KR20200080951A - Dialogue system having multiple dialogue domains - Google Patents

Abstract

Provided is a dialogue system having multiple dialogue domains. According to the present invention, the dialogue system having multiple dialogue domains comprises: a user interface receiving a user comment from a user through a network and transmitting a system comment to the user; a dialogue domain management unit having multiple dialogue domains; a user comment preprocessing unit referring to a natural language understanding unit having action ontology and fact ontology to perform natural language processing on the user comment and configure and vectorize the user comment into a semantic frame and using a result of natural language processing for the user comment to select one of the multiple dialogue domains; a natural language analysis unit detecting an entity, an intention, and a substitute term from the semantic frame for the configured user comment and searching for and detecting an entity indicated by the detected substitute term; and a system comment execution unit generating a system comment based on the entity referring to the substitute term of the user comment.

Description

Dialogue system having multiple dialogue domains}

The present invention relates to a conversation system, and more particularly, to a conversation system having a plurality of conversation domains.

The present invention was derived from a study conducted by Bitlux Co., Ltd. and jointly researched by Saltlux Co., Ltd. as part of the SW Computing Industry Source Technology Development Project (SW) by the Ministry of Science, ICT and Future Planning. [Research period: 2018.04.01~2019.03.31, Research management agency: Information and Communication Technology Promotion Center, Research project name: Machine learning-based medical HER cloud service development, task identification number: 2017-0-03355]

The conversation system is a system that generates and provides a system speech in response to a user's utterance, and the conversation system is provided with a conversation domain for providing a conversation environment possible in a specific domain or situation. When the user's remark includes an anaphora such as a person pronoun, an instructive pronoun, or a relational pronoun, in the conversation system, it may be difficult to understand the meaning of the surrogate and to understand the user's intention.

The technical problem of the present invention is to provide a conversation system having a plurality of conversation domains so as to grasp the meaning of a substitute term.

A conversation system having a plurality of conversation domains according to an aspect of the technical idea of the present invention for achieving the above technical problem includes: a user interface that receives a user's speech from a user through a network and transmits the system's speech to the user; A conversation domain management unit having a plurality of conversation domains; By referring to the natural language understanding unit having an action ontology and a fact ontology, natural language processing is performed on the user speech, and the user speech is composed and vectorized into a semantic frame, and the plurality of conversation domains are used by using the natural language processing result for the user speech. A user speech preprocessing unit selecting one of the two; A natural language analysis unit that detects entities, intentions, and substitutes from the constructed semantic frame for the user's speech, and searches for and detects the entities indicated by the detected substitutes; And a system utterance executing unit executing to generate the system utterance based on an entity indicating a substitute word of the user utterance.

The conversation domain management unit may further include a cluster model generation unit configured to construct and vectorize domain conversations of each of the plurality of conversation domains into a semantic frame to generate a single cluster model for each of the plurality of conversation domains.

The natural language analysis unit includes an entity detection unit, an intention detection unit, and a surrogate decomposition unit that detect entities, intentions, and surrogates from the semantic frame for the configured user speech, and the surrogate decomposition unit is detected by the entity detection unit A candidate selection unit for selecting candidate substitutes that are candidates for the detected substitutes from the entity, and candidate replacement units for replacing the detected substitutes with the candidate substitutes; And a similarity calculating unit calculating the similarity between the user's speech and the selected conversation domain in which the detected substitute is replaced with the candidate substitutes, with reference to the sensed intention.

The similarity calculating unit may calculate a distance between the vectorized user speech and a single cluster model for the selected conversation domain.

A semantic frame context storage unit for storing a semantic frame context having a semantic frame, a sensed entity, a sensed intention, and a sensed substitute for the user's speech; and further comprising the candidate selection unit storing the semantic frame context The candidate substitutes may be selected from the sensed entity of the semantic frame context stored in the unit.

The surrogate decomposition unit further includes a cluster model selection unit, and the cluster model selection unit selects the selected conversation domain when the similarity between the user speech and the selected conversation domain substituted with the candidate substitution terms is greater than or equal to a preset value. , When less than the preset value, calculates the similarity between the user's speech substituted with the candidate substitutes and the unselected chat domain among the plurality of chat domains, for an unselected chat domain equal to or greater than the preset value. A single cluster model and a single cluster model for the selected conversation domain can be combined to generate a multiple cluster model.

When the multi-cluster model is generated by the cluster model selection unit, the user remark pre-processing unit may allow the generated multi-cluster model to be selected instead of the single cluster model of the selected conversation domain.

The cluster model generation unit includes: a model semantic frame conversion unit configured to configure each of the domain conversations of each of the plurality of conversation domains as a semantic frame with reference to the action ontology and the fact ontology; A model vector generator for generating action vectors and fact vectors through keywords corresponding to the action ontology and the fact ontology in the semantic frame for each of the configured domain conversations; And a semantic frame vector with reference to the action vector and the fact vector generated in the semantic frame for each of the configured domain conversations, and combining the semantic frame vector for each of the domain conversations to each of the plurality of conversation domains. And a semantic frame vector generator for generating a single cluster model for.

