KR20220155957A - Method and apparatus for providing chatbot service based on slot filling - Google Patents

Abstract

Disclosed are a method and an apparatus for providing a chatbot service based on slot filling. The chatbot service providing method according to one embodiment of the present invention comprises the steps of: determining intent corresponding to user conversation; using an intent mapping list to select an execution block corresponding to the intent; performing slot filling corresponding to the execution block; and based on the slot filling, performing a conversation process corresponding to the intent. The intent mapping list is provided for each execution block and corresponds to at least one intent. Accordingly, a chatbot can effectively respond to various user conversation patterns.

Description

Method and apparatus for providing chatbot service based on slot filling {METHOD AND APPARATUS FOR PROVIDING CHATBOT SERVICE BASED ON SLOT FILLING}

The present invention relates to a method and apparatus for providing a chatbot service, and in particular, to a slot filling-based chatbot service providing technology capable of slot filling information that is difficult for a user to say from the beginning or information added while having a conversation with a user. it's about

Unless otherwise indicated herein, material described in this section is not prior art to the claims in this application, and inclusion in this section is not an admission that it is prior art.

A chatbot refers to a computer program or artificial intelligence equipped with the ability to communicate by text or voice. It ranges from simple chatbots that use only word or phrase matching to complex and sophisticated natural language processing technology to conduct natural conversations like humans. There are various levels of chatbots, including chatbots that apply .

A conversational AI agent system using a chatbot identifies the service the user wants in response to the user's input and provides a response. In this case, when the user provides an input through a natural sentence, natural language processing technology may be applied to the input for processing. To this end, several utterance example sentences with the same user intent are collected and classified into one intent. It is possible to provide a chatbot service capable of handling various spontaneous speeches of users by mapping bot responses corresponding to each of the thus classified intents or defining operations such as service API calls.

When one bot response unit (execution block) is mapped for each user's utterance unit (intent), bot responses and corresponding intents increase as service functions increase. It is assigned to intents, and when intent is classified, a plurality of intents having a slight score difference are generated, which increases the possibility of error in intent classification.

Even if the number of intents does not increase, if various utterance patterns are registered in the intent to recognize the user's various natural utterances, similar patterns are registered in multiple intents, so that the natural language processor accurately identifies similar utterances between different intents. Difficult to process, prone to errors in intent classification.

That is, if utterances are classified according to the situation in which intents and bot response units are mapped 1:1, the performance of the natural language processor decreases as the number of intents increases or the number of example utterances registered in the intent increases. occurrence, and when utterances are classified to ensure the performance of the natural language processor, all similar utterance patterns are defined in one intent, which makes the implementation of blocks mapped to the intent very complicated, and many actually provided functions are included in one block. As it becomes a mapping situation, difficulties in management are inevitable.

In addition, most chatbots process bot actions by searching for all intents when the user utters them. In the case of searching for all intents in this way, the problem of processing results of natural language processors and bot responses and actions There is a problem that complicates the configuration.

Therefore, there is an urgent need for a new chatbot service providing technology that can more freely map intents and bot response units.

Korean Patent Publication No. 10-2019-0006403, published on January 18, 2019 (Name: Voice processing method and system supporting it)

An object of the present invention is to enable a chatbot to effectively respond to various user conversation patterns by setting the same intent to operate differently based on slot filling defined in the execution block when the same intent is used in different execution blocks.

In addition, an object of the present invention is to perform slot filling for a completely new slot as well as a slot existing in an intent using several entities mapped to one slot, through which information that is difficult for a user to speak from the beginning. It is to enable slot filling for information added during the conversation.

Also, an object of the present invention is to simplify a process for maintaining a context flow by providing a slot reuse function, and to more effectively manage slot values set for maintaining a context flow.

In addition, an object of the present invention is to enable N:N (N is a natural number) mapping between intents classified as suitable for a natural language processor and execution blocks processing the intents, so that a single It is to be able to execute a function, and conversely, to be able to perform a completely different function with similar utterances.

Also, an object of the present invention is to enable various intent-execution block mapping relationships by using an intent mapping list including an intent identifier and a slot condition for each one or more intents.

In addition, the object of the present invention is not limited to the objects as described above, and it is obvious that another object may be derived from the following description.

To achieve the above object, a method for providing a chatbot service based on slot filling according to an embodiment of the present invention includes determining an intent corresponding to a user conversation; selecting an execution block corresponding to the intent using an intent mapping list; performing slot filling corresponding to the execution block; and performing conversation processing corresponding to the intent based on the slot filling, wherein the intent mapping list is provided for each execution block and corresponds to at least one intent.

In this case, performing slot filling may perform slot filling based on at least one of an entity mapped to a slot, slots corresponding to the execution block, and intent mapping information of the slot.

In this case, the intent mapping information is selected from among result values of the intent name mapped to the slot, the slot name of the intent mapped to the slot, the slot value of the intent mapped to the slot, and the slot value of the intent mapped to the slot. may contain at least one.

In this case, slots corresponding to the execution block may include slots of intents connected to the execution block through the intent mapping list.

In this case, in the step of performing slot filling, a slot value may be extracted by performing synonym matching on an entity mapped to the slot.

In this case, one slot may be mappable with a plurality of entities.

At this time, the step of performing slot reuse configuration for maintaining a slot context in consideration of a scenario including an execution block may be further included.

In this case, the slot values of the intents connected to the execution block for which the slot reuse setting is performed may not be deleted even when the scenario is executed again after the end of the scenario.

In this case, the intent mapping list may include an intent identifier and a slot condition for each intent.

In addition, an apparatus for providing a chatbot service based on slot filling according to an embodiment of the present invention includes one or more processors; and an execution memory storing one or more programs executed by the one or more processors. The at least one program determines an intent corresponding to a user conversation, selects an execution block corresponding to the intent using an intent mapping list, and performs slot filling corresponding to the execution block. and performs conversation processing corresponding to the intent based on the slot filling, wherein the intent mapping list is provided for each execution block and corresponds to at least one intent.

In this case, the at least one program may perform slot filling based on at least one of an entity mapped to a slot, slots corresponding to the execution block, and intent mapping information of the slot.

In this case, the intent mapping information is selected from among result values of the intent name mapped to the slot, the slot name of the intent mapped to the slot, the slot value of the intent mapped to the slot, and the slot value of the intent mapped to the slot. may contain at least one.

In this case, slots corresponding to the execution block may include slots of intents connected to the execution block through the intent mapping list.

In this case, the at least one program may extract a slot value by performing synonym matching on an entity mapped to the slot.

