to another domain. More specifically, the weight module 5 according to the third extended example of the present invention, when a series of input data is input and input data previously input is assigned to one domain, A probability distribution table (domain transition probability distribution) that can be transferred from a domain to a domain in a domain set, and applying this to calculate the probability to be assigned to each domain. For example, it can be explained as follows.

Schematic example of domain transition probability distribution according to the third extension example

Domain transition probability distribution table D1 (game) D2 (movie) D3 (chicken) ... D1 (game) 60% 20% 4% D2 (movie) 15% 75% 5% D3 (chicken) 3% 2% 80% ...

Above equation Score domain score, P (D _k | x) is a weight, P probability that the input data x belong to the _{D k, W (D k,} x (i)) (D k | D j) is D _k To D _j .

In this way, the probability that a next input data can be classified into a domain having a high probability of transition from the domain of the previous input data is calculated based on the domain transition probability, which is a probability of transition between the domains at the time of domain conversion. Lt; / RTI > to determine the domain score. For example, the probability that a domain will be transferred to an entertainer story when sharing a movie story is high, but it can be applied to the calculation of a domain score that the probability that a domain is transferred to a story of a chicken after a story of a movie is low. The domain transition probability can be calculated based on data on past domain transitions and can be changed according to the input time difference of input data.

Also, at the time when the domain moves from the existing D _k to the next domain D _j , it is necessary to first calculate the domains with high probability of domain transition with D _k without calculating the similarity with all domains except D _k in order to calculate D _j The amount of computation can be reduced.

[ Fourth Extended Example - Domain reinforcement learning]

In a situation where a series of input data is input, the weight module 5 according to the fourth extended example of the present invention is configured to assign a weight based on the probability of transition from the matching domain to the other domain At the same time, reinforcement learning may be performed on the weight, so that the weight is updated according to the user's response. The compensation to be used in the reinforcement learning includes the user's reaction to the output data 201 of the time t output by the matching of the specific domain pipeline 40 of the domain matching apparatus 3, ) Of the input data 200 at the time t + 1 that is input again after the user has received the input data 200 may be used.

6 is a schematic diagram showing a signal flow of a natural language processing apparatus according to a fourth extended example of the present invention. As shown in FIG. 6, the natural language processing apparatus according to the fourth extended example of the present invention includes a weight module 5 and a sentiment analysis module 5, which are configured between a natural language processing apparatus at time t and a natural language processing apparatus at time t + 6). The weighting module 5 applies the domain score calculated in the domain matching apparatus 3 at time t or the information about the specific domain to the domain score in the domain matching apparatus 3 at time t + As the module, it can be configured to perform the same role as the weight module 5 of the first, second and third extension examples. However, the weight of the weight module 5 in the fourth extended example may be configured to be updated by the emotion analysis value of the sentiment analysis module 6. The sentiment analysis module 6 is connected to the classification module 2 at time t + 1 and performs sentiment analysis of the input data 200 at time t + 1 using the delimited phrases of the input data 200 and the sentiment dictionary Value, and provide the generated emotion analysis value to the weight module 5.

7 is a schematic diagram showing reinforcement learning of a natural language processing apparatus according to a fourth extended example of the present invention. As shown in FIG. 7, the natural language processing apparatus according to the fourth extended example of the present invention uses the weight module 5, which is configured between the natural language processing apparatus at time t and the natural language processing apparatus at time t + 1, as Agent, The weights generated in the weighting module 5 are used as actions of the Agent in accordance with the degree of similarity and weight between the query data set included in each specific domain pipeline 40 of the domain matching apparatus 3 and the input data 200, The sentiment analysis value generated by the sentiment analysis module 6 by the sentiment analysis module 6 is expressed as Reward by the input data 200 of the time t and the input data 200 of the time t + A reinforcement learning algorithm can be constructed.

The existing reinforcement learning is mainly used in a domain such as a game, and it is very rare that the reinforcement learning is used in a natural language processing apparatus as in the fourth extension example of the present invention. In addition, even if the reinforcement learning is applied in the conventional natural language processing, it is hard to find an Agent, Action, State, Reward, and Environment set as in the fourth extension example of the present invention. If the weight of the weight module 5 is set to Action and the input data 200 of the time t is configured to State as in the fourth extension example of the present invention, There is an effect that the probability that a natural language processing apparatus having a pipeline will gradually and specifically match a specific domain pipeline becomes higher as the weights are dynamically updated according to the phrase configuration of the specific input data 200. [ In addition, since the weighting of the weight module 5 is determined as Action and the sensitivity analysis value of the input data 200 at time t + 1 is configured as Reward, the specific domain pipeline matching the intention of the user can be matched An effect is generated.