The system speech execution unit includes a standard execution unit and a surrogate execution unit, and the standard execution unit, when there is no substitute word detected in the user speech, based on the entity sensed in the user speech and the sensed intention And the surrogate execution unit may execute to generate the system remark when there is a surrogate detected in the user remark.

The conversation system according to the present invention can grasp the meaning of a substitute word and a user's intention even when the user's speech includes a substitute word, and even if the conversation domain does not have a conversation domain corresponding to all conversation processes, two different By creating a multi-cluster model that combines the above-mentioned dialogue domains, it can directly respond to a new type of dialogue process.

1 is a block diagram illustrating a conversation system having a plurality of conversation domains according to an exemplary embodiment of the present invention.
2A is a block diagram showing the configuration of a conversation domain management unit of a conversation system having a plurality of conversation domains according to an exemplary embodiment of the present invention, and FIG. 2B is a block diagram for explaining the operation of the conversation domain management unit.
3A and 3B are conceptual views illustrating a semantic frame and a semantic frame context used in a conversation system having a plurality of conversation domains according to an exemplary embodiment of the present invention.
Fig. 4 is a block diagram illustrating an operation of a natural language analysis unit of a conversation system having a plurality of conversation domains according to an exemplary embodiment of the present invention.
5A is a block diagram illustrating a process of generating a multi-cluster model in a conversation system having a plurality of conversation domains according to an exemplary embodiment of the present invention, and FIG. 5B is a conceptual diagram illustrating the concept of a multi-cluster model .
Fig. 6 is a block diagram illustrating an operation of a system speaking execution unit of a conversation system having a plurality of conversation domains according to an exemplary embodiment of the present invention.
7 illustrates an exemplary conversation progression made in a conversation system having a plurality of conversation domains according to an exemplary embodiment of the present invention.

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings. The embodiments of the present invention are provided to more fully describe the present invention to those skilled in the art. The present invention can be applied to various changes and may have various forms, and specific embodiments will be illustrated in the drawings and described in detail. However, this is not intended to limit the present invention to a specific disclosure form, and it should be understood that all modifications, equivalents, and substitutes included in the spirit and scope of the present invention are included. In describing each drawing, similar reference numerals are used for similar components. In the accompanying drawings, the dimensions of the structures are enlarged or reduced than actual ones for clarity of the present invention.

The terms used in this specification are only used to describe specific embodiments, and are not intended to limit the present invention. Singular expressions include plural expressions, unless the context clearly indicates otherwise. In this specification, terms such as “include” or “have” are intended to indicate that a feature, number, step, operation, component, part, or combination thereof described on the specification exists, and that one or more other features are present. It should be understood that the existence or addition possibilities of fields or numbers, steps, actions, components, parts or combinations thereof are not excluded in advance.

Unless otherwise defined, all terms used herein, including technical or scientific terms, have the same meaning as commonly understood by a person skilled in the art to which the present invention pertains. Terms such as those defined in a commonly used dictionary should be interpreted as having meanings consistent with meanings in the context of related technologies, and are not to be interpreted as ideal or excessively formal meanings, unless explicitly defined herein. .

In the following drawings and descriptions, one block, for example, a component indicated or described as'~ unit' or'~ module' may be a hardware block or a software block. For example, each of the components may be an independent hardware block that exchanges signals with each other, or may be a software block executed in one processor.

In order to fully understand the configuration and effects of the present invention, preferred embodiments of the present invention will be described with reference to the accompanying drawings.

1 is a block diagram illustrating a conversation system having a plurality of conversation domains according to an exemplary embodiment of the present invention.

Referring to FIG. 1, a conversation system having a plurality of conversation domains (1, hereinafter a conversation system) receives a utterance from the user 10 through the network 50 and transmits a utterance to the user (UI, 100). The network 50 includes wired Internet service, local area network (LAN), broadband communication network (WAN), intranet, wireless Internet service, mobile computing service, wireless data communication service, wireless Internet access service, satellite communication service, wireless LAN, Bluetooth, etc. It can include anything that can send and receive data over wired or wireless. When the network 50 is connected to a smartphone or tablet, the network 50 may be a wireless data communication service such as 3G, 4G, 5G, wireless LAN such as Wi-Fi, Bluetooth, or the like.

The user interface 100 may provide an interface for accessing the conversation system 1 through a terminal or the like used by the user 10. The user 10 may transmit a utterance of the user 10 to the conversation system 1 through the user interface 100, and the conversation system 1 provides through the user interface 100 The system's comments on the user's comments can be received. It is a sentence that the user 10 or the conversation system 1 wants to convey to the other person in the process of speaking. The user's speech can be called an input sentence, and the system's speech can be called an output sentence.

The conversation system 1 includes a conversation execution manager 200 connected to the user interface 100 to receive an input sentence from the user 10 and transmit an output sentence to the user 10 through the user interface 100. do. In addition, the conversation execution manager 200 may control the components included in the conversation system 1 to proceed with the conversation process.