In this case, one slot may be mappable with a plurality of entities.

At this time, the at least one or more programs may perform slot reuse configuration for maintaining a slot context in consideration of a scenario including an execution block.

In this case, the slot values of the intents connected to the execution block for which the slot reuse setting is performed may not be deleted even when the scenario is executed again after the end of the scenario.

In this case, the intent mapping list may include an intent identifier and a slot condition for each intent.

According to the present invention, when the same intent is used in different execution blocks, the chatbot can effectively respond to various user conversation patterns by setting to operate differently based on the slot filling defined in the execution block.

In addition, the present invention can perform slot filling for a completely new slot as well as a slot existing in an intent using several entities mapped to one slot, and through this, information or conversation that is difficult for the user to speak from the beginning Slot filling may also be enabled for information added in progress.

In addition, the present invention simplifies a process for maintaining a context flow by providing a slot reuse function, and can more effectively manage slot values set for maintaining a context flow.

In addition, the present invention enables N:N (N is a natural number) mapping between intents classified as suitable for a natural language processor and execution blocks processing the intents, so that one function can be executed with various different utterances. and conversely, similar utterances can perform completely different functions.

Also, according to the present invention, it is possible to define various intent-execution block mapping relationships by using an intent mapping list including an intent identifier and a slot condition for each one or more intents.

The effects of the present embodiments are not limited to the effects mentioned above, and other effects not mentioned will be clearly understood by those skilled in the art from the description of the claims.

1 is an operational flowchart illustrating a method of providing a chatbot service based on slot filling according to an embodiment of the present invention.
2 is a diagram illustrating an example of an entity according to the present invention.
FIG. 3 is a diagram illustrating an example of an intent to which the entity shown in FIG. 2 is applied.
FIG. 4 is a diagram illustrating an example of intent mapping information based on the intent shown in FIG. 3 .
5 is an operational flow diagram illustrating an example of a slot filling process according to the present invention.
6 is a block diagram illustrating an apparatus for providing a chatbot service based on slot filling according to an embodiment of the present invention.
7 is a block diagram showing a chatbot service system according to an embodiment of the present invention.
8 is a block diagram of a 'coffee order' scenario according to an embodiment of the present invention.
FIG. 9 is a diagram illustrating an example of intents used in the scenario shown in FIG. 8 .
10 is a block diagram showing connections between a plurality of intents and a plurality of execution blocks.
FIG. 11 is a diagram illustrating an example of the execution block coffee selection shown in FIG. 9 .
FIG. 12 is a diagram illustrating an example of hot ice checking of the execution block shown in FIG. 9 .
13 is a diagram illustrating an example of execution block takeout confirmation shown in FIG. 9 .
9 is a block diagram showing a connection relationship between execution blocks shown in FIGS. 9 to 13 .
10 is a diagram illustrating an example of connecting blocks by defining a waiting state for a user response in an execution block.
16 is a diagram illustrating an example of mapping intents defined for each speech pattern to blocks.
17 is a diagram illustrating a case of 1:1 mapping between intents and execution blocks.
18 is a diagram illustrating an example of intent classification and block mapping according to an embodiment of the present invention.
FIG. 19 is a diagram showing execution blocks corresponding to the responses shown in FIG. 27 .
FIG. 20 is a diagram showing an example of priorities sorted according to inclusion relationships of conditions corresponding to the coffee ordering scenarios illustrated in FIGS. 8 to 9 .
21 is a block diagram illustrating an example of modules necessary for executing a chatbot.
22 is a diagram illustrating a process of searching for intents and blocks while the chatbot is running.
23 is an operational flowchart illustrating a method of providing a chatbot service according to another embodiment of the present invention.
24 is a graph showing performance of a natural language processor according to an increase in the number of speech patterns.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. However, detailed descriptions of well-known functions or configurations that may obscure the gist of the present invention will be omitted in the following description and accompanying drawings. In addition, it should be noted that the same components are indicated by the same reference numerals throughout the drawings as much as possible.

The terms or words used in this specification and claims described below should not be construed as being limited to a conventional or dictionary meaning, and the inventors should appropriate the concept of terms to describe their invention in the best way. It should be interpreted as a meaning and concept consistent with the technical spirit of the present invention based on the principle that it can be defined in the following way. Therefore, the embodiments described in this specification and the configurations shown in the drawings are only one of the most preferred embodiments of the present invention, and do not represent all of the technical ideas of the present invention, so various alternatives can be made at the time of this application. It should be understood that there may be equivalents and variations. In addition, terms such as first and second are used to describe various components, and are used only for the purpose of distinguishing one component from another, and are not used to limit the components.

In the conventional chatbot service, slot values are extracted by using only the user's initial utterance or performing synonym matching based on only one entity mapped to the slot. There was a disadvantage that slot filling was difficult for the added information.

In addition, when slot filling is performed, the same intent processes slots identically in execution blocks with different functions simply by setting whether a slot dependent on an intent is required or a prompt message is set. It can also lead to inefficient results.

Therefore, the present invention intends to propose a chatbot service providing technology based on slot filling that can solve the above problems.

1 is an operational flowchart illustrating a method of providing a chatbot service based on slot filling according to an embodiment of the present invention.

Referring to FIG. 1 , the method for providing a chatbot service based on slot filling according to an embodiment of the present invention determines an intent corresponding to a user conversation (S110).

In this case, an intent corresponding to the user dialogue may be selected from one or more candidate blocks connected to the current block through a graph or from among intent candidates connected to the main block.

In this case, the intent corresponding to the user dialogue may be determined by considering the first intent candidate linked to the candidate block with priority over the second intent candidate linked to the main block.

In addition, in the method for providing a chatbot service based on slot filling according to an embodiment of the present invention, an execution block corresponding to an intent is selected using an intent mapping list (S120).

In this case, the intent mapping list is provided for each of the execution blocks and may correspond to one or more intents.

In this case, the intent mapping list may include an intent identifier and a slot condition for each intent.

In this case, the slot condition may correspond to conditional statements related to one or more slots.

At this time, the slot condition may include a null condition.

In this case, the intent mapping list may include intent identifiers and slot conditions corresponding to a plurality of intents (N:1), and execution blocks having different intents have the same intent identifier. If the slot conditions are different, it can be connected (1:N) to different execution blocks.

In this case, the slots corresponding to the execution block may include slots of intents connected to the execution block through the intent mapping list.