8 is a schematic diagram showing a state flow of reinforcement learning of the natural language processing apparatus according to the fourth extended example of the present invention. As shown in FIG. 8, in the reinforcement learning of the natural language processing apparatus according to the fourth extended example of the present invention, states of State, Action, and Reward can be continuously performed. Reinforcement learning of this natural language processing apparatus may be continued to this signaling stage or may be configured to include a terminal state S _n as shown in FIG. 8.

9 is a schematic diagram showing a signal flow of reinforcement learning of the natural language processing apparatus according to the fourth extended example of the present invention. 9, reinforcement learning of the natural language processing apparatus according to the fourth extended example of the present invention is performed by assigning a weight module 5 configured between the natural language processing apparatus at time t and the natural language processing apparatus at time t + The input data 200 of the time t to the state ₀ and the input data 200 of the time t + 1 to the sentiment analyzing module 6 as the action ₀ of Agent, the weight generated by the weight module 5, A reinforcement learning algorithm can be constructed in which the sensibility analysis value generated by emotion analysis is Reward ₁ . Then, the input data 200 at time t + 1 can be restarted to State ₁ to start the reinforcement learning cycle.

According to the fourth extended example of the present invention, it is possible to match a specific domain pipeline that best matches each input data by using reinforcement learning, which has been difficult to apply to a natural language processing apparatus.

[ Fifth Extended Example Reinforcement learning using -Q]

The reinforcement learning of the natural language processing apparatus according to the fourth extended example of the present invention may cause a problem that the reward is sparse. The problem of Reward being lean is that in the course of training in a better way to choose which action (weight) to take when given a certain state (input data) And there is a risk that the training will not be performed because the input data at time t + 1 appears after inputting.

Therefore, in order to solve such a problem, it can be configured to perform reinforcement learning using the predicted value Q. FIG. 10 shows a Q function relating to the predicted value of reinforcement learning of the natural language processing apparatus according to the fifth extended example of the present invention. 10, in the reinforcement learning using the Q function of the natural language processing apparatus according to the fifth extended example of the present invention, the input data 200 is set to the State, and the weight set by the weight module 5, which is the agent, It is possible to use a function Q (state, action) for outputting Reward Quality expected as an Action. Using the Q function, it is possible to reinforce learning by using the predicted value of Reward before performing a specific action.

In Equation (6), Q denotes a predicted value for Reward, Max Q denotes a maximum value Q can have, s denotes a state at time t, and a 'denotes an action at time t + 1. "" Means time t + 1. For the action giving the maximum value of Q, the following optimal policy can be proposed.

In Equation (7), π ^* can mean an Optimal Policy for Q. At this time, the following method may be proposed to express Q (s, a) at time t as Q (s ', a') at time t + 1.

When the sum of all future rewards which can be received by the weighting module 5 as an agent is defined as R, and the final state is defined as n, R can be defined as Equation 8 above. Expressing R as a variable for time t, it can be calculated as follows.

Also, Expression (9) can be expressed as a variable for time t + 1 as follows.

Therefore, R _t can be expressed as follows.

The Optimal value for this R _t can be roughly expressed as follows.

In Equation (12), "*" may mean Optimal, and max may mean a maximum value.

Therefore, according to Equation (12), the Q function can be expressed as Equation (12) as follows.

In Equation 13, Q hat is the predicted value of Q, s, a is state and action of time t, r is reward of time t, s ', a' is state and action of time t + 1 . According to Equation (13), the predicted value for the current Q value can be updated using the maximum value of the predicted value of the Q value of the current reward and t + 1. Therefore, in the reinforcement learning of the natural language processing apparatus according to the fourth extended example of the present invention, the Q value can be reinforcement learned as shown in Equation (13).

With respect to the concrete algorithm, the following table shows a reinforcement learning algorithm using the Q value according to the fifth extended example of the present invention.

Observe current state s
Do forever:
-Select an action a and execute it
-Receive immediate reward r
-Observe the new state s'
-Update the table entry for

as follows:

-s <- s'

Algorithm
For each s, a initialize table entry

Observe current state s
Do forever:
-Select an action a and execute it
-Receive immediate reward r
-Observe the new state s'
-Update the table entry for

as follows:

-s <- s>

The reinforcement learning algorithm using the Q value in the reinforcement learning of the natural language processing apparatus according to the fifth extended example of the present invention

[ Example 6 extended - reinforcement learning using learning rate]

Reinforcement learning according to the fourth and fifth extended examples of the present invention may cause a difficulty in learning because it is likely that the weighting of the action and the matching of a specific domain pipeline are not actually the same. For example, when an action is determined such that a high weight is given to the domain for the 'weather' with respect to the phrase information of the specific input data, the domain matching apparatus 3 searches the input information of the input data for each specific domain pipeline The similarity degree with respect to the query data set included is considered together with the weight, so that the domain for 'weather' can be configured to be stochastic.

