KR20220153852A - Natural language processing apparatus for intent analysis and processing of multi-intent speech, program and its control method - Google Patents

Abstract

The present invention relates to a technology for recognizing intent of a user in speech including multi-intent in a more accurate manner. More specifically, the present invention relates to a technology, which receives user speech, stores a first corpus and a second corpus, receives connection information between the first and second corpora, generates a combination corpus, in which the first and second corpora are combined, based on the received connection information, trains at least one natural language processing model based on the combination corpus, and processes the input user speech based on the at least one natural language processing model. The present invention has an advantage of being able to effectively generate a corpus for recognizing multi-intention speech.

Description

Natural language processing device, program, and control method for intent analysis and processing of multi-intent utterances

The present invention relates to a natural language processing apparatus and a control method for analyzing and processing multi-intent utterances, and more specifically, to a parallel intention utterance in which several intentions are input in succession and the result of one intention is another intention. It relates to an apparatus and control method for effectively processing dependent intent utterances required for processing.

A chatbot is a system in which a user and a computer exchange daily conversations on various topics in the form of natural language. These chatbots are used as element technologies that provide a natural conversation interface by being integrated with purpose-oriented systems such as information acquisition and FAQ systems.

Natural language processing technology is required to understand a user's speech, properly execute a command, or perform a response to the speech. Natural language processing technology is one of the core functions of artificial intelligence as a technology for interaction between computer and human language.

In user speech, there are many cases in which two or more intentions are included in one sentence. For example, in one sentence “Give me a fortune and tell me the weather”, a first intention to ask for a fortune and a second intention to request weather information are included. In this way, cases in which individual intentions are simply listed are classified as parallel intention utterances.

In contrast, when the result value for one intent is necessary to process another intent, it can be classified as a dependent intent utterance. For example, there may be a user utterance "Find a recipe for kimchi stew (first intention) and text me (second intent)." In order to perform a task related to the second intention, a result of the task related to the first intention is required.

In the case of parallel intention utterances, a plurality of intentions can be distinguished through a method of dividing front and rear sentences around a specific keyword. However, in the case of dependent intention utterances, there is a problem that segmentation cannot be performed based on specific keywords.

Accordingly, there is a demand for development of a natural language processing device and control method capable of clearly distinguishing and processing dependent intended utterances included in user utterances.

An object to be solved by the present invention is to provide an apparatus and method capable of identifying an accurate user intent in multi-intent utterances by distinguishing between peer-intended utterances and subordinate-intended utterances.

Another problem to be solved by the present invention is to provide an apparatus and method for automatically generating a corpus for recognizing dependent intended utterances.

The technical problems to be achieved in the present invention are not limited to the technical problems mentioned above, and other technical problems not mentioned will be clearly understood by those skilled in the art from the description below. You will be able to.

According to one aspect of the present invention to solve the above or other problems, a memory for storing instructions; and a processor configured to execute the stored instructions, wherein the processor: receives user utterance, stores a first corpus and a second corpus, receives connection information between the first and second corpus, and receives the input Based on the received connection information, a combined corpus that is a corpus combining the first and second corpus is generated, at least one natural language processing model is trained based on the combined corpus, and based on the at least one natural language processing model Provided is a natural language processing device that processes the input user utterance.

The processor may divide the input user speech into a plurality of speeches based on a plurality of intentions included in the input user speech.

The processor may determine whether the input user utterance is an intended parallel utterance, and if it is determined that the input user utterance is an intended parallel utterance, the input user utterance may be divided into a plurality of partial parallel intended utterances based on the equal divider. have.

The processor determines whether the input user utterance is a dependent intended utterance if it is determined that it is not the equivalent intended utterance, and if it is determined that the input user utterance is the dependent intended utterance, the input user utterance is divided into a plurality of sub-divisions based on the dependent divider. It can be divided into partial dependent intent utterances.

The plurality of partial dependent intended utterances may include a data partial dependent intended utterance and an action partial dependent intended utterance.

The first corpus may be a data part corpus, and the second corpus may be an action part corpus.

The connection information may include information identifying at least one of the first and second corpora.