At this time, the execution block corresponding to the intent, when the intent corresponding to the user dialogue is connected to a plurality of execution blocks, the first execution block having an inclusion relationship of the slot condition corresponding to the large set is included in the large set. A second execution block having an inclusion relationship of a slot condition corresponding to an included small set may be prioritized and selected. In this case, the inclusion relationship of the slot condition may be the inclusion relationship of slots included in the slot condition. For example, an execution block corresponding to $level == "regular member" and $service == "subscription" may be selected with priority over an execution block corresponding to $service == "subscription".

In addition, the slot filling-based chatbot service providing method according to an embodiment of the present invention performs slot filling corresponding to an execution block (S130).

In this case, slot filling may refer to a process of receiving a slot value that can be input into a slot from a user.

In this case, slot filling may be performed based on at least one of an entity mapped to a slot, slots corresponding to an execution block, and intent mapping information of the slot.

For example, if the user accurately said the slot value through a conversation with the chatbot, slot filling can be performed based on the entity mapped to the slot, and the user can send a sentence corresponding to the intent corresponding to the current execution block. In this case, slot filling can be performed based on slots corresponding to execution blocks.

In this case, a slot value may be extracted by performing synonym matching on entities mapped to the slot. That is, entities mapped to slots may indicate which slot values can be input to slots.

For example, assuming that the 'coffee' entity is mapped to the $coffee slot included in the 'order_coffee' intent, the slot value corresponding to the 'coffee' entity may be input to the slot value of the $coffee slot.

For another example, assuming that the 'juice' entity is mapped to the $juice slot included in the 'order_juice' intent, the slot value corresponding to the 'juice' entity can be input to the slot value of the $juice slot. .

In this case, one slot may be mappable with a plurality of entities.

At this time, in the present invention, by mapping a plurality of entities to one slot, it is possible to respond to various situations that may occur through user speech.

For example, if you can order both coffee and juice, such as 'order drink', rather than an intent with a defined category, such as 'order coffee' or 'order juice', the $drink slot included in the 'order_drink' intent By mapping a plurality of 'coffee' entities, 'juice' entities, and 'drink' entities, the user can define various beverages to ignite.

At this time, according to an embodiment of the present invention, a slot to which an entity is not mapped may exist. In this case, a slot value may be freely set in the user's utterance based on a method other than entity synonym matching.

At this time, the synonym matching for the entity is to match whether there is a value set as a synonym for the entity in the user's speech. If a synonym for the entity is detected in the user's speech, a slot value corresponding to the entity is selected. can be entered.

For example, it can be assumed that the join entity, register entity, cancel entity, and withdraw entity shown in FIG. 2 are mapped to the $action slot included in the 'member_action' intent. At this time, if the utterance "I want to join" is input from the user, entity synonym matching may be performed and the representative word 'join' corresponding to the join entity may be input as a slot value.

At this time, in the present invention, slot filling may be performed based on intent mapping information of a slot in order to additionally perform slot filling even for intents that do not exist in the intent corresponding to the current execution block.

At this time, in the present invention, when performing slot filling, slot filling can be performed not only for slots existing in the intent but also for completely new slots based on intent mapping information. In this case, similarly, an entity can be designated and used in a slot. can Through this process, it is possible for the chatbot to flexibly respond to and handle various situations that may occur due to user speech.

For example, when the user says "HDMI connection is not available", the chatbot can respond with "Please enter the serial number on the back to check the model."

For another example, when the user says "Please cancel the purchase", the chatbot sends an authentication number to the user's mobile phone and then responds with "Please enter the authentication number sent to the mobile phone for identity verification".

That is, slot filling may also be performed on information that is difficult for the user to say from the start or information added during a conversation.

In this case, the intent mapping information is selected from among result values of the intent name mapped to the slot, the slot name of the intent mapped to the slot, the slot value of the intent mapped to the slot, and the slot value of the intent mapped to the slot. may contain at least one.

Hereinafter, a process of performing slot filling based on intent mapping information will be described in detail with reference to FIGS. 2 to 4 .

For example, assuming that an entity is defined as shown in FIG. 2 and an intent and a slot are defined as shown in FIG. 3, the join entity and the cancel entity are mapped to the 'join_confirm' slot. Intent mapping information for the 'join_confirm' slot may be defined as shown in FIG. 4 .

For example, in the case of the 'join_confirm' intent located first among the intents shown in FIG. 4, it may be mapped using an intent name mapped to a slot.

At this time, referring to FIG. 3, since the 'join_confirm' intent includes the 'join_confirm' slot, a slot value corresponding to the 'join_confirm' intent can be input.

As another example, in the case of the second 'positive' intent among the intents shown in FIG. may have been

That is, if the user answered "yes" to the bot's question "Do you want to join?", it can be classified as a 'positive' intent based on FIG. It can be input as a slot value of the 'join_confirm' slot.

As another example, in the case of the 'member_action' intent located third among the intents shown in FIG. 4, it may be mapped using the intent name mapped to the slot and the slot name of the intent mapped to the slot. .

That is, if the user answers "Would you like to join?" to the bot's question "Do you want to join?", the intent may be classified as 'member_action' based on FIG. 3 . At this time, the 'action' slot of the 'member_action' intent is mapped with the 'join' entity that can match 'join' as a synonym, so the 'join' corresponding to the entity representative word is the slot value of the 'join_confirm' slot. can be entered as

As another example, in the case of the 'member_action' intent located fourth among the intents shown in FIG. 4, the intent name mapped to the slot, the slot name of the intent mapped to the slot, and the intent mapped to the slot. It may be mapped using both the slot value and the result value of the slot value of the intent mapped to the slot.

That is, if the user answers "Do you want to register for swimming?" to the bot's question "Do you want to join?", it can be classified as a 'member_action' intent based on FIG. 3 . At this time, the 'action' slot of the 'member_action' intent is mapped with the 'register' entity that can match 'register' as a synonym, and the slot value of the 'action' slot of the 'member_action' intent is 'register'. Since 'join' is mapped as the result value in case of 'join', 'join' can also be input to the slot value of the 'join_confirm' slot.

Hereinafter, with reference to FIG. 5, a slot filling process according to an embodiment of the present invention will be described in detail.

According to one embodiment of the present invention, while processing the components in the execution block in order, it is determined whether slot filling is required for the slot designated by the slot component (S502), and if slot filling is required because the slot value is empty, the chatbot After requesting input from the user through a prompt message (Bot Prompt) (S504), a response from the user may be slot listened (S506).

Then, when a user message is input by the user's speech (S508), synonym matching with the synonyms of the entity mapped to the target slot for which slot filling is to be performed may be performed on a sentence corresponding to the input message (S510). .