If the action is applied to the environment in such a stochastic manner, there is a possibility that learning can not be performed properly. This is because, if a certain action is performed stochastically according to the selection of a specific action, a proper feedback occurs stochastically.

Accordingly, in the sixth extended example of the present invention, a learning rate can be further applied to solve this problem. In updating the Q (s, a) of the time t, the learning module is applied to learn the weight module 5 with a specific ratio of 0 to 1 for Q (s') at time t + (s) are also considered together at a certain rate.

The update of Q (s, a) to which the learning rate is applied can be performed according to the following equation.

In Equation (14), k may mean a learning rate.

According to Equation (14) above, the reinforcement learning algorithm of the natural language processing apparatus according to the sixth extended example of the present invention can be roughly configured as follows.

Observe current state s
Do forever:
-Select an action a and execute it
-Receive immediate reward r
-Observe the new state s'
-Update the table entry for

as follows:

-s <- s'

Algorithm
For each s, a initialize table entry

Observe current state s
Do forever:
-Select an action a and execute it
-Receive immediate reward r
-Observe the new state s'
-Update the table entry for

as follows:

-s <- s>

The reinforcement learning algorithm using the learning rate k in the reinforcement learning of the natural language processing apparatus according to the sixth extended example of the present invention

If the learning rate is applied according to the sixth extended example of the present invention as described above, reinforcement learning can be performed well even if the action is performed stochastically.

[ Example 7 expansion - Neural  Reinforcement learning using network 1]

The reinforcement learning of the natural language processing apparatus according to the fourth, fifth and sixth extended examples of the present invention may cause a problem that the degree of freedom of the input data is high or the number of cases increases, which makes learning difficult.

Therefore, in order to solve such a problem, reinforcement learning of the natural language processing apparatus according to the seventh extended example of the present invention can use a neural network. FIGS. 11 and 12 show reinforcement learning using a neural network according to a seventh extended example of the present invention. As shown in FIG. 11, input data of a state and a weight of an action are input to an input layer of a neural network, and an approximation of a predicted reward Q value is output from an output layer You can apply the receiving model. 12, only the input data of the state is input to the input layer of the neural network, and an approximate value of the Q value, which is a reward predicted for the entire action, is output to the output layer (ouput layer) Can be applied.

According to the seventh extended example of the present invention, when the reinforcement learning using the neural network is applied, the degree of freedom of the input data is excessively high, and when the calculation amount of the number of cases in the entire environment is excessively large, Q is outputted in an approximate manner, Effect is generated.

12, the target y of the neural network can be set to Q ^* , which is an optimal Q value, and the cost function to be minimized in the neural network can be described as follows.

In the above equation, cost (W) is the cost function for the neural network W, and Ws is the output value from the output layer when input data is input to the input layer. y is the optimal Q value. At this time, y can be described as follows.

y is the optimal Q value, Q is the reward of time t, and maxQ (s') is the maximum value of Q that can be received at time t + 1. We change this to a more formal formula:

은 Q에 대한 예측값으로 주어진 s 또는 s,a에 의해서 주어지고 θ라는 상태를 전제로 하게 된다. θ는 뉴럴 네트워크 또는 뉴럴 네트워크의 weight를 의미할 수 있다. Q^*은 optimal Q값을 의미한다. 따라서 뉴럴 네트워크의 cost function은 아래와 같이 표현될 수 있다. In the above equation, Ws is the output value of the neural network,

Is given by s or s, a given as a predictive value of Q, and is assumed to be θ. θ can be the weight of a neural network or a neural network. Q ^* means the optimal Q value. Therefore, the cost function of the neural network can be expressed as follows.

According to Equation (18), the reinforcement learning of the natural language processing apparatus proceeds in the direction of finding a value of θ that minimizes the difference between the predicted value of Q and the optimal Q value.

The algorithm for reinforcement learning of the natural language processing apparatus according to the seventh extended example of the present invention can be exemplarily constructed as follows according to the above equation (18).

Perform a gradient descent step on (y_j - Q(s_j,a_j;θ))²
end for
Algorithm
Initialize action - value function Q with random weights θ
Initialize state s ₁ = {x ₁ }
for t = 1, T do
With probability ε select a random action a _t,
otherwise select a _t = max _a Q ^* (s _t , a; θ)
Execute action a _t in Weight module and observe reward r _t
Set s _{t + 1} = s _t , a _t

Set

Perform a gradient descent step on (y _j - Q (s _j , a _j ; θ)) ²
end for
In the reinforcement learning of the natural language processing apparatus according to the seventh extended example of the present invention, a reinforcement learning algorithm using a neural network

In the above algorithm, the probability ε may mean a greedy algorithm for selecting an action. s is the state, a is the action, Q is the Q value, θ is the weight of the neural network, r is the reward, and y is the target.