The conversation system 1 includes a user speech pre-processing unit 300. The user speech preprocessing unit 300 may perform natural language processing on the user speech by referring to the natural language understanding unit 500. Natural language processing is a morphological analysis process that analyzes a target word into a morpheme that is a minimum semantic unit, and selects the most suitable form of part-of-speech among the analyzed morpheme results. Part-of-speech process to select suitable analysis results, phrase unit analysis process to analyze chunks of noun phrases, verb phrases, adverbs phrases, clause-by-clause analysis procedures to break sentences such as Chinese and complex sentences into single-unit units, and constituent components constituting sentences It may include a syntax analysis process to determine the structure of the sentence by analyzing the hierarchical relationship.

The user speech pre-processing unit 300 may select any one of a plurality of conversation domains 610, 620, and 630 of the conversation domain management unit 600 using the natural language processing result of the user speech. Each of the plurality of conversation domains 610, 620, and 630 may be a conversation model configured for a conversation situation. The user speech pre-processing unit 300 converts the user speech into a semantic frame (SF in FIG. 3A), a user speech semantic frame change unit 310, and a speech vector generator that vectorizes the user speech converted into the semantic frame SF. 320). The user speech pre-processing unit 300 may convert the user speech into a semantic frame SF and vectorize it by utilizing the action ontology 510 and the fact ontology 520 of the natural language understanding unit 500.

The semantic frame (SF) is a theory that connects linguistic semantics to encyclopedic knowledge, and utilizes the action ontology 510 and the fact ontology 520 in the natural language processing of speech. It is a standardized frame that can extract and map each entity and verb. For example, the language Sell is incomprehensible without the context of buyers, sellers, products, prices, etc., so the semantic frame (SF) is composed of frames in order to interpret the language. The semantic frame (SF) may be applied to lexeme (the smallest unit of meaning), but is not limited thereto, and applied to grammatical construction as a core semantic analysis element, complex linguistic units, or constructive grammar ( construction grammar).

The semantic frame storage unit 550 stores a mapping frame for converting a user's speech into a semantic frame (SF). In this specification, the semantic frame SF may mean a result of mapping each entity and verb extracted from the remark to a mapping frame stored in the semantic frame storage unit 550.

Ontology refers to a technology that identifies and conceptualizes the characteristics of objects or events in the world and makes them into a kind of database, and clearly defines and describes in detail the concept of a resource that is the target of an information system. It can help to find accurate information about related objects and events. The ontology may be composed of four types, for example, Class, Instance, Relation, and Property. Class refers to "computer", "keyboard", and "monitor", which are the names we usually give to things or concepts. Instance is a practical form of class, such as "Samsung Electronics SM-500 computer", "LG Electronics X keyboard", "Samsung Electronics SM-520 monitor", etc., can be generally viewed as instances. have. A property is a value that contains specific properties or information of a class or instance. In the Samsung Electronics 12-inch SM-520 monitor, the instance "Samsung Electronics SM-520 monitor" has a property of 12 inches. Relation may express a relationship between a class and an instance. As in "People are Animals", the "Person" class and the "Animal" class can be expressed as ""Person" isA "Animal" through a relation such as'isA'. Can.

Action Ontology (510) is a data set that defines verbs used in everyday life. The action ontology 510 is an organic data set that has connectivity between data, and can determine similar words, counterwords, and personalities for each verb through such connectivity, and perform intention analysis through verbs.

Fact Ontology (520) is a data set in which nouns and entities used in daily life are defined. Similar to the action ontology 510, nouns and entities stored in the data set constituting the fact ontology 520 have connectivity with each other, and by utilizing such connectivity, synonyms and counterwords and corresponding entities have All available attribute information can be used.

The operation of the speech vector generation unit 320 is almost similar to the model vector generation unit 654 to be described with reference to FIGS. 2A and 2B, and detailed description thereof will be omitted. The model vector generation unit 654 vectorizes the domain conversation (DH) of each of the conversation domains 610, 620, and 630 converted to the semantic frame (SF), and the speech vector generation unit 320 is the semantic frame (SF) You can vectorize the user comments converted to. In some embodiments, any one of the speech vector generator 320 and the model vector generator 654 may be omitted. For example, the user speech pre-processing unit 300 does not include the speech vector generator 320 separately, and calls the model vector generator 654 to vectorize the user speech converted to the semantic frame SF. have. Or, for example, the cluster model generator 650 does not include the model vector generator 654 separately, and calls the speech vector generator 320 to generate a domain conversation (DH) converted into a semantic frame (SF). Can be vectorized.

Similarly, in some embodiments, any one of the user-speaking semantic frame changing unit 310 and the model semantic frame converting unit (652 in FIG. 2A) may be omitted. For example, the user speech pre-processing unit 300 does not include the user speech semantic frame change unit 310 separately, and calls the model semantic frame conversion unit 652 to convert the user speech to the semantic frame SF. Can. Or, for example, the cluster model generation unit 650 does not include the model semantic frame conversion unit 652 separately, and calls the user-spoken semantic frame change unit 310 to transmit the domain conversation (DH) to the semantic frame (SF). Can be converted to

The natural language analysis unit 400 may include an entity detection unit 410, an intent detection unit 420, and an anaphora decomposition unit 430.