Thereafter, it is determined whether a slot value extracted by entity synonym matching exists (S512), and if the slot value extracted by entity synonym matching exists as a result of the determination in step S512, action validation is performed to determine the slot value After verifying (S522), if the verification succeeds, the next block component can be executed.

At this time, as a result of the determination in step S512, if the entity synonym matching fails and the slot value extracted by the entity synonym matching does not exist, the slot value is extracted based on the block intents declared in the execution block. It can (S514).

That is, a slot value can be extracted based on slots included in intents declared in an execution block.

Thereafter, it is determined whether the extracted slot value exists based on the intents declared in the execution block (block intent) (S516), and as a result of the determination in step S516, the intents declared in the execution block (block intent) are determined. If there is a slot value extracted based on intent), action validation is performed to verify the slot value (S522), and if the validation succeeds, the next block component can be executed.

At this time, as a result of the determination in step S516, the extraction using the block intents declared in the execution block fails, so the slot value extracted based on the block intents declared in the execution block does not exist. Otherwise, a slot value may be extracted using intent mapping information of the target slot (S518).

Thereafter, it is determined whether a slot value extracted using the intent mapping information exists (S520), and if a slot value extracted using the intent mapping information exists as a result of the determination in step S520, action validation is performed to After verifying the slot value (S522), if the verification succeeds, the next block component can be executed.

At this time, as a result of the determination in step S520, if the extraction using the intent mapping information fails and the slot value of the target slot cannot be filled, it may be determined whether to retry slot filling (S524).

As a result of the determination in step S524, if slot filling can be retried, the user is guided to retry and requested to input again (S526), and the process can be performed again from step S504.

At this time, as a result of the determination in step S524, if the number of retries is exceeded or retries are impossible, it may be treated as failure of slot filling.

In addition, according to one embodiment of the present invention, even if the slot value for the target slot is filled and action validation is performed in accordance with step S522, but validation of the slot value fails, whether to retry slot filling can be determined. It can (S528).

As a result of the determination in step S528, if the slot filling can be retried, a retry is guided and re-input is requested (S530), and the process can be performed again from step S504.

At this time, as a result of the determination in step S528, if the number of retries is exceeded or retries are impossible, it may be treated as failure of slot filling.

In addition, although not shown in FIG. 1, the method for providing a chatbot service based on slot filling according to an embodiment of the present invention may perform slot reuse setting to maintain a slot context in consideration of a scenario including an execution block. .

That is, when a user speaks, it is possible to determine whether to reuse a slot value of a corresponding slot or to receive a new input according to reuse setting information of a slot connected to a block to be executed.

At this time, slot reuse configuration may be performed in consideration of various situations according to the type of scenario.

For example, in the case of the 'coffee order' scenario, the 'coffee order change' scenario may operate after the scenario ends and all slot values for the coffee order are input. In this case, you shouldn't set up slot reuse to run the 'coffee order change' scenario and ask again from the beginning for slot filling.

As a detailed example, in the 'coffee order' scenario, when a user's utterance "coffee please" is first received, the chatbot may respond with "what kind of coffee would you like?"

Thereafter, when the user says "{Americano:coffee} please" is received, the chatbot may respond with "Would you like it warm or iced?"

Thereafter, when a utterance "Please give me {ice:hotice}" is received from the user, the chatbot may respond with "Do you want to eat and go?"

After that, when the user says "{Takeout: takeout}", the chatbot uses the slot values entered in each slot ($coffee, $hotice, $takeout) to say "Iced Americano takeout has been ordered. ', the 'coffee order' scenario may end by executing the utterance confirming the order.

At this time, assuming that slot reuse is set for each slot ($coffee, $hotice, $takeout) and a user says "I'll change it to {latte:coffee}" is received, 'change coffee order' In the 'scenario, the slot values of $hotice and $takeout, which are slots that the user did not say, are maintained, so the chatbot can immediately respond with "Iced Latte Takeout".

That is, in this case, since slot filling must be performed again for the slot values of $hotice and $takeout, slots can be configured not to be reused.

Therefore, if slot reuse is not set for each of the slots ($coffee, $hotice, $takeout), and the utterance "{cappuccino:coffee} please" is received from the user again after the above order, the user The slot values of $hotice and $takeout, which are unspoken slots, are initialized and deleted, and the chatbot can ask the user to refill the deleted slot values.

In the past, input and output properties had to be set for each intent unit to maintain the slot context, and the set value had to be managed for each previous/next intent, so it is difficult to manage if the slot context is maintained or not depending on the scenario. there was

In the present invention, this problem can be solved through slot reuse configuration, and the context by the previous intent can be maintained more effectively.

In this case, the slot values of the intents connected to the execution block for which the slot reuse setting is performed may not be deleted even when the scenario is executed again after the end.

For example, according to the present invention, in the case of slots in which slot reuse is set when extracting slots from intents mapped to correspond to user conversations, if the slots are not spoken by the user, the already filled slot values are stored in memory without deleting them. can be dealt with In this case, since the slot context needs to be maintained even after the scenario ends, slot values existing in the context can be maintained without being deleted.

Also, for slots for which slot reuse is not set, when extracting a slot from an intent mapped to a user conversation, the slot value may be initialized if the user did not speak.

Through this processing process, the context can be maintained without managing separate input and output properties, so the service can be provided efficiently without burdening the chatbot server.

In this case, the slot value may be directly changed by using an assign action or a reset action, if necessary, in consideration of conditions set in the slot.

In addition, the method for providing a chatbot service based on slot filling according to an embodiment of the present invention performs conversation processing corresponding to an intent based on slot filling (S140).

As such, the present invention defines slot information mainly in the intent from the viewpoint of natural language processing, and divides the slot filling information necessary to execute the function to be defined in the execution block, so that different execution blocks using the same slot are each It can operate differently according to the slot filling information defined in the execution block.

That is, the slot information defined in the intent may not be completely fixed, but may have a meaning of a default setting, and may correspond to a concept that may be redefined by an execution block.

According to the configuration described above, the method for providing a chatbot service according to an embodiment of the present invention sets various user conversations by setting the same intent to operate differently based on the slot filling defined in the execution block when used in different execution blocks. You can make chatbots respond effectively to patterns.

In addition, the method for providing a chatbot service according to an embodiment of the present invention can perform slot filling not only for a slot existing in an intent but also for an entirely new slot using several entities mapped to one slot. Slot filling may also be enabled for information that is difficult for the user to say from the beginning or information that is added while having a conversation.

In addition, the chatbot service providing method according to an embodiment of the present invention simplifies the process of maintaining the context flow by providing a slot reuse function, and can more effectively manage slot values set for maintaining the context flow.