Each of the first and second corpora includes a plurality of spoken phrases, and the combined corpus is a corpus obtained by extracting and matching one spoken phrase included in the first corpus and one spoken phrase included in the second corpus. can be

According to another aspect of the present invention to solve the above or other problems, receiving a user speech; storing the first corpus and the second corpus; receiving connection information between the first and second corpus; generating a combined corpus, which is a corpus obtained by combining the first and second corpus, based on the input connection information; training at least one natural language processing model based on the combination corpus; and processing the input user utterance based on the at least one natural language processing model.

The processing of the user speech may include dividing the input user speech into a plurality of speeches based on a plurality of intentions included in the input user speech.

The processing of the user utterance may include determining whether the input user utterance is a peer-to-peer utterance; and dividing the input user utterance into a plurality of partial equivalent intended utterances if it is determined that the utterance is intended to be equivalent utterance.

The processing of the user utterance may include: determining whether the input user utterance is a subordinate intended utterance when it is determined that the user utterance is not a parallel intended utterance; and dividing the input user utterance into a plurality of partial dependent intended utterances if it is determined that the dependent intended utterance is the dependent intended utterance.

The plurality of partial dependent intended utterances may include a data partial dependent intended utterance and an action partial dependent intended utterance.

The first corpus may be a data part corpus, and the second corpus may be an action part corpus.

The connection information may include information identifying at least one of the first and second corpora.

Each of the first and second corpora includes a plurality of spoken phrases, and the combined corpus is a corpus obtained by extracting and matching one spoken phrase included in the first corpus and one spoken phrase included in the second corpus. can be

Effects of the natural language processing device, program, and control method according to the present invention are described below.

According to at least one of the embodiments of the present invention, there is an advantage in that a corpus for recognizing multi-intended utterances can be effectively generated.

In addition, according to at least one of the embodiments of the present invention, there is an advantage in that natural language processing can be performed by accurately determining a user's intention in a subordinate intention utterance.

A further scope of the applicability of the present invention will become apparent from the detailed description that follows. However, since various changes and modifications within the spirit and scope of the present invention can be clearly understood by those skilled in the art, it should be understood that the detailed description and specific examples such as preferred embodiments of the present invention are given as examples only.

1 is a block diagram of a natural language processing apparatus 100 according to an embodiment of the present invention.
2 shows an example of generating a combined corpus by combining two corpora according to an embodiment of the present invention.
FIG. 3 shows an example of a combined corpus 210 generated based on the first and second corpora 200 - 1 and 200 - 2 of FIG. 2 and the connection information 203 .
4 is a diagram illustrating a method in which the corpus generator 106 generates a subordinate segmented corpus according to an embodiment of the present invention.
5 is a diagram illustrating a natural language processing model constituting the speech processing unit 104 according to an embodiment of the present invention.
6 is a diagram illustrating an example of a peer classifier 501 and a single-subordinate classifier 502 according to one embodiment of the present invention.
7 is a diagram illustrating an example of a peer divider 503 and a subordinate divider 504 according to one embodiment of the present invention.
8 is a diagram illustrating a user speech processing sequence by the speech processing unit 104 according to an embodiment of the present invention.
9 illustrates an example in which a peer splitter 503 divides utterances based on intent according to an embodiment of the present invention.
10 illustrates an example in which a dependent splitter 504 splits a dependent intent utterance according to one embodiment of the present invention.
11 illustrates an example in which the speech processing unit 104 processes single-intention speech according to an embodiment of the present invention.
FIG. 12 illustrates a conceptual diagram of processing of the speech processing unit 104 when an equivalent intention and a subordinate intention coexist according to an embodiment of the present invention.
13 is a diagram illustrating a configuration of a natural language processing device 100 according to an exemplary embodiment.