[ Example 8 expansion  - Neural  Network reinforcement learning 2]

가 Q에 대한 예측값,

가 optimal Q값(타겟)을 의미할 수 있다. The reinforcement learning using the neural network of the natural language processing apparatus according to the seventh extended example of the present invention may cause a problem that the target is not kept constant. As described in Equation (18), reinforcement learning of the natural language processing apparatus proceeds in the direction of seeking a value that minimizes the difference between the predicted value of Q and the optimal Q value in the cost function of the neural network. the θ value of the optimal Q value, which is the target of the cost function, is updated together with the update of the θ value, so that the target is not kept constant and learning becomes difficult. In Equation 18,

Is the predicted value for Q,

May mean an optimal Q value (target).

로 다르게 구성하고, 특정 시간이 지난 뒤에 또는 특정 조건이 만족되면

를 θ로 업데이트하는 방법을 이용할 수 있다. 도 13은 본 발명의 제8확장예에 따른 자연어 처리 장치의 뉴럴 네트워크를 이용한 강화학습의 네트워크 구조를 도시한 것인데, 도 13에 도시된 바와 같이, 뉴럴 네트워크 θ는 Q의 예측값(Ws)에 이용하고, 뉴럴 네트워크

는 optimal Q값에 이용하도록 구성될 수 있다. 이에 따르면 위의 수학식 18은 아래의 수학식과 같이 수정될 수 있다. Therefore, reinforcement learning using a neural network of a natural language processing apparatus according to an eighth extended embodiment of the present invention can solve such a problem by setting the neural network used in the predicted value of Q and the neural network used in the target as?

And after a certain time has elapsed or when certain conditions are satisfied

To &amp;thetas; can be used. 13 shows a network structure of reinforcement learning using a neural network of a natural language processing apparatus according to an eighth extended example of the present invention. As shown in FIG. 13, a neural network? Is used for a predicted value Ws of Q And a neural network

May be configured to use for the optimal Q value. According to this, equation (18) can be modified as shown in the following equation.

According to Equation (19), reinforcement learning of the natural language processing apparatus proceeds in the direction of finding a value of θ that minimizes the difference between the predicted value of Q and the fixed optimal Q value. Also, when the? Value is updated, only the predicted value for Q is updated and the optimal Q value is fixed.

According to Equation (19) above, the reinforcement learning algorithm of the natural language processing apparatus according to the eighth extension example of the present invention can be configured as an example as follows.

with weights

Initialize state s₁ = {x₁}
for t =1, T do
With probability ε select a random action a_t,
otherwise select a_t = max_aQ^*(s_t,a;θ)
Execute action a_t in 가중치 모듈 and observe reward r_t
Set s_t+1 = s_t,a_t

Set

Perform a gradient descent step on (y_j - Q(s_j,a_j;θ))² with respect to network parameters θ
Every n steps reset

end for
Algorithm
Initialize action - value function Q with random weights θ
Initialize target action - value function

with weights

Initialize state s ₁ = {x ₁ }
for t = 1, T do
With probability ε select a random action a _t,
otherwise select a _t = max _a Q ^* (s _t , a; θ)
Execute action a _t in Weight module and observe reward r _t
Set s _{t + 1} = s _t , a _t

Set

Perform a gradient descent step on (y _j - Q (s _j , a _j ; θ)) ² with respect to network parameters θ
Every n steps reset

end for
In the reinforcement learning of the natural language processing apparatus according to the eighth extended embodiment of the present invention, a reinforcement learning algorithm using a neural network

In the example of the algorithm for reinforcement learning of the natural language processing apparatus according to the eighth extended example of the present invention, the target and the neural network used by the predictive value are different from each other, so that the effect that the target is not varied by the weight update of the neural network is generated. The reinforcement learning works properly because the target is fixed. Also, since the target neural network is updated together with each specific step, the problem that the target is fixed for a long time without being updated can be solved.

In the reinforcement learning of the natural language processing apparatus according to the fourth, fifth, sixth, seventh and eighth embodiments of the present invention, reward is described as being processed by the emotion analysis value for the input data at time t + 1. However, the scope of the present invention is not limited to this, and how long input data of time t + 1 is input and how long input data of time t + 1 is inputted can be reordering of reinforcement learning.

[ Variation example ]

In a modification of the present invention, the function of the weight module 5 may be configured to be included in the domain matching apparatus 3. FIG. 14 shows a signal flow of a natural language processing apparatus according to a modification of the present invention. As shown in Fig. 14, according to a modification of the present invention, the function of the weight module 5 is configured to be included in the domain matching apparatus 3 so that a weight is generated in the domain matching apparatus 3 at time t, to the domain matching apparatus 3 of t + 1. The weight generated in the domain matching apparatus 3 according to the modified embodiment of the present invention may include the weights of the first, second and third extended examples of the present invention.