6 is a block diagram showing a chatbot service providing apparatus according to an embodiment of the present invention.

Referring to FIG. 6 , an apparatus for providing a chatbot service according to an embodiment of the present invention includes a processor 620 and an execution memory 630. According to the embodiment, the chatbot service providing device may further include a communication unit 610 .

The communication unit 610 may serve to transmit information related to a chatbot service with a plurality of other devices through a communication network such as a network. At this time, the network provides a path for transferring data between devices, and is a concept encompassing both existing networks and networks that can be developed in the future.

For example, the network is an IP network that provides large data transmission and reception services and data services without disconnection through the Internet Protocol (IP), and All IP, an IP network structure that integrates different networks based on IP. ) network, and can be a wired network, Wibro (Wireless Broadband) network, 3G mobile communication network including WCDMA, 3.5G mobile communication network including HSDPA (High Speed Downlink Packet Access) network and LTE network, 4 including LTE advanced It may be achieved by combining one or more of a household mobile communication network, a satellite communication network, and a Wi-Fi network.

For example, the network includes a wired and wireless local area network that provides communication between various information devices within a limited area, a mobile communication network that provides communication between mobile bodies and between mobile bodies and the outside of the mobile body, and a satellite that provides communication between earth stations and earth stations. It may be a satellite communication network or any one of wired and wireless communication networks, or a combination of two or more. Meanwhile, the transmission method standards of the network are not limited to existing transmission method standards, and may include all transmission method standards to be developed in the future.

The execution memory 630 stores one or more programs executed by the processor 620 .

At this time, at least one program determines an intent corresponding to the user dialogue, selects an execution block corresponding to the intent using the intent mapping list, performs slot filling corresponding to the execution block, and performs slot filling. Based on this, conversation processing corresponding to the intent is performed.

In this case, the intent mapping list is provided for each of the execution blocks and may correspond to at least one or more intents.

In this case, the at least one program may perform slot filling based on at least one of an entity mapped to a slot, slots corresponding to the execution block, and intent mapping information of the slot.

In this case, the intent mapping information is selected from among result values of the intent name mapped to the slot, the slot name of the intent mapped to the slot, the slot value of the intent mapped to the slot, and the slot value of the intent mapped to the slot. may contain at least one.

In this case, slots corresponding to the execution block may include slots of intents connected to the execution block through the intent mapping list.

In this case, the at least one program may extract a slot value by performing synonym matching on an entity mapped to the slot.

In this case, one slot may be mappable with a plurality of entities.

At this time, the at least one or more programs may perform slot reuse configuration for maintaining a slot context in consideration of a scenario including an execution block.

In this case, the slot values of the intents connected to the execution block for which the slot reuse setting is performed may not be deleted even when the scenario is executed again after the end of the scenario.

In this case, the intent mapping list may include an intent identifier and a slot condition for each intent.

The device shown in FIG. 6 may receive various types of information such as numbers and text information, and transmit input signals to the processor 620 through an input unit in relation to setting various functions and controlling functions of the device. In addition, the input unit may include at least one of a keypad and a touchpad that generate an input signal according to a user's touch or manipulation. At this time, the input unit of the device is configured in the form of a single touch panel (or touch screen) together with the display unit to simultaneously perform input and display functions. In addition, the input unit may use any type of input means that may be developed in the future, in addition to input devices such as a keyboard, keypad, mouse, joystick, and the like.

In addition, the display unit may display information about a series of operation states and operation results that occur while performing functions of the device. Also, the display unit may display menus of the device and user data input by the user. Here, the display unit includes a liquid crystal display (LCD), a thin film transistor LCD (TFT-LCD), a light emitting diode (LED), an organic light emitting diode (OLED), It can be composed of an active organic light emitting diode (AMOLED, Active Matrix OLED), a Retina Display, a flexible display, and a 3D display. At this time, when the display unit is configured in the form of a touch screen, the display unit may perform some or all of the functions of the input unit.

In addition, the storage unit of the device is a device for storing data, includes a main memory device and an auxiliary memory device, and may store application programs required for functional operation of the device. This storage unit may largely include a program area and a data area. Here, when the device activates each function in response to a user's request, the corresponding application programs are executed under the control of the processor 620 to provide each function.

In addition, the communication unit 610 may perform a function for transmitting and receiving data through a network. Here, the communication unit 610 may include an RF transmitter for up-converting and amplifying the frequency of a transmitted signal, and an RF receiver for low-noise amplifying a received signal and down-converting the frequency. The communication unit 1810 may include at least one of a wireless communication module and a wired communication module. And, the wireless communication module is a component for transmitting and receiving data according to a wireless communication method, and when the device uses wireless communication, data is transmitted using any one of a wireless network communication module, a wireless LAN communication module, and a wireless fan communication module. can be sent and received. In addition, the wired communication module is for transmitting and receiving data by wire. The wired communication module may access a network through a wire and transmit/receive data to and from the device. That is, the device accesses the network using a wireless communication module or a wired communication module, and can transmit/receive data with other devices through the network.

In addition, the processor 620 may be a process device that drives an operating system (OS) and each component.

According to the configuration described above, the slot filling-based chatbot service providing apparatus according to an embodiment of the present invention is set to operate differently based on the slot filling defined in the execution block when the same intent is used in different execution blocks By doing so, the chatbot can effectively respond to various user conversation patterns.

In addition, an apparatus for providing a chatbot service based on slot filling according to an embodiment of the present invention can perform slot filling not only for a slot existing in an intent but also for a completely new slot using several entities mapped to one slot. Through this, slot filling can be made possible for information that is difficult for the user to say from the beginning or information that is added while having a conversation.

In addition, the apparatus for providing a chatbot service based on slot filling according to an embodiment of the present invention simplifies the process of maintaining the context flow by providing a slot reuse function, and more effectively manages slot values set for maintaining the context flow. can

In addition, according to the present invention, according to natural language processing performed by a conversational AI agent system, intents corresponding to natural language inputs are classified according to similar properties, and a plurality of classified intents are mapped to a block, which is a processing unit of a bot. It's about technology.

For a useful chatbot service, it is necessary to continuously improve the ability to process natural language for terms by business or service domain and the processing ability for provided functions or tasks.

To this end, it is necessary to train the natural language processor so that various domain-specific terms and sentences can be processed by the natural language processor. At this time, the natural language processor needs to classify intents so that collisions between similar utterances do not occur. In other words, it is necessary to define intent from the perspective of Natural Language Understanding (NLU), rather than defining intent only by function or task.