The entity detection unit 410 may detect entities from a natural language processing result of a user's speech. Entities exist to set vocabulary keywords used in the conversation process. Entities can be set in a conversation domain for a substitute term in four types: Thing, Location, Time, and People. Thing is an object, Location is a location and place, Time is a time and date, People is an entity related to a person, and can be defined in a conversation domain for a substitute.

The intention detection unit 420 may detect an intention from a natural language processing result of a user's speech. Intention is used as the intention of a speech, and multiple paraphrases can be set for one intention. Intent Classification (Intent Classification or Intent Detect) refers to the process of finding a sentence with high similarity among intentions set in a conversation process to determine the meaning of a specific sentence.

The surrogate decomposition unit 430 may detect an anaphora from a natural language processing result of a user's speech, and search for an entity indicated by the surrogate. The pronoun may be a noun referring to a noun or a noun phrase mentioned above.

The conversation domain manager 600 may include a plurality of conversation domains 610, 620, and 630, and a plurality of single cluster models SCM1, SCM2, and SCM3 generated for each of the conversation domains 610, 620, and 630. have. The cluster model generation unit 650 generates single cluster models SCM1, SCM2, and SCM3 for each of the conversation domains 610, 620, and 630, and at least two of a plurality of single cluster models SCM1, SCM2, and SCM3. With reference to, a multiple cluster model (CCM in FIG. 5B) may be configured.

The surrogate decomposition unit 430 may refer to the semantic frame context (SFC of FIG. 3B) stored in the semantic frame context storage unit 750 in order to search for an entity for a surrogate when a surrogate appears in the user's speech. .

Semantic frame context (SFC) is semantic frame (SF) for conversations (user remarks and system remarks) made through conversation domains 610, 620, and 630, and recognition information for each surrogate (detected) It may have information about a substitute term (DA in FIG. 3B), a sensed entity (DE in FIG. 3B), and a sensed intention (DI in FIG. 3B).

The term decomposition unit 430 extracts an entity valid for the current term by referring to Thing, Location, Time, and People, which are the type information of the entities of the conversation domains 610, 620, and 630 in the semantic frame context (SFC). can do. By replacing the sensed entity DE with a surrogate, the intention detection unit 420 performs intention detection to extract an entity having a higher degree of similarity. The reliability of the conversation may be increased by inversely asking the user 10 from the system remark execution unit 700 based on the information on the substitute word DA detected by the substitute term decomposition unit 430.

The system remark generating unit 800 may generate remarks of a system, such as a reverse question to be performed by the system remark execution unit 700, in a natural language with reference to the natural language generation unit 850, and the remarks of the system generated in the natural language May be delivered to the user 10 through the user interface 100 by the conversation execution manager 200.

2A to 7, detailed configuration and operation of the conversation system 1 having a plurality of conversation domains will be described with reference to FIG. 1 together.

2A is a block diagram showing the configuration of a conversation domain management unit of a conversation system having a plurality of conversation domains according to an exemplary embodiment of the present invention, and FIG. 2B is a block diagram for explaining the operation of the conversation domain management unit. In FIG. 2A, only one conversation domain (conversation domains 1, 610, hereinafter referred to as the first conversation domain) among the plurality of conversation domains 610, 620, and 630 shown in FIG. 1 is shown for convenience of illustration. The description of FIG. 2B can be equally applied to other conversation domains, for example, the second conversation domain (conversation domains 2 and 620) and the third conversation domain (conversation domains 3 and 630).

2A and 2B, the conversation domain management unit 600 includes a first conversation domain 610 and a cluster model generation unit 650. The first conversation domain 610 may include a first domain conversation storage unit (domain conversation storage units 1, 612) in which domain conversations DH are stored. Although not separately illustrated, each of the second conversation domain 620 and the third conversation domain 630 may include a second domain conversation storage unit and a third domain conversation storage unit.

The cluster model generation unit 650 may include a model semantic frame conversion unit 652, a model vector generation unit 654, and a semantic frame vector generation unit 656.

The model semantic frame conversion unit 652 has an action ontology 510 that the natural language understanding unit 500 has domain conversations DH stored in the first domain conversation storage unit 612 with respect to the first conversation domain 610. And the fact ontology 520 may be used to map to a semantic frame (SF).

The model vector generation unit 654 may generate an action vector (AV) and a fact vector (FV) through keywords corresponding to the action ontology 510 and the fact ontology 520 configured in the semantic frame SF.

The semantic frame vector generation unit 656 may obtain a semantic frame vector, which is a vector value referring to the action vector (AV) and the fact vector (FV), for the semantic frame SF constituting each of the domain conversations DH, Semantic frame vectors for each of the domain conversations DH (SSFV in FIG. 5B, semantic frame vector (set 1)) are combined and stored in the first domain conversation storage 612 of the first conversation domain 610. A first cluster model (cluster model 1, SCM1) that is a cluster model for domain conversations (DH) may be generated.

In the same way, the second cluster model (cluster model 2, SCM2) for the second conversation domain 620 shown in FIG. 1 and the third cluster model (cluster model 3, SCM3) for the third conversation domain 630 are shown. Can be created.