Hereinafter, the embodiments disclosed in this specification will be described in detail with reference to the accompanying drawings, but the same or similar elements are given the same reference numerals regardless of reference numerals, and redundant description thereof will be omitted. The suffixes "module" and "unit" for components used in the following description are given or used together in consideration of ease of writing the specification, and do not have meanings or roles that are distinct from each other by themselves. In addition, in describing the embodiments disclosed in this specification, if it is determined that a detailed description of a related known technology may obscure the gist of the embodiment disclosed in this specification, the detailed description thereof will be omitted. In addition, the accompanying drawings are only for easy understanding of the embodiments disclosed in this specification, the technical idea disclosed in this specification is not limited by the accompanying drawings, and all changes included in the spirit and technical scope of the present invention , it should be understood to include equivalents or substitutes.

Terms including ordinal numbers, such as first and second, may be used to describe various components, but the components are not limited by the terms. These terms are only used for the purpose of distinguishing one component from another.

It is understood that when an element is referred to as being "connected" or "connected" to another element, it may be directly connected or connected to the other element, but other elements may exist in the middle. It should be. On the other hand, when an element is referred to as “directly connected” or “directly connected” to another element, it should be understood that no other element exists in the middle.

Singular expressions include plural expressions unless the context clearly dictates otherwise.

In this application, terms such as "comprise" or "have" are intended to designate that there is a feature, number, step, operation, component, part, or combination thereof described in the specification, but one or more other features It should be understood that the presence or addition of numbers, steps, operations, components, parts, or combinations thereof is not precluded.

When a user's utterance includes a plurality of intentions, it may be classified into an equivalent intention utterance and a subordinate intention utterance according to a relationship between the plurality of intentions.

The relationship between a plurality of intents included in a parallel intent utterance is that (1) utterances representing each intent exist separately in one sentence, (2) each partial utterance constitutes a complete intent by itself, ( 3) There is a feature that a plurality of intentions can be processed simultaneously.

As an example of a parallel intention utterance, “tell me the fortune (first intention), tell me the weather (second intention)”.

The relationship between a plurality of intents included in a dependent intent utterance is: (1) utterances representing each intent may be mixed with each other, (2) each partial utterance may not constitute a complete intent by itself, and (3) each There is a feature that there is a processing order between intentions.

An example of a dependent intent utterance is "Text me (first intent), find a recipe for kimchi stew (second intent) and send it to me (first intent)". As can be seen in this example, since multiple intentions are mixed in one sentence, it is not an easy process to separate utterances based on intentions.

Each of a plurality of intents constituting the dependent intent utterance is called a partial intent. And, the dependent intention utterance may be configured to include a data part and an action part. The data part means the intention to produce data, and the action part means the action intention to receive and process the data (result data) produced by the data part.

Accordingly, the method for processing dependent intent utterances, unlike the existing method for processing parallel intent utterances, (1) requires special processing to generate a corpus for training, (2) requires a separate module for intent segmentation, (3) A scheduler that preserves the processing order and delivers it to the actual execution module and receives the result value is required.

Meanwhile, parallel intention utterances and subordinate intention utterances may be mixed and included in user utterances. A subordinate intent utterance may be included within a peer intent utterance. For example, the user's speech "Tell me your fortune, find a recipe for kimchi stew, and send it to me as a text message." “Look for a recipe for kimchi stew” and “Find a recipe for kimchi stew and send it to me as a text message” have an equal intention utterance relationship, but “find a recipe for kimchi stew” and “send me a text message” have a relationship of subordinate intention utterances.

Conversely, there may be a case in which a parallel intended utterance is included in a subordinate intended utterance.

1 is a block diagram of a natural language processing apparatus 100 according to an embodiment of the present invention.

Specifically, the natural language processing device 100 includes a speech input unit 101, a corpus database 102, a connection information input unit 103, a speech processing unit 104, a model training unit 105, and a corpus generator 106. can be configured. The components shown in FIG. 1 are not essential to implement the natural language processing device 100, so the natural language processing device 100 described in this specification includes more or fewer components than the components listed above. can have

The speech input unit 101 is a component that receives user speech. The input user speech may be transmitted to the speech processing unit 104 .

The corpus database 102 is a component that stores at least one corpus. A corpus or corpus (English: corpus, plural: corpora) refers to a set of language samples for a specific purpose for natural language research.

The corpus database 102 according to an embodiment of the present invention may store at least one of first and second corpus and combined corpus.