On the other hand, in order to efficiently process functions or tasks, detailed processing steps or flows need to be defined as execution blocks, and a connection (order relationship) between the defined execution blocks needs to be defined. At this time, it is preferable that the connection relationship be defined in the form of a graph for freely connecting various execution blocks rather than in the form of a list or tree. At this time, it is preferable that the execution blocks are defined to suit the function provided by the service rather than from the perspective of natural language processing.

In this way, it is necessary to be able to freely (N:N) map intents and blocks defined in a form suitable for each purpose in order to efficiently provide a chatbot service that meets the user's intention. In other words, one intent can be written to several blocks, and one block must be able to handle multiple intents. To this end, intents and blocks must be mapped N:N rather than 1:1. .

When processing a large amount of data in which a large amount of utterance patterns are classified into various intents, collisions between similar utterances occur, which requires verification and processing.

At this time, similar utterance patterns can be grouped into the same intent, and conditions can be given based on slots to enable different bot responses.

In addition, if it is possible to map a plurality of user speech units to one bot response unit (execution block) so that they can be processed with the same response, even though they are not similar speech, it is possible to process them with the same response even if the speech type is different. This enables efficient intent management.

Hereinafter, a chatbot service is a concept covering not only a voice service based on a user's voice but also a messenger service based on text input from a user.

Chatbot design and implementation can be largely divided into three parts: chatbot builder, natural language understanding (NLU), and dialog management (DM). At this time, the chatbot builder structures and defines data for the dialog flow processing (DM) to operate and provides it, and the dialog flow processing (DM) processes the bot response based on this data.

A chatbot process may require two workspaces to run. One is a workspace where the bot designer defines the bot driving process, and the other is an execution area where the defined process is executed.

7 is a block diagram showing a chatbot service system according to an embodiment of the present invention.

Referring to FIG. 7 , the chatbot service system includes a chatbot 710, a natural language processor 720, a chatbot resource database 730, and a chatbot builder 740.

The chatbot 710 receives a spoken sentence from the user, generates a response thereto, and provides a chatbot response to the user.

The natural language processor 720 receives spoken sentences from the chatbot, receives intent definition information from the chatbot resource database 730, generates intent classification results and slot extraction results, and provides them to the chatbot 710.

The chatbot resource database 730 receives intent definition information, block information, response information, and scenario information from the chatbot builder, provides block information, response information, and scenario information to the chatbot 710, and the natural language processor 720 provides intent definition information.

The chatbot builder 740 receives input on chatbot resources from the bot designer, creates a chatbot 710, and provides intent definition information, block information, response information, and scenario information to the chatbot resource database 730. .

In the chatbot builder 740, learning utterances and slots, which are important information for bot processing, are defined as intents, the defined intents are mapped to execution blocks, conditions are set in the execution blocks, and bot actions are performed accordingly. can be defined In addition, blocks that can be processed in the response waiting state are set in the chatbot builder 740, and these blocks can be defined as graph-type scenarios.

At this time, the bot designer works to define the bot driving process through the chatbot builder 740. First, a block diagram in which bot scenarios for major business processes are defined in a graph form may be provided.

8 is a block diagram of a 'coffee order' scenario according to an embodiment of the present invention.

Referring to FIG. 8 , in the block speech scenario of the coffee ordering chatbot, a user's speech "I will order coffee" is first received (811).

In addition, in the block speech scenario of the coffee ordering chatbot, the chatbot executes the speech "What kind of coffee would you like?" (821).

In response to the chatbot's utterance, in the coffee ordering chatbot's block utterance scenario, the user's desired coffee name is uttered (812).

In response to the user's utterance, in the block utterance scenario of the coffee ordering chatbot, the chatbot executes the utterance "Would you like something warm or iced?" (822).

In response to the conversation of the chatbot, the user selects hot or ice in a scenario where the coffee ordering chatbot triggers a block (813).

In response to the user's speech, in the block speech scenario of the coffee ordering chatbot, the chatbot executes the speech "Are you going to eat?" (823).

In response to the conversation of the chatbot, the user selects in-store dining or takeout in a scenario where the coffee ordering chatbot triggers a block (814).

When all three slots of $ (coffee name), $ (hot ice), and $ (takeout) are input, the chatbot executes an utterance to confirm the order in a block utterance scenario of the coffee ordering chatbot (1324).

In response to the chatbot's utterance, in the coffee ordering chatbot's block firing scenario, the user utters 'yes' (815) or 'no' (816), and if the user utters 'yes', the chatbot executes the coffee order executes a custom action (825) and executes an utterance to confirm the order (826), and if the user utters 'No', the chatbot executes the utterance "Cancel the order" ( 827), and then the operation ends.

In FIG. 8 , "User" represents a user's utterance, "templete" represents a chatbot's utterance, and "Custom Action" represents a bot action process defined by an execution block.

In the example shown in FIG. 8, $ (coffee name), $ (hot ice), $ (takeout), and $ (order number) correspond to required slots, and although not explicitly shown in FIG. 8, coffee Option slots may be further used in the block firing scenario of the ordering chatbot. In this case, the required slot and/or the optional slot may be generated from the user's speech, or may be slots indicating the status of other chatbot systems or generated according to conditions in the block.

Based on the diagram shown in FIG. 8, all speech patterns that can be processed in each block can be collected, and the patterns can be grouped in a form that is easy for the natural language processor to classify. In addition, when an intent is classified based on grouped information and an intent is set by tagging slots for each utterance pattern, the utterance patterns based on the grouped and classified intents are matched to the learning intent of the natural language processor (NLU) learning data. can be used as

Scenarios can be viewed as connecting blocks in the form of a graph, and a list of intents that can be processed can be defined in each block.

FIG. 9 is a diagram illustrating an example of intents used in the scenario shown in FIG. 13 .

Referring to FIG. 9, the intents used in the coffee ordering scenario include the intent order_coffee (910), the intent order_hotice (920), the intent order_takeout (930), the intent order_togo (940), and the intent order_eatin (950). can be found to contain.

At this time, slot $coffee, slot $hotice, and slot $takeout are defined for intent order_coffee(910), slot $hotice is defined for intent order_hotice(920), and slot $hotice is defined for intent order_takeout(930). $takeout is defined, and slots may not be defined for intent order_togo (940) and intent order_eatin (950).

When the bot processes the intents shown in FIG. 9, they are processed in units called blocks (processing blocks), and a plurality of intents can be mapped to and processed in one block. Also, one intent may be mapped to a plurality of blocks.

10 to 13 are diagrams illustrating examples of execution blocks used in the scenario shown in FIG. 9 .