The reinforcement learning in the case where the function of the weight module 5 is included in the domain matching apparatus 3 according to the modification of the present invention can be configured as shown in Figs. 15 and 16 are schematic diagrams showing a signal flow of reinforcement learning of a natural language processing apparatus according to a modification of the present invention. 15 and 16, the reinforcement learning of the natural language processing apparatus according to the modified embodiment of the present invention is performed by using the domain matching apparatus 3 constituted in the natural language processing apparatus at time t as Agent, the domain matching apparatus 3, the matching of the domain is generated from the Agent of Action _0, the input data 200 of the time t in State _0, the time sensitivity of input data sentimental analysis module (200) (6) of the t + 1 is generated by sensibility analysis A reinforcement learning algorithm can be constructed in which the analysis value is Reward ₁ . Then, the input data 200 at time t + 1 can be restarted to State ₁ to start the reinforcement learning cycle. The entire environment may include a query data set included in each specific domain pipeline of the response module 4 and input data 200 of each time.

Domain Matching Method for Multidomain Natural Language Processing

FIG. 17 is a flowchart illustrating a domain matching method of a natural language processing apparatus for multi-domain natural language processing according to an embodiment of the present invention, in a domain matching method for multi-domain natural language processing according to an embodiment of the present invention. As shown in FIG. 17, the domain matching method according to an embodiment of the present invention may include a dividing step S10, a domain score generating step S11, and a matching step S12.

In the classification step S10, the classification module 2 receives the input data 200 in a normal text form, and the general character type input in the classification module 2 is input to the NLP module 21 as an entity, And is divided into intent. The input data may be divided into words or word combinations (phrase information) by the segmentation step.

The domain score generation step S11 is a step in which the domain matching apparatus 3 calculates the degree of similarity between the query data set of each specific domain pipeline 40 based on the vectorized phrase information received from the classification module 2, This is the step of generating the score.

The matching step S12 is a step of matching the input data to the domain having the highest domain score among the respective domain pipelines.

According to the domain matching method according to an embodiment of the present invention, the amount of calculation of the task of receiving input data and allocating it to each specific domain pipeline is very small and the processing speed is improved.

FIG. 18 shows an example of the result of each step of the domain matching method according to an embodiment of the present invention. According to Fig. 18, input data such as "Is it possible to deliver seasoned chicken to Gangnam Station?" Is segmented into phrases such as 'Gangnam Station', 'Seasoned chicken', 'Delivery' The domain matching device 3 calculates a domain score, which is a score for how specific the at least one phrase information is to a specific domain pipeline, to determine 'food delivery' = 1, 'clothing order' = 0.5, 'weather' = 0.3, and 'time' = 0.2. In step S12, the domain matching apparatus 3 matches the input data 200 to the "food delivery" domain pipeline, which is a particular domain pipeline that matches the highest or the domain score satisfying a certain condition.

According to an embodiment of the present invention, the similarity between the input data and the set of question data included in each specific domain pipeline can be used in the domain score calculation of S11.

[ First extended example ]

FIG. 19 is a flowchart illustrating a domain matching method according to a first extended example of the present invention, with respect to a first extended example of a domain matching method for multi-domain natural language processing according to an embodiment of the present invention. As shown in FIG. 19, the domain matching method according to the first extended example of the present invention may further include a weight generating step (S10 ') before S11. In addition, the domain score generation step S11 may be configured to further apply a weight in calculating the domain score (S11 ').

The domain matching apparatus may perform a weight generation step S10 'before S11' for generating a weight for applying a weight for matching a domain in the calculation of a domain score in step S11 of calculating a domain score, Lt; RTI ID = 0.0 &gt; and / or &lt; / RTI &gt;

The weight in the weight generation step S10 'may mean a weight generated by the weighting module 5 or the domain matching device 3 at time t-1 and provided to the domain matching device 3 at time t. In the weight generation step S10 ', a weight can be generated and used for S11'. Applying the weight in S11 applies the domain score calculated in the domain matching apparatus 3 at time t or the information about the specific domain to the domain score in the domain matching apparatus 3 at time t + 1 or the matching of a specific domain .

The input data input at time t affects the domain score of the matched domain with respect to the input data at time t + 1, and the input data is continuously input The more weighted weights are given to the domain scores of a particular domain, the less weights are weighed over time, and the weights are weighted when we need to match another domain with a higher domain score of another particular domain . That is, the weight may mean the probability that the domain will be maintained according to the time difference in which the input data is input.