The cluster model for one conversation domain may be called a single cluster model. For example, the single cluster model for the first chat domain 610 is the first cluster model (SCM1), the single cluster model for the second chat domain 620 is the second cluster model (SCM2), and the third The single cluster model for the conversation domain 630 may be a third cluster model (SCM3). Conversely, a cluster model for two or more conversation domains may be referred to as a multi-cluster model (CCM in FIG. 5B).

3A and 3B are conceptual views illustrating a semantic frame and a semantic frame context used in a conversation system having a plurality of conversation domains according to an exemplary embodiment of the present invention.

Referring to FIG. 3A, when a user speech is input, in the natural language processing process of the user speech in the user speech pre-processing unit 300 referring to the natural language understanding unit 500, the user speech semantic frame changing unit 310 is the semantic frame (SF) can be configured. The composition of the semantic frame SF means a method of extracting entities and verbs, respectively, and mapping them to a standardized frame by using the action ontology 510 and the fact ontology 520 for user speech.

When "I have an appointment at Gangnam Station tomorrow morning" is input as a user's remark, the verb "I have an appointment" is extracted, and as an'action (promise)' through the action ontology 510, {Agent: "Me", Location : "Gangnam Station", Time: "Tomorrow Morning"}. Attributes, each attribute maps semantic type, synonym, etc. for keywords by using fact ontology 520, semantic frame (SF) Can be configured.

The action vector (AV in FIG. 2B) is a vectorization of the attributes of the verb “action” (promise), through the action ontology 510, and is mapped to keywords by using the fact ontology 520. The vectorized attributes are the fact vector (FV).

Semantic frames (SF), action vectors (AV), and fact vectors (FV) may be generated for user speech and domain conversation (DH in FIG. 2B). Specifically, the semantic frame (SF), the action vector (AV), and the fact vector (FV) for the user speech can be generated by the user speech semantic frame change unit 310 and the speech vector generator 320 shown in FIG. 1. The semantic frame (SF), action vector (AV), and fact vector (FV) for the domain conversation (DH) are generated by the model semantic frame conversion unit 652 and the model vector generation unit 654 shown in FIG. 2A. can do.

Referring to FIG. 3B, semantic frames (SF) generated for speeches in a conversation process, sensed intentions (DI) for each speech, sensed entities (DEs), and sensed substitutes (DA) Information about can be configured as a semantic frame context (SFC).

That is, one semantic frame (SF) is configured for one speech of the user 10 or the system, and the semantic frame context (SFC) can be configured for one conversation process proceeding through one selected conversation domain. .

Fig. 4 is a block diagram illustrating an operation of a natural language analysis unit of a conversation system having a plurality of conversation domains according to an exemplary embodiment of the present invention.

Referring to FIG. 4, the natural language analysis unit 400 includes an entity detection unit 410, an intention detection unit 420, and a substitute term decomposition unit 430. When a substitute word is not found in the user's speech, the entity detected by the entity detection unit 410 and the intention detected by the intention detection unit 420 are transmitted to the standard execution unit 710 of the system speech execution unit 700, and the system The speech of the system to be generated by the speech execution unit 700 may be generated in natural language with reference to the natural language generation unit 850 in the system speech generation unit 800 shown in FIG. 1, and the speech of the system generated in the natural language is The conversation execution manager 200 may be delivered to the user 10 through the user interface 100.

The term decomposition unit 430 may include a candidate selection unit 432, a candidate substitution unit 434, a similarity calculation unit 436, and a cluster model selection unit 438.

When a substitute word is found in the user's remark, the candidate selection unit 432 is entities that can be candidates for the substitute word through the semantic frame context (SFC of FIG. 3B) stored in the semantic frame context storage unit 750. A list of candidate substitutes can be selected.

The candidate replacement unit 434 replaces each of the entities that are candidate candidate substitutes, and the similarity calculator 436 replaces the user's speech and the selected conversation domain, for example, the first conversation domain (for example, the first conversation domain). Similarity calculation between the first cluster model SCM1 for 610) may be performed. The similarity calculation unit 436 calculates the distance between the vectored user speech and the first cluster model SCM1 in the speech vector generation unit 320, so that the user speech and the selected conversation domain, for example, the first conversation domain 610 ) Can be calculated, and the intention detected by the intention detector 410 can be referenced. Also, the similarity calculation unit 436 may include a similarity between the user's speech substituted with each of the candidate clusters and the second cluster model SCM2 for the unselected chat domain, for example, the second chat domain 620, and/or Alternatively, the similarity between the third cluster model SCM3 for the third conversation domain 630 and the user's speech substituted with each of the candidate substitutes may be calculated.

The cluster model selection unit 438 selects a cluster model based on the similarity calculation result performed by the similarity calculation unit 436 and transmits it to the large-word execution unit 720 of the system speech execution unit 700. The speech of the system to be generated by the execution unit may be generated in natural language with reference to the natural language generation unit 850 in the system speech generation unit 800 shown in FIG. 1, and the speech of the system generated in the natural language may be a conversation execution manager ( 200, it may be delivered to the user 10 through the user interface 100.