In one embodiment of the present invention, the first corpus may be a data part corpus, and the second part intention corpus may be an action part corpus.

The data part corpus means a corpus composed of a plurality of spoken phrases related to the data part. That is, the data part corpus may include only spoken phrases related to the intention of producing data.

Further, the action part corpus means a corpus including a command for performing an action, and may be composed of a plurality of utterance phrases related to the action part.

The data part corpus and action part corpus may be corpus input by a user. That is, assuming that a chatbot is designed, it will be a corpus input by a chatbot developer.

The combination corpus means a corpus matched by extracting one spoken phrase from the first and second partial intention corpus. Hereinafter, with reference to FIG. 3, a method for generating a combined corpus will be described later.

The connection information input unit 103 is a component that receives connection information between the first and second corpora from a user. Connection information may refer to specific information for combining the first and second corpora.

The speech processing unit 104 processes the input user speech based on at least one natural language processing model. A natural language processing model according to an embodiment of the present invention is an artificial intelligence model, and the model training unit 105 described below may continuously train and optimize the natural language processing model. User utterance processing according to an embodiment of the present invention, when a user utterance includes a plurality of intents, a process of determining whether it is a peer intent or a subordinate intent, a process of dividing an utterance based on a plurality of peer intents, a plurality of It may include at least one of dividing utterances based on the dog's dependent intentions and outputting the divided utterances.

The natural language processing model may include at least one of a peer classifier, a subordinate classifier, a peer divider, and a subordinate divider. Details of each model will be described later.

The model training unit 105 is a component that trains at least one natural language processing model.

The corpus generator 106 is a component that generates a combined corpus by combining the first and second corpus based on the connection information. A detailed combination method will be described below with reference to FIGS. 2 and 3 .

2 shows an example of generating a combined corpus by combining two corpora according to an embodiment of the present invention.

The first corpus 200-1 is a data part corpus, and the second corpus 200-2 is an action part corpus. The first and second corpora 200-1, 200-2 may be stored on the corpus database 102, and may be a corpus input from a user (such as a chatbot developer).

The first corpus 200-1 may include a first utterance list 202-1 related to the first intention 201-1. Likewise, the second corpus 200-2 may include a second utterance list 202-2 related to the second intention 201-2. It is proposed that the speech phrase included in the action part corpus according to one embodiment of the present invention include a predetermined special symbol (__DATA__ in the illustrated example) at a position where the data part speech phrase is inserted by combination. This is to generate a combinatorial corpus in a natural form.

For example, in the example shown in FIG. 2 , when the 3rd spoken phrase 205-3 “air condition” and the 7th spoken phrase 205-7 “give me __DATA__” are combined, the 7th spoken phrase 205-7 The third speech phrase 205-3, "air state", may be inserted into the '__DATA__' part of the speech phrase 205-7, and a combined utterance "Deliver the air state to me" may be generated.

In the illustrated example, the first and second utterances 202-1 may include identification information for identifying an entity and a representative word included in each entity. An entity refers to a unit of a bundle of information that is important to understand from a user's speech. And, the representative word means at least one sub-item defined by the user as being included in the entity.

For example, if there are pepperoni, shrimp, and cheese pizzas as a pizza menu in a pizza shop, representative words such as “pepperoni”, “shrimp”, and “cheese” can be defined in an entity called “pizza menu”. After defining the “pizza menu” entity like this, when the user mentions “pepperoni”, it can be recognized as saying “pizza menu”. In the illustrated example, the boxed “air condition”, “this week”, “integrated standby”, “yongin”, “everland”, “sister's mother's room”, and “me” may be identified entities or representative words.

Identification information for identifying an entity or representative word is information about whether the entity or representative word included in the first and second utterances 202-1 exists and/or, if present, the location where the corresponding entity or representative word exists. can mean

Subsequently, the connection information input unit 103 may receive connection information 203 from the user. The connection information 203 means information for combining the first and second corpora 200-1 and 200-2 by connecting them to each other. The connection information 203 according to an embodiment of the present invention may include information identifying at least one of the first and second corpora. For example, as shown in FIG. 3, the connection information 203 is stored as attribute information of the second corpus 200-2, and in a state where it is stored as attribute information of the second corpus 200-2. It may be configured to include information 204 identifying the first corpus 200-1, which is the counterpart of the combination.