[ Second Extended Example ]

In the domain matching method according to the second extended example of the present invention, in a situation where a series of input data is input, the weighting module 5 determines the weighting value of the domain matching device 3 . &Lt; / RTI &gt; X (i-1), x (i-2), x (i-3) ... when the current input data, that is, the i- If x (iM) is equal to D _k all of the allocated domain, which can be seen that the M input data is continuous in a domain, so that the domain matching unit 3 is a weight of W (D _k, x ( i)) may be multiplied by P (D _k | x) calculated in the similarity method to maintain the domain matching.

20 is a flowchart illustrating a domain matching method according to a second extended example of the present invention. 20, the domain matching method according to the second extended example of the present invention includes a classification step S10, a weight generation step S20, a domain score calculation step S21, a domain sustainability determination step S22, A domain maintenance step S23, a domain score calculation step S24, and a domain matching step S25.

In the classification step S10, the classification module 2 receives the input data 200 in a normal text form, and the general character type input in the classification module 2 is input to the NLP module 21 as an entity, And is divided into intent. The input data may be divided into words or word combinations (phrase information) by the segmentation step.

The weight generating step S20 is a step in which the weighting module 5 or the domain matching device 3 calculates the weighting factor of the domain score of the time domain t based on the domain score of the immediately preceding same domain domain, And generating a weight to be applied. In the second extended example of the present invention, the immediately preceding same domain period may mean a time period of time t-1, t-2, ..., t-n matched with the same domain pipeline as the domain of t-1.

The domain score calculation step S21 uses the similarity degree with the query data set of the specific domain pipeline matched in the immediately preceding same domain interval calculated on the basis of the phrase information received in S10 and the weight of S20 calculated by the domain matching apparatus 3 To generate a domain score.

The domain sustainability determination step S22 is a step of determining whether or not the domain matching apparatus 3 at time t holds the domain or not. If the domain score generated at S21 is greater than or equal to a predetermined specific value or a specific probability The process may be branched to S23 so as to be matched again to the specific domain pipeline matching the immediately preceding domain period. In addition, if the domain score generated in S21 is less than or equal to a predetermined specific value or a specific probability, the process may be branched to S24 so as to match the specific domain pipeline other than the specific domain pipeline matched in the immediately preceding same domain period.

In the domain maintenance step S23, if the domain score generated in step S21 is greater than or equal to a preset specific value or a predetermined probability, the domain matching apparatus 3 is set so as to match again the specific domain pipeline matched in the immediately preceding same domain period. Is the step of maintaining the domain.

The domain score calculation step S24 is performed so that the domain score generated in S21 is matched to a specific domain pipeline other than the specific domain pipeline matched in the immediately preceding same domain period when the domain score is less than or equal to a predetermined specific value or a specific probability. The domain matching apparatus 3 calculates similarity with the query data set of the other specific domain pipeline based on the phrase information received in S10 and uses the similarity and the weight of S20 to calculate the domain for each remaining specific domain pipeline And calculating the score.

The domain matching step S25 is a step of matching the input data to the domain having the highest domain score among the remaining specific domain pipelines. And the matched domain is transferred to another domain.

When applying this method, for example, when input data classified as a chicken delivery domain is processed (eg, how much is chicken / golden olive coupon applies), input data (eg, what time is it / how much?) is input, it is possible to classify it into a domain that has been processed previously and converge it into a conversation of consistent subject.

Further, in matching the domain pipeline through the input data, the effect of reducing the amount of computation is generated. When the previous input vector (vectorized input data) x (i-1) belongs to D _k , the product of P (D _k | x (i)) and W (D _k , x examples the effects that can be configured to immediately assign the x (i-1) to D _k, without having to calculate the degree of similarity with other domains except for D _k greater than 50%) is generated.

[ Third Expansion Example ]

In a domain matching method according to the third extended example of the present invention, when a situation in which a series of input data is input and a situation is to be transferred from the currently matched domain pipeline to another domain pipeline, the weight module 5 May be configured to weight the domain matching device 3 based on the probability of transitioning from the matching domain to the other domain with the incoming input data. More specifically, in the domain matching method according to the third extension example of the present invention, when a series of input data is input and each of the previously input data is assigned to one domain, A probability distribution table (domain transition probability distribution) that can be transferred from a domain in which data is allocated to a domain in a domain set is constructed and applied to calculate a probability to be assigned to each domain.

More specifically, FIG. 21 is a flowchart illustrating a domain matching method according to a third extended example of the present invention. As shown in FIG. 21, the domain matching method according to the third extended example of the present invention may further include a transition probability application step S24 'before S24, which is the domain score calculation step of FIG.

The transition probability application step S24 'is a step of applying the transition probability to the weight so that the transition probability can be included in the weight used in S24. Transition probability means information about the probability of transition from one domain to another.