10 is a block diagram showing connections between a plurality of intents and a plurality of execution blocks.

Referring to FIG. 10 , it can be seen that the intent order_coffee 910 , the intent order_hotice 920 , and the intent order_takeout 930 are connected to the execution block coffee selection 1010 . In this case, the intent mapping list of the execution block coffee selection 1010 may include intent identifiers corresponding to the intent order_coffee 910 , the intent order_hotice 920 , and the intent order_takeout 930 .

At this time, the intent order_hotice (920) and the intent order_takeout (930) are also connected to the execution block hotice confirmation (1020).

At this time, the intent order_takeout (930) is also connected to the execution block takeout confirmation (1030).

The intent order_togo (940), the intent order_eatin (950), the intent followup.yes (960), and the intent followup.no (970) are connected to the action block takeout confirmation (1030).

As shown in FIG. 10, intents are mapped to blocks for their execution, and a condition is defined as a grammar expression based on a variable called a slot for each mapped intent, so that the bot classifies utterances only by classifying intents for each user utterance. Appropriate detailed processing is possible through conditions without processing.

As shown in FIG. 10, a plurality of intents may be connected to one block (N: 1) or a plurality of blocks may be connected to one intent (1:N), so that there is freedom between intents and blocks. Mapping (N: N) is possible.

Furthermore, one execution block may have a connection relationship with other execution blocks that can be executed later. For example, runblock coffee selection 1010 can be linked with runblock ice confirmation 1020 and runblock takeout confirmation 1030, where runblock hotice confirmation 1020 confirms runblock takeout 1030 ) can be associated with

FIG. 11 is a diagram showing an example of the execution block coffee selection shown in FIG. 10 .

Referring to FIG. 11, the execution block coffee selection 1010 may include a connected intent list, slot, slot filling, and action information.

In this case, the connected intent list may be included in the intent mapping list, and the intent mapping list may include an intent identifier and a slot condition for each intent. In this case, the slot condition may correspond to conditional statements related to one or more slots and may include a null condition.

The slots ($coffee, $hotice, $takeout) shown in FIG. 11 may include the union of all slots corresponding to the intents (order_coffee, order_hotice, order_takeout) to which the execution block coffee selection 1010 is connected.

The slot filling order information may include information on the filling order of the slots of the execution block coffee selection 1010 and ignition information necessary for filling the corresponding slot first.

Action information may include a condition and a response when the condition is satisfied. At this time, the condition may correspond to one or more conditional sentences using slots, and the response may include a link calling another block.

For example, the execution block coffee selection 1010 may connect the execution block hot ice check 1020 based on the condition determination using the slot $hotice.

For example, the execution block coffee selection 1010 may connect the execution block takeout check 1030 based on the condition determination using the slot $takeout.

FIG. 12 is a diagram illustrating an example of hot ice checking of the execution block shown in FIG. 10 .

Referring to FIG. 12 , the execution block hot ice check 1020 may include a connected intent list, slot, slot filling, and action information.

The slots ($hotice, $takeout) shown in FIG. 12 may include the union of all slots corresponding to intents (order_hotice, order_takeout) to which the execution block hotice check 1020 is connected.

The slot filling priority information may include information on the filling priority of the slots of the execution block hot ice check 1020 and ignition information necessary for filling the corresponding slot first.

Action information may include a condition and a response when the condition is satisfied. At this time, the condition may correspond to one or more conditional sentences using slots, and the response may include a link calling another block.

For example, the execution block hot ice check 1020 may connect the execution block takeout check 1030 based on the condition determination using the slot $takeout.

FIG. 13 is a diagram illustrating an example of execution block takeout confirmation shown in FIG. 10 .

Referring to FIG. 13 , execution block takeout confirmation 1030 may include a connected intent list, slot, slot filling, and action information.

The slot ($takeout) shown in FIG. 13 may include the union of all slots corresponding to the intents (order_takeout, order_togo, order_eatin, followup.yes, followup.no) to which the execution block takeout confirmation 1030 is connected. have.

Slot filling priority information may include information on the filling priority of the slots of the execution block takeout check 1030 and ignition information necessary for filling the corresponding slot first.

Action information may include a condition and a response when the condition is satisfied. At this time, the condition may correspond to one or more conditional sentences using slots, and the response may include a link calling another block.

14 is a block diagram showing a connection relationship between execution blocks shown in FIGS. 10 to 13 .

Referring to FIG. 14 , it can be seen that an execution block coffee selection 1010, an execution block hot ice confirmation 1020, and an execution block takeout confirmation 1030 are connected in a graph form.

In this case, the connection between the execution blocks may mean a connection with next execution blocks capable of processing the user's speech when entering the user's speech standby state in a specific current block. That is, since a situation that can occur when a user speech is received in a specific current block is execution blocks connected to the current block, only intents connected to these execution blocks can be searched as intents requiring user speech processing. In the state of waiting for user utterance, if you collect only the intents included in the sub-blocks connected based on the current block and search for the intents through the natural language processor, they correspond to utterances much more quickly and accurately than when all intents are searched. Intent can be identified.

In the example shown in FIG. 14, since the execution block coffee selection 1010 is connected to the execution block hot ice selection 1020 and the execution block takeout confirmation 1030, the user utterance in the state of execution block coffee selection 1010 If input, the next state may be execution block hot ice selection (1020) or execution block takeout confirmation (1030).

In addition, when user utterance is input in the state of execution block hot ice selection 1020, the next state may be execution block takeout confirmation 1030.

If a user intent is identified from the user's speech, the blocks may be searched by arranging them in the order of priority of intent conditions of blocks including the identified intent and/or a priority defined by a bot designer.

When a block that satisfies the conditions is searched for, it moves to the corresponding block, performs an operation defined according to the corresponding block setting, and returns to a state of waiting for a user response according to the condition of the action.

Connecting blocks to blocks may be defined as enumerating blocks to be processed next from the current block when a user response waiting state is entered, and several user response waiting states may be defined in one block.

15 is a diagram illustrating an example of connecting blocks by defining a waiting state for a user response in an execution block.

Referring to FIG. 15 , it can be seen that one block can be connected to a plurality of other blocks in the form of a graph.

Block 1 shown in FIG. 15 is connected to Block 2, Block 3, Block 4, and Block 5. At this time, Block 1 is connected to Block 2 and Block 3 by Condition 4, and by Condition 5, Block 3, It can be connected with Block 4 and Block 5. Accordingly, after Block 1, situations of Block 2, Block 3, Block 4, and Block 5 may occur.

Block 2 is connected to Block 6. Therefore, after Block 2, the situation of Block 6 may occur.