5A is a block diagram illustrating a process of generating a multi-cluster model in a conversation system having a plurality of conversation domains according to an exemplary embodiment of the present invention, and FIG. 5B is a conceptual diagram for explaining the concept of a multi-cluster model .

Referring to FIG. 5A, when a user speech (UU) is input, one conversation domain is selected from the natural language processing result for the user speech (UU) (S100). For example, the first conversation domain 610 may be selected from the natural language processing result for the user's speech (UU). When the first conversation domain 610 is selected, the first cluster model SCM1 of the first conversation domain 610 may be used in the conversation process.

When a substitute word is detected in the user's speech (UU), a list of candidate substitute words that are candidate entities for the substitute word is selected from the semantic frame (SF) of the semantic frame context (SFC of FIG. 3B ). And replaces the selected candidate substitutes (S200).

The similarity between each of the user speeches (UUs) substituted with the substitutes sensed as candidate substitutes and the selected conversation domain is calculated (S300). For example, the similarity may be calculated by measuring the distance between each of the vectorized user speeches (UUs) in which the substitutes detected as candidate substitutes are replaced, and the first cluster model (SCM1) of the first conversation domain 610. Can. When the calculated similarity is equal to or greater than a preset value, the system speech may be executed and generated with reference to the selected conversation domain, for example, the first conversation domain 610 (S600).

When the calculated similarity is less than a predetermined value, the similarity between each of the vectorized user speech (UUs), each of which is substituted with the substitutes sensed as candidate substitutes, and each of the unselected conversation domains is calculated (S400). For example, the similarity between each of the vectorized user speeches (UUs) in which the substitutes detected as candidate substitutes are replaced with the second cluster model (SCM2) of the second conversation domain 620, candidate substitutes Similarity is calculated between each of the vectorized user speeches (UUs), each of which has been substituted with a third term, and the third cluster model (SCM3) of the third conversation domain 630.

Among the unselected conversation domains, the calculated similarity is selected to be greater than or equal to a preset value, and a multiple cluster model is generated together with the selected conversation domain (S500). For example, when the similarity between each of the user utterances (UUs), each of which is detected as candidate substitutes, and the second cluster model (SCM2) of the second conversation domain 620 is equal to or greater than a preset value, Combining the first cluster model (SCM1) of the first chat domain (610), which is the selected chat domain, and the second cluster model (SCM2) of the second chat domain (620), creates a multiple cluster model (CCM in FIG. 5B). can do.

Thereafter, a system remark may be executed and generated using a multi-cluster model (CCM) (S600). In addition, when a multi-cluster model (CCM) is generated, a conversation process may be performed using the multi-cluster model (CCM) generated instead of the first cluster model (SCM1) of the selected first conversation domain 610. For example, the user speech pre-processing unit (300 in FIG. 1) may select a multi-cluster model (CCM) generated instead of the first cluster model (SCM1) of the first conversation domain 610 to be used in the conversation process. That is, the user speech pre-processing unit 300 may change the selected conversation domain from the first conversation domain 610 to the first conversation domain 610 and the second conversation domain 620.

5B, each of the plurality of conversation domains 610, 620, and 630 has domain conversations DH1, DH2, and DH3, and the action ontology 510 and the fact ontology 520 of the natural language understanding unit 500 ) To convert the domain conversations (DH1, DH2, DH3) into a semantic frame (SF in FIG. 3A). The semantic frames SF in which the domain conversations DH1, DH2, and DH3 of each of the plurality of conversation domains 610, 620, and 630 are converted may constitute a set SSF1, SSF2, and SSF3, respectively. That is, the semantic frames SF that convert each of the domain conversations DH1 of the first conversation domain 610 constitute the first semantic frame set SSF1, and the domain conversations of the second conversation domain 620 The semantic frames SF converting each of the fields DH2 constitute a second semantic frame set SSF2, and the semantic frames SF converting each of the domain conversations DH2 of the third conversation domain 620 They may form a third semantic frame set (SSF3). The first to third semantic frame sets (SSF1, SSF2, and SSF3) constitute first to third semantic frame vector sets (SSFV1, SSFV2, and SSFV3) with reference to action vectors (AV) and fact vectors (FV), respectively. Can. The first semantic frame vector set SSFV1 constructed from the first conversational domain 610 may be generated as a first cluster model SCM1, which is a cluster model, and a second semantic frame vector set constructed from the second conversational domain 620. (SSFV2) may be generated as a second cluster model (SCM2), which is a cluster model, and the third semantic frame vector set (SSFV3) constructed from the third conversation domain 630 may be a third cluster model (SCM3), which is a cluster model. Can be generated. Each of the first to third cluster models SCM1, SCM2, and SCM3 is a single cluster model of one conversation domain, that is, each of the first to third conversation domains 610, 620, and 630.

By combining single cluster models for different conversation domains, multiple cluster models (CCM) can be generated. For example, a multi-cluster model (CCM) may be generated by combining the first cluster model (SCM1) and the second cluster model (SCM2), and in the process of executing and generating system speech, a single cluster of a selected conversation domain. Instead of a model, multiple cluster models (CCM) can be used.