In this way, the probability that a next input data can be classified into a domain having a high probability of transition from the domain of the previous input data is calculated based on the domain transition probability, which is a probability of transition between the domains at the time of domain conversion. Lt; / RTI &gt; to determine the domain score. For example, the probability that a domain will be transferred to an entertainer story when sharing a movie story is high, but it can be applied to the calculation of a domain score that the probability that a domain is transferred to a story of a chicken after a story of a movie is low. The domain transition probability can be calculated based on data on past domain transitions and can be changed according to the input time difference of input data.

Also, at the time when the domain moves from the existing D _k to the next domain D _j , it is necessary to first calculate the domains with high probability of domain transition with D _k without calculating the similarity with all domains except D _k in order to calculate D _j The amount of computation can be reduced.

[ Fourth Extended Example ]

The weighting module 5 of the domain matching method according to the fourth extended example of the present invention calculates the weight based on the probability of transition from the matching domain to the other domain And a reinforcement learning is performed on the weight, so that the weight is updated in accordance with the reaction of the user. The compensation to be used in the reinforcement learning includes the user's reaction to the output data 201 of the time t output by the matching of the specific domain pipeline 40 of the domain matching apparatus 3, ) Of the input data 200 at the time t + 1 that is input again after the user has received the input data 200 may be used.

More specifically, FIG. 22 is a flowchart illustrating a domain matching method according to a fourth extended example of the present invention. As shown in FIG. 22, after the domain matching step S12 of FIG. 17, the emotional analysis step S40 and the reinforcement learning step S41 may be further included.

The sensibility analysis step S40 is a step for generating sensibility analysis values by performing sensibility analysis of new input data inputted by a user after output data is outputted in a domain pipeline matched after the domain matching step S12.

The reinforcement learning step S41 is a step of updating the sensitivity score of S40 to reward, setting the domain matching device 3 as an agent, setting the domain score as action, and updating the domain score of specific phrase information according to the emotion analysis value . For example, when the sensitivity of the new input data is classified as "good ", the domain score calculation method of the domain matching apparatus 3 is performed such that a high weight is assigned to the matching domain in S12 with respect to the phrase information of the input data in S10 Can be updated. For example, when the sensitivity of the new input data is classified as "bad ", the domain score calculation of the domain matching apparatus 3 is performed so that a weighted value is assigned to the matching domain in S12 with respect to the phrase information of the input data in S10 The method can be updated.

23 is a flowchart showing another example of the domain matching method according to the fourth extended example of the present invention. 23, in the reinforcement learning of the domain matching method according to the fourth extended example of the present invention, the reinforcement learning step S41 sets the emotion analysis value of S40 as reward and the weight module 5 as agent, The weight may be set to action, and the weight for the specific phrase information may be updated in accordance with the sentence analysis value. For example, when the sensitivity of new input data is classified as "good ", the weight calculation method of the weight module 5 is updated so that a high weight is assigned to the matching domain in S12 for the phrase information of the input data of S10 . For example, when the sensitivity of the new input data is classified as "bad ", the weight calculation method of the weight module 5 is performed such that a weighted value is assigned to the matching domain in S12 with respect to the phrase information of the input data in S10 Can be updated.

In the above description, it is described that, after the input data is inputted, the domain matching apparatus 3 matches the domain and assigns the input data to the specific domain pipeline 40. However, as shown in FIG. 24, And the scope of the present invention is not limited thereto. 24 is a schematic diagram showing a signal flow of a natural language processing apparatus according to a modification of the present invention. As shown in FIG. 24, each domain pipeline 40 directly receives the input data 200 divided by the phrase information by the classification module, and generates output data in each domain pipeline 40 And selects one output data by selecting and sorting a plurality of output data in the domain matching apparatus 3 by rating or ranking.

Although the chatbot has been described above for the convenience of explanation, the scope of the present invention is not limited to the chatbot, but an artificial intelligence module for voice analysis (voice bots), an artificial intelligence module for natural language analysis, It is within the scope of the present invention that the present invention is applied to a natural language processing apparatus such as an artificial intelligence module.

As described above, those skilled in the art will appreciate that the present invention may be embodied in other specific forms without departing from the spirit or essential characteristics thereof. It is therefore to be understood that the above-described embodiments are to be considered in all respects only as illustrative and not restrictive. The scope of the present invention is defined by the appended claims rather than the detailed description, and all changes or modifications derived from the meaning and scope of the claims and their equivalents should be construed as being included in the scope of the present invention.

The features and advantages described herein are not all inclusive, and in particular, many additional features and advantages will be apparent to those skilled in the art in view of the drawings, specification, and claims. Moreover, it should be noted that the language used herein is primarily chosen for readability and for purposes of teaching, and may not be selected to delineate or limit the subject matter of the invention.