The corpus generator 106 generates a combined corpus based on the first and second corpus 200 - 1 and 200 - 2 and the connection information 203 .

FIG. 3 shows an example of a combined corpus 210 generated based on the first and second corpora 200 - 1 and 200 - 2 of FIG. 2 and the connection information 203 .

As shown in the figure, the combined corpus 210 generated by the corpus generator 106 includes the speech phrases 205-1 to 205-4 included in the first corpus 200-1 and the first corpus 200-1. 2 Speech phrases 205-5 to 205-8 included in the corpus 200-2 may be extracted and matched one by one to be generated.

In the example shown in FIG. 3 , in the first combined utterance 301-1 included in the combined corpus 210, the first spoken phrase 205-1 and the fifth spoken phrase 205-5 are each included in the first corpus. It is a result extracted from (200-1) and the second corpus (200-2) and matched. In the same way, when one-to-one matching is performed while sequentially extracting one from the first corpus 200-1 and the second corpus 200-2, the combined corpus 210 shown in FIG. 3 can be generated.

The combinatorial corpus 210 generated in this way may be used for learning a subordinate classifier. Hereinafter, the combination corpus 210 used for learning the dependent classifier is referred to as a 'subordinate classification corpus'. In the illustrated figure, highlighted (bold and underlined) parts represent data part utterances, and the remaining parts represent action part utterances.

The corpus generator 106 according to an embodiment of the present invention further proposes to produce a 'subordinate segmentation corpus' further including discrimination information for distinguishing the data unit and the action unit in the generated subordinate classification corpus. The division information may be, for example, information indicating a start point and an end point of the data unit. More specifically, the classification information may include analysis information through a Beginning-Inside-Outside (BIO) tagging method. The BIO tag refers to a tag that divides an utterance into a plurality of units and marks the start point and end point of each unit.

4 is a diagram illustrating a method in which the corpus generator 106 generates a subordinate segmented corpus according to an embodiment of the present invention. A subordinate segmentation corpus according to an embodiment of the present invention refers to a corpus further including discrimination information in the subordinate classification corpus.

According to the BIO tagging method used in FIG. 4 , the data part is replaced with 'BI' based on the morpheme unit, and the action part is replaced with O based on the morpheme unit. In addition, the first morpheme of the data part is 'B', and consecutive morphemes can be replaced with 'I' in a repeated form.

4 shows an example of BIO tagging the first combined utterance 301-1, but other combined utterances may be similarly BIO tagged to generate a subordinate segmented corpus.

5 is a diagram illustrating a speech processing model constituting the speech processing unit 104 according to an embodiment of the present invention. The speech processing unit 104 according to an embodiment of the present invention may be configured to include a peer classifier 501, a single-dependent classifier 502, a peer divider 503, and a subordinate divider 504. The components shown in FIG. 5 are not essential to implement the speech processing unit 104, and may have more or fewer components than those listed above.

If a utterance is given, the peer classifier 501 determines whether it is a peer-intended utterance, and outputs a discrimination result. For example, whether it is a parallel intent utterance may be output in a binary classification form. A specific example in which the peer classifier 501 determines parallel intended utterances will be described later with reference to FIG. 6 .

The model training unit 105 trains the peer classifier 501 based on the peer classification corpus. The peer classification corpus refers to a corpus of utterances for which it is already known whether or not the utterance is intended to be equivalent. For example, the peer classifier 501 is an artificial intelligence model, and the model training unit 105 may train the peer classifier 501 through supervised learning. In this case, supervised learning means learning to find an output suitable for a given input using data having input values and corresponding output values.

Given an utterance, the single-subordinate classifier 502 determines whether it has a single intent or a dependent intent. A specific example of determining whether the single-dependent classifier 502 has a single intention or a dependent intention will be described later with reference to FIG. 6 . In this case, it is assumed that the utterance input to the single-subordinate classifier 502 is either a single intent or a dependent intent, and that a peer intent utterance is not input. The model training unit 105 trains the single-subordinate classifier 502 by combining the single-intention corpus and the subordinate classification corpus generated by the corpus generator 106 . In this case, the single-dependent classifier 502 may be an artificial intelligence model, and training may be performed through supervised learning as described above.