Fig. 6 is a block diagram illustrating an operation of a system speaking execution unit of a conversation system having a plurality of conversation domains according to an exemplary embodiment of the present invention.

Referring to FIG. 6, the system speech execution unit 700 includes a standard execution unit 710 and a substitute execution unit 720.

The standard execution unit 710 allows the system to generate a system speech when there is an entity and an intention in the user speech when no substitute is detected.

The surrogate execution unit 720, when a surrogate is detected, generates a remark of the system for the user's remark.

The system's remarks can be generated, for example, by text, smart card, or action script.

The system speech execution unit 700 generates a speech of the system in the substitute speech execution unit 720 or the standard execution unit 710 depending on the presence or absence of a substitute word for each user speech in the conversation process.

7 illustrates exemplary conversation progress in a conversation system having a plurality of conversation domains according to an exemplary embodiment of the present invention.

Referring to FIG. 7, a unit of dialogue consisting of system remarks and user remarks is set to turn, and the conversation process may be repeatedly performed while referring to the contents described with reference to FIGS. 1 to 6. In FIG. 7, the system speech is made first, and the user's speech on the system speech is illustrated as one turn, but the present invention is not limited thereto, and the user's speech and the system speech on the speech may be set as one turn.

For example, when the conversation process starts, a user remarks,'Are you doing a barbecue pizza L size discount event?' for a system remark that is a text'Hello' with a sensed intention of'Greeting' at Turn 0. When this is input, a semantic frame may be generated by mapping the detected entities (Bulgogi Pizza (Thing), discount event (Thing)) and the detected intention (Event Promotions) based on the action of'make'.

A system remark that is text'Yes discount event is in progress' may be output by referring to the semantic frame for the user remark mapped in turn 0 in turn 1 in turn 1 (Turn 1). When a user's remark'How much is that?' is input, the entity detected based on the action of'how much' (if the thing is a substitute,'bulgogi pizza','discount event as candidate substitutes) By selecting'and replacing it, how much is Bulgogi Pizza?' and'How much is the discount event?', similarity can be analyzed to extract entities with higher similarity.

In Turn 2, the system utterance may be generated, for example, as a system utterance, for example, as a text, a smart card, or an action script. In Turn 2, when two entities with high similarity are extracted, a case in which a user asks for text or provides a smart card is exemplified. Text means outputting a system statement such as "What did you say?" to the user, and smart card means outputting a system statement so that the user can select an option. Action scripts can be used when information to output a system remark needs to be fetched through an external API. On the other hand, when a user's remark'I'm talking about bulgogi pizza' is input, an entity detected based on an action called'how much' by referring to the semantic frame context composed of the semantic frames for the previous system remark and user remark A semantic frame may be generated by mapping (Bulgogi Pizza (Thing)) and the sensed intention (Buy Pizza).

In turn 3, create an action script that fetches price and discount information for bulgogi pizza through an external API, and corresponds to the semantic frame generated from the user's remark in turn 2'current 50 % Discount and it is 15,000 won.' If'Give me that' is input as a user's remark, if the entity (Thing) detected based on the action of'Give' is a substitute,'Bulgogi Pizza' as a candidate substitute, ' Discount event','discount', and '15,000 won' are selected, and the similarity is analyzed after replacing it.

As a result of the similarity analysis performed in Turn 3, when only one'Bulgogi Pizza' is extracted as an entity with high similarity, an entity called'Bulgogi Pizza' is combined with the action of'Ordered', resulting in Bulgogi Pizza An order for and a system statement of'It was ordered as a bulgogi pizza' may be output.

Referring to FIGS. 1 to 7 together, the first conversation domain 610 is a domain in which a conversation model for pizza orders is defined, and the second conversation domain 620 is a domain in which conversation models for chicken orders are defined, When the third conversation domain 630 is a domain in which a conversation model for clothing order is defined, FIG. 7 illustrates that the conversation process is performed by selecting the first conversation domain 610.

For example, 1 conversation domain 610 is a conversation model that includes information on a discount event, and the second conversation domain 620 and a 3rd conversation domain 630 are conversation models that do not include information on a discount event. Can be

If the user's remark (UU) for the chicken order is input and the second conversation domain 620 is selected, when the user's remark (UU) for the discount event is entered in the conversation process, the user's remark (UU) and the second When the similarity analysis of the conversation domain 620 is performed, the similarity calculated between the result of vectorizing the user speech (UU) converted to the semantic frame SF and the second cluster model 2 (SCM2) may be less than a preset value. . In this case, by analyzing the similarity between the user talk (UU) and the unselected first chat domain 610 and the third chat domain 630, the first chat domain 610 including information on the discount event When the similarity between user speeches (UU) is equal to or greater than a preset value, the first cluster model (SCM1) of the first chat domain 610 and the second cluster model (SCM2) of the second chat domain 620 are combined, After generating a multi-cluster model (CCM) having information about chicken orders and discount events, a system remark may be executed and generated with reference to the multi-cluster model (CCM).