The foregoing description of embodiments of the invention has been presented for purposes of illustration; It is not intended to be exhaustive or to limit the invention to the precise form disclosed. Those skilled in the art will appreciate that many modifications and variations are possible in light of the above teachings.

Some portions of this description describe embodiments of the present invention relating to symbolic representations and algorithms of informational operations. These algorithmic descriptions and representations are generally used by those skilled in the data processing arts to efficiently convey the essence of their work to the other skilled artisan. While such operations are described functionally, computationally, or logically, they are understood to be implemented by a computer or equivalent electrical circuitry, microcode, or the like. Furthermore, this is also conveniently demonstrated without loss of generality, sometimes to represent an arrangement of such operations as a module. The operations described above and their associated modules may be implemented in software, firmware, hardware, or any combination thereof.

Any of the steps, operations, or processes described herein may be performed or implemented in conjunction with one or more hardware or software modules, alone or in combination with other devices. In one embodiment, a software module is implemented with a computer program product comprised of a computer-readable medium comprising computer program code, and the computer program code is executable to perform any or all of the processes, steps, May be executed by a computer processor.

Furthermore, embodiments of the invention may relate to an apparatus for performing the operations herein. These devices may include a general-purpose computing device that may be specially constructed and / or selectively activated or reconfigured by a computer program stored within the computer for the required purpose. Such a computer program may be stored in any type of media suitable for storing computer readable storage media or type of instructions, and may be coupled to a computer system bus. Further, any computing system referred to herein may comprise a single processor, or it may be a structure employing a multiprocessor design for increased computing power.

Finally, the language used herein has been chosen primarily for readability and for purposes of teaching, and may not be selected to describe or limit the subject matter of the invention.

The scope of the invention is, therefore, not to be limited by the Detailed Description, but is to be defined by the claims of any application based thereon. Accordingly, the disclosure of embodiments of the invention is illustrative and not restrictive of the scope of the invention, which is set forth in the following claims.

1: Natural language processing device
2: Separation module
3: Domain matching device
4: Response module
5: Weight module
6: Sentimental analysis module
21: NLP module
22: vectorization module
40: Domain Pipeline
200: input data
201: Output data

Claims (3)

A first classifying module for receiving input data and generating the phrase information of the input data;
A second segmentation module for receiving a query data set of a specific domain pipeline that generates output data for a particular domain and generating the query information of the query data set;
In order to determine whether or not the phrases classified in the first classification module are characteristic to a specific domain, the probability of occurrence of a word or occurrence of a phrase in the query data set is compared with the probability of occurrence of a word or occurrence of a word in a general corpus, A domain-specific score calculation module for determining a domain-specific score, which is a score of a characteristic word of a corresponding word in the domain;
A domain score calculation module for calculating a domain score for each domain pipeline using the similarity of the input information of the input data and the query information of the query data set of the specific domain pipeline and the domain specific score; And
A domain matching module for matching and inputting the phrase information of the input data to the domain pipeline having the highest domain score;
/ RTI &gt;
Domain matching device using domain - specific score.
A first segmenting step of receiving the input data and generating the phrase information of the input data;
A second segmentation module receiving a query data set of a specific domain pipeline that generates output data for a specific domain and generating the query information of the query data set;
The domain-specific score calculation module may determine whether a word classified in the first classification module is characteristic to a specific domain, a probability of occurrence of a word in the query data set or a probability of occurrence of a word in the query data set, A domain-specific score calculation step of comparing the probability of appearance and determining a domain-specific score which is a score of a characteristic word of the corresponding word in the corresponding domain through a difference;
Domain score calculation module calculates a domain score for each domain pipeline using the similarity of the input information of the input data to the query information of the query data set of the specific domain pipeline and the domain specific score, Calculating step; And
A domain matching step in which the domain matching module matches and inputs the phrase information of the input data to the domain pipeline having the highest domain score;
/ RTI &gt;
Domain matching method using domain - specific score.
A first segmenting step of receiving the input data and generating the phrase information of the input data;
A second segmentation module receiving a query data set of a specific domain pipeline that generates output data for a specific domain and generating the query information of the query data set;
The domain-specific score calculation module may determine whether a word classified in the first classification module is characteristic to a specific domain, a probability of occurrence of a word in the query data set or a probability of occurrence of a word in the query data set, A domain-specific score calculation step of comparing the probability of appearance and determining a domain-specific score which is a score of a characteristic word of the corresponding word in the corresponding domain through a difference;
Domain score calculation module calculates a domain score for each domain pipeline using the similarity of the input information of the input data to the query information of the query data set of the specific domain pipeline and the domain specific score, Calculating step; And
A domain matching step in which the domain matching module matches and inputs the phrase information of the input data to the domain pipeline having the highest domain score;
/ RTI &gt;
A domain matching method using a domain specific score is a program stored in a recording medium executed on a computer.

Images