6 is a diagram illustrating an example of a peer classifier 501 and a single-subordinate classifier 502 according to one embodiment of the present invention. According to the illustrated figure, when a predetermined text is entered as an input 601 in the peer classifier 501 and the single-dependent classifier 502, a softmax result ( 602) may be output.

The softmax result 602 for the peer classifier 501 may be output in a binary classification form indicating whether it is a peer intended utterance or not.

The softmax result 602 for the subordinate classifier 502 may be output in the form of a binary classification indicating whether it is a single intent or a dependent intent.

Referring back to FIG. 5 , when a given utterance is a parallel intent utterance, the peer divider 503 divides the utterance based on the intent and outputs the divided utterance. For example, as in the above example in the background of the invention, if the utterance "Find a recipe for kimchi stew (first intention) and send it to me as a text message (second intention)" is input, "Find a recipe for kimchi stew" The utterance is divided into a first intention of “Send me a text message” and a second intention of “Send me a text message.”

The model training unit 105 trains the equal divider 503 with the equal division corpus. The equal division corpus means a corpus in which a parallel intent utterance and a result obtained by dividing the equal intent utterance based on the intention (correct answer) are combined.

When a given utterance is a dependent intended utterance, the subordinate divider 504 may separate and output utterances based on intention. At this time, the separated utterances may be divided into data part utterances and action part utterances.

7 is a diagram illustrating an example of a peer divider 503 and a subordinate divider 504 according to one embodiment of the present invention. According to the illustrated diagram, a given utterance is converted (701) into at least one embedding and entered into an input of biLSTM (700, Bidirectional LSTM). The biLSTM (700) is a model that uses LSTM, a model that overcomes the limitations of recurrent neural networks (RNNs) through gate techniques, using both forward and backward results. As an output of the biLSTM 700, tagging information 702 for segmenting utterances may come out.

The model shown in FIG. 7 is only one example of a BIO tagging-based segmentation module, and various BIO tagging-based segmentation modules may be applied to the present invention.

8 is a diagram illustrating a user speech processing sequence by the speech processing unit 104 according to an embodiment of the present invention.

When user utterance is input through the utterance input unit 101, first, the peer classifier 501 determines whether the input user utterance has a peer intent (S801). Subsequently, in step S802, the speech processing unit 104 proceeds to step S803 if the result of determining S801 is equal intent, and proceeds to step S804 if it is not equal intent.

In step S803, the equal divider 503 divides the input utterance into two based on intention. The divided utterance may return to step S801 and repeatedly perform the above-described process.

9 illustrates an example in which a peer splitter 503 divides utterances based on intent according to an embodiment of the present invention.

As an input to the equal divider 503, an equal intention utterance 901 "tell me the weather and then tell me the news" is entered. As an output of the parallel divider 503, a result 902 obtained by dividing the parallel intended utterance 901 into two utterances is output.

First, the peer divider 503 may determine utterance attribute information 903 of the utterance intended to be equal 901 and output it. The utterance attribute information 903 may include at least one of utterance type, intent, and utterance. In the utterance attribute information 903 of the illustrated example, the “parallel” type means that the input utterance is a parallel intended utterance 901 .

The peer divider 503 may output divided utterance attribute information 904-1 and 904-2 for each of the divided utterances. The split utterance attribute information 904-1 and 904-2 includes type information indicating whether the split utterance is a single utterance (if the type is single, it is a single utterance; if the type is parallel, it is a parallel intended utterance), single utterance (if the type is single ), intent information indicating the intention of the corresponding split utterance and utterance information of the split utterance may be included.

Referring again to FIG. 8 , in step S804, the single-subordinate classifier 502 determines whether the user utterance has a single-subordinate intent. In step S805, the utterance processing unit 104 proceeds to step S806 in the case of dependent utterances, and proceeds to step S807 in the case of single-intent utterances.