Accordingly, the conversation system 1 having a plurality of conversation domains according to the present invention can grasp the meaning of a substitute term and the user's intention even when a substitute word is included in the user's speech, and identify a conversation domain corresponding to all conversation processes. Even if you do not have it, you can create a multi-cluster model that combines two or more different conversational domains, so you can immediately respond to a new type of conversational process.

As described above, the present invention has been described in detail with reference to preferred embodiments, but the present invention is not limited to the above embodiments, and various modifications and changes by those skilled in the art within the technical spirit and scope of the present invention. This is possible.

DESCRIPTION OF SYMBOLS 1: Conversation system having multiple conversation domains, 10: user, 50: network, 100: user interface, 200: conversation execution manager, 300: user speech preprocessing unit, 310: user speech semantic frame conversion unit, 320: speech vector Generation unit, 400: natural language analysis unit, 410: entity detection unit, 420: intention detection unit, 430: a large-word decomposition unit, 500: natural language understanding unit, 510: action ontology, 520: fact ontology, 550: semantic frame storage unit, 600 : Conversation domain management unit, 610, 620, 630: conversation domain, 650: cluster model generation unit, 700: system speech execution unit, 750: semantic frame context storage unit, 800: system speech generation unit, 850: natural language generation unit, SCM1 , SCM2, SCM3: Cluster model, CCM: Multiple cluster model, DH: Domain conversation, SF: Semantic frame, AV: Action vector, FV: Fact vector, SFC: Semantic frame context

Claims (9)

A user interface that receives a user's speech from the user via a network and sends a system speech to the user;
A conversation domain management unit having a plurality of conversation domains;
With reference to a natural language understanding unit having an action ontology and a fact ontology, the natural language processing of the user speech is performed, the user speech is composed and vectorized into a semantic frame, and the plurality of conversation domains are obtained using the natural language processing result for the user speech User speech pre-processing unit for selecting one of the;
A natural language analysis unit that detects entities, intentions, and substitutes from the constructed semantic frame for the user's speech, and searches for and detects the entities indicated by the detected substitutes; And
And a system speech execution unit executing to generate the system speech based on an entity indicating a substitute word of the user speech. According to claim 1,
The conversation domain manager further comprises a cluster model generation unit configured to construct and vectorize domain conversations of each of the plurality of conversation domains into a semantic frame to generate a single cluster model for each of the plurality of conversation domains. system. According to claim 2,
The natural language analysis unit includes an entity detection unit, an intention detection unit, and a surrogate decomposition unit that detect entities, intentions, and surrogates, respectively, from the semantic frame for the configured user speech,
The large term decomposition unit,
A candidate selection unit for selecting candidate substitutes that are candidates for the detected substitutes from the entity detected by the entity detection unit,
A candidate replacement unit for substituting the detected substitutes with the candidate substitutes, respectively; And
And a similarity calculating unit for calculating the similarity between the user's speech and the selected conversational domain in which the sensed substitute is replaced with the candidate substitutes, referring to the sensed intention. According to claim 3,
And the similarity calculating unit calculates a distance between the vectorized user speech and a single cluster model for the selected conversation domain. According to claim 3,
Further comprising; a semantic frame context storage unit for storing a semantic frame context having a semantic frame, a sensed entity, a sensed intention, and a sensed synonym configured for the user speech,
The candidate selection unit, the dialogue system, characterized in that for selecting the candidate substitutes from the sensed entity of the semantic frame context stored in the semantic frame context storage unit. According to claim 2,
The term decomposition unit further includes a cluster model selection unit,
The cluster model selection unit,
If the similarity between the user remarks and the selected conversation domains respectively substituted by the candidate substitutes is greater than or equal to a preset value, the selected conversation domain is selected, and if less than the preset value, the users substituted by the candidate substitutes respectively Multi-cluster model by combining the single cluster model for the unselected chat domain and the single cluster model for the selected chat domain by calculating the similarity between the speech and the unselected chat domain among the plurality of chat domains. Characterized in that the dialogue system. The method of claim 6,
When the multi-cluster model is generated by the cluster model selection unit, the user remark pre-processing unit causes the generated multi-cluster model to be selected instead of the single cluster model of the selected chat domain. According to claim 2,
The cluster model generation unit,
A model semantic frame converter configured to configure each of the domain conversations of each of the plurality of conversation domains as a semantic frame with reference to the action ontology and the fact ontology;
A model vector generator for generating action vectors and fact vectors through keywords corresponding to the action ontology and the fact ontology in the semantic frame for each of the configured domain conversations; And
A semantic frame vector is obtained by referring to the action vector and the fact vector generated in the semantic frame for each of the configured domain conversations, and the semantic frame vector for each of the domain conversations is combined to each of the plurality of conversation domains. Semantic frame vector generation unit for generating a single cluster model for; a conversation system having a plurality of conversation domains comprising a According to claim 1,
The system speech execution unit includes a standard execution unit and a substitute execution unit,
The standard execution unit executes to generate the system speech based on the entity and the sensed intention detected in the user speech when there is no substitute for the user speech,
And the surrogate execution unit executes to generate the system remark if there is a surrogate detected in the user remark.

Images