In one embodiment of the present invention, it is proposed to further determine whether a single intention or a dependent intention is determined in step S804, as well as a specific intention. For example, the single-subordinate classifier 502 not only determines that the utterance "Find a recipe for kimchi stew and send it to me via KakaoTalk" belongs to the dependent intention utterance, but also determines that the utterance "Find a recipe for kimchi stew" belongs to "MI-findcook". belongs to the first intention, and the utterance “Send it to me via KakaoTalk” belongs to the second intention called “MI-readyToSendTalk”.

In step S806, the dependent divider 504 divides the dependent intent utterance into two based on the intent.

10 illustrates an example in which a dependent splitter 504 splits a dependent intent utterance according to one embodiment of the present invention.

As an input to the subordinate divider 504, a subordinate intention utterance 1001 "Find a recipe for kimchi stew and send it to me via KakaoTalk" is entered. As an output of the dependent divider 504, a result 1002 obtained by dividing the dependent intended utterance 1001 into two utterances is output.

First, the dependent divider 504 may determine utterance attribute information 1003 for the dependent intended utterance 1001 and output it. The utterance attribute information 1003 may include at least one of utterance type, intent, and utterance. In the utterance attribute information 1003 of the illustrated example, the “multi-sub” type means that the input utterance is a dependent intended utterance 1001.

The subordinate divider 504 may output divided utterance attribute information 1004-1 and 1004-2 for each of the divided utterances. The split utterance attribute information 1004-1 and 1004-2 is a type indicating whether the split utterance is a data part utterance or an action part utterance (if the type is partial-data, a data part utterance; if the type is partial-action, an action part utterance). It may include (type) information, intent information representing the intention of utterance of the data part/utterance of the action part, and utterance information of divided utterances.

Referring to the output result 1002 of the subordinate divider 504, it is possible to check whether each divided utterance is a data part utterance or an action part utterance as well as a spoken phrase divided according to the intention.

Referring again to FIG. 8 , if it is determined that the utterance is a single intention utterance in step S804, the final output may be constructed immediately in step S807.

11 illustrates an example in which the speech processing unit 104 processes single-intent speech according to an embodiment of the present invention.

As an input to the speech processing unit 104, a single intentional speech 1101 "Will you tell me a fortune?" Further, as an output of the speech processing unit 104 , speech attribute information 1103 of the single-intent speech 1101 is output.

The utterance attribute information 1103 may include at least one of utterance type, intent, and utterance. The “single” type in the utterance attribute information 1003 of the illustrated example means that the input utterance is a single intention utterance 1101 .

On the other hand, when the equivalent intention and the subordinate intention coexist, the above-described processes of FIGS. 9 and 10 may be sequentially performed.

FIG. 12 illustrates a conceptual diagram of processing of the speech processing unit 104 when an equivalent intention and a subordinate intention coexist according to an embodiment of the present invention.

First, as an input to the equal divider 503, an equal-subordinate intention mixed utterance 1201, "Please look for a fortune, find a recipe for kimchi stew, and send it to me by text message" is entered. Also, as an output of the equal divider 503, a result 1202-1 obtained by dividing two equal intention utterances is output. The split result 1202-1 of the two equivalent intent utterances may be input to the subdivision splitter 504 again. And as a result 1202-2 of the subordinate splitter 504, it is divided into two subordinate intent utterances.

13 is a diagram illustrating a configuration of a natural language processing device 100 according to an embodiment.

Referring to FIG. 13 , the natural language processing device 100 includes a processor 1301 and a memory 1302 . Memory 1302 stores one or more instructions executable by processor 1301 . Processor 1301 executes one or more instructions stored in memory 1302 . Processor 1301 may execute one or more operations described above with respect to FIGS. 1-12 by executing instructions. In addition, the configuration of the present invention described above together with FIG. 1 may be a configuration implemented by instructions executed by the processor 1301.

Although the embodiments of the natural language processing apparatus and control method according to the present invention have been described above, this is described as at least one embodiment, and thus the technical idea of the present invention and its configuration and operation are not limited. The scope of the technical idea of the invention is not limited / limited by the drawings or the description referring to the drawings. In addition, the concepts and embodiments of the present invention presented in the present invention can be used by those skilled in the art as a basis for modifying or designing other structures to achieve the same purpose of the present invention. , Modified or changed equivalent structure by a person of ordinary skill in the art to which the present invention belongs is bound by the technical scope of the present invention described in the claims, and does not depart from the spirit or scope of the invention described in the claims. Various changes, substitutions and modifications are possible within the limits.

Claims (17)

In the natural language processing device,
memory for storing instructions; and
a processor configured to execute the stored instructions, the processor comprising:
Receive user input,
storing a first corpus and a second corpus;
Receive connection information between the first and second corpus,
Creating a combined corpus, which is a corpus combining the first and second corpora, based on the input connection information;
Train at least one natural language processing model based on the combinatorial corpus;
Processing the input user utterance based on the at least one natural language processing model;
natural language processing device.
According to claim 1,
The processor divides the input user speech into a plurality of speeches based on a plurality of intentions included in the input user speech.
natural language processing device.
The method of claim 2, wherein the processor,
Determining whether the input user utterance is a paired intended utterance;
Dividing the input user utterance into a plurality of partial equivalent intended utterances based on the equal intended utterance based on the equal intended utterance,
natural language processing device.
The method of claim 3, wherein the processor,
If it is determined that the utterance is not the equivalent intended utterance, it is determined whether the input user utterance is a dependent intended utterance;
Dividing the input user utterance into a plurality of partial dependent intended utterances based on the dependent divider when it is determined that the dependent intended utterance is the dependent intended utterance.
natural language processing device.
The method of claim 4, wherein the plurality of partially dependent intended utterances,
Including data part dependent intent utterances and action part dependent intent utterances,
natural language processing device.
According to claim 1,
The first corpus is a data part corpus, and the second corpus is an action part corpus,
natural language processing device.
The method of claim 6, wherein the connection information,
Including information identifying at least one of the first and second corpus,
natural language processing device.
The method of claim 7, wherein each of the first and second corpora includes a plurality of spoken phrases,
The combined corpus is a corpus obtained by extracting and matching one spoken phrase included in the first corpus and one spoken phrase included in the second corpus,
natural language processing device.
receiving a user's speech;
storing the first corpus and the second corpus;
receiving connection information between the first and second corpus;
generating a combined corpus, which is a corpus obtained by combining the first and second corpus, based on the input connection information;
training at least one natural language processing model based on the combination corpus; and
Processing the input user utterance based on the at least one natural language processing model,
Control method of natural language processing device.
The method of claim 9, wherein the processing of the user utterance comprises:
Dividing the input user speech into a plurality of speeches based on a plurality of intentions included in the input user speech,
Control method of natural language processing device.
11. The method of claim 10, wherein processing the user utterance comprises:
determining whether the inputted user utterance is a peer-intent utterance; and
Further comprising the step of dividing the input user utterance into a plurality of partial equivalent intended utterances if it is determined that it is the equivalent intended utterance.
Control method of natural language processing device.
The method of claim 11, wherein the processing of the user utterance comprises:
determining whether the input user utterance is a dependent intended utterance if it is determined that the utterance is not a peer intended utterance; and
Further comprising the step of dividing the input user utterance into a plurality of partial dependent intended utterances, if it is determined that the dependent intended utterance is determined.
Control method of natural language processing device.
13. The method of claim 12, wherein the plurality of partially dependent intended utterances,
Including data part dependent intent utterances and action part dependent intent utterances,
Control method of natural language processing device.
According to claim 9,
The first corpus is a data part corpus, and the second corpus is an action part corpus,
Control method of natural language processing device.
The method of claim 14, wherein the connection information,
Including information identifying at least one of the first and second corpus,
Control method of natural language processing device.
16. The method of claim 15, wherein each of the first and second corpora includes a plurality of spoken phrases,
The combined corpus is a corpus obtained by extracting and matching one spoken phrase included in the first corpus and one spoken phrase included in the second corpus,
Control method of natural language processing device.
A computer program stored in a medium to execute the method of any one of claims 9 to 16 in combination with hardware.

Images