Block 3 is connected to Block 6 and Block 7. Accordingly, situations of Block 6 and Block 7 may occur after Block 3.

Block 5 is connected to Block 6 and Block 7. Accordingly, situations of Block 6 and Block 7 may occur after Block 5.

For example, if the current state corresponds to Block 2, it is sufficient to search for intents targeting only the intents corresponding to Block 6 for subsequent utterances, and the possibility of occurrence is high. There is no need to perform user utterance matching for intents corresponding to Block 7 that do not exist.

When one intent and one execution block are mapped, one intent may need to include a plurality of firing patterns.

For example, the intent (I1: intent to sign up) is the learning sentences "I want to write" (pattern classification P1), "How do I write" (pattern classification P1), "I'm curious about how to sign up" (pattern classification P3) and "Inquiry on how to join" (pattern classification P3), and the intent (I2: withdrawal intention) includes the learning sentences "I want to stop writing" (pattern classification P2), "I don't want to write anymore" (pattern classification P2), It may include "I'm curious about how to cancel" (pattern classification P3) and "Inquire about how to cancel" (pattern classification P3).

At this time, one block may be connected to the intent (I1: intent to join) and one block may be connected to the intent (I2: intent to leave), but the performance of the natural language processor may deteriorate as the intents include a plurality of speech patterns may be lowered

Therefore, in order to prevent degradation of the performance of the natural language processor, it is desirable to avoid including a plurality of speech patterns in one intent. For example, the intent (I1: subscription intent) includes only the learning sentences "I want to write" (pattern classification P1) and "how to write" (pattern classification P1), and the intent (I2: withdrawal intent) It includes only the learning sentences "I want to stop writing" (pattern classification P2) and "I don't want to use anymore" (pattern classification P2), and the intent (I3: inquiry intent) is the learning sentences "I'm curious about the <sign up> method" ( Pattern classification P3), "Inquire about how to <register>" (pattern classification P3), "I'm curious about the <withdrawal> method" (pattern classification P3), and "Inquire about how to <subscribe>" (pattern classification P3). have.

In this way, when an intent is classified so that it can be effectively processed by a natural language processor, it is necessary to connect the intents and processing blocks in more diverse forms. To this end, each of the block's intents may have a condition, and this condition may be defined as statements comparing slot values. That is, block and intent mapping information may be defined in a block in the form of a list (intent mapping list) in which intents and slot conditions are paired.

16 is a diagram illustrating an example of mapping intents defined for each speech pattern to blocks.

Referring to FIG. 16, it can be seen that three intents are connected to two blocks as a pair of intent and slot conditions.

At this time, the slots used for the slot condition may include not only slots defined in the current intent, but also slots of other blocks set while executing a previous conversation, or slots provided by the system.

In the example shown in FIG. 16, Intent 1 corresponds to Speech Pattern 1, Intent 2 corresponds to Speech Pattern 2, and Intent 3 corresponds to Speech Pattern 2. It can be seen that 3) corresponds to the ignition pattern 3 (Pattern 3).

In the example shown in FIG. 16, the intent mapping list of Block 1 includes a pair of Intent 1 and Condition 1, and a pair of Intent 3 and Condition 2. ), there are two elements of the pair. Therefore, Block 1 is connected to two Intents (Intent 1 and Intent 3).

In the intent mapping list of Block 2, there are two elements: a pair of Intent 2 and Condition 3 and a pair of Intent 3 and Condition 4. exist. Therefore, Block 2 is connected with two intents (Intent 2 and Intent 3).

17 is a diagram illustrating a case of 1:1 mapping between intents and execution blocks.

Referring to FIG. 17 , it can be seen that each of the three intents I1, I2, and I3 includes four firing patterns.

That is, intent I1 (subscription intention) includes speech patterns "I want to try", "How do I write", "I'm curious about how to sign up", and "Inquire how to sign up".

Intent I2 (withdrawal intent) includes the utterance patterns "I want to stop using", "I don't want to use anymore", I'm curious about how to unsubscribe, and "Inquire about how to unsubscribe".

Intent I3 (intent to inquire about regular membership) includes speech patterns "I want to become a regular member", "I want to be promoted to a regular member", "I'm curious about how to become a regular member", and "How to become a regular member".

When intents (intents) and responses (execution blocks) are mapped 1:1, when a sentence like "*** I'm curious" appears in multiple intents, the natural language processor (NLU) can distinguish between these similar sentences. difficult. For example, a sentence such as "I am new to sign up, ask how to sign up" is processed by the natural language processor as a sentence similar to "inquire how to sign up" of the withdrawal intent, depending on the circumstances, resulting in false intention recognition.

That is, the user's intent classification may be different from the intent classification that is easy to be processed by a natural language processor (NLU). One actual user's intent may be divided into several intents that are easy to be processed by a natural language processor, and conversely, several The intentions of the user may be classified into one intention that is easy for the natural language processor to process.

18 is a diagram illustrating an example of intent classification and block mapping according to an embodiment of the present invention.

Referring to FIG. 18 , it can be seen that sentences are classified into a form that is easy to process by a natural language processor (NLU), and responses are mapped according to conditions by tagging necessary slots.

In the example shown in FIG. 18 , four intents (I1, I2, I3, and I4) are linked to three responses (R1, R2, and R3).

At this time, Intent 1 (I1) is connected to the subscription information response (R1). At this time, the slot condition may be a null condition.

At this time, Intent 4 (I4) is connected to a subscription guide response (R1), a withdrawal guide response (R2), and a member guide response (R3). At this time, intent 4 (I4) can be connected to the subscription information response (R1) by the slot condition $service == "subscription". At this time, intent 4 (I4) may be connected to the withdrawal guidance response (R2) by the slot condition $service == "withdrawal". At this time, intent 4 (I4) can be connected to regular member information response (R3) by slot conditions $level == "regular member" and $service == "subscription".

At this time, Intent 2 (I2) is connected to the withdrawal guidance response (R2). At this time, the slot condition may be a null condition.

At this time, Intent 3 (I3) is connected to the regular member information response (R3). At this time, the slot condition may be a null condition.

Each of the responses shown in FIG. 18 may correspond to one execution block.

At this time, the execution block may be defined including intent, slot conditions, and response processing of the chatbot.

FIG. 19 is a diagram showing execution blocks corresponding to the responses shown in FIG. 18 .

Referring to FIG. 19 , each execution block includes an intent mapping list including a pair of at least one intent and a slot condition and action information.

As shown in FIG. 19 , the slot condition may include a null condition corresponding to “no condition” and may include conditional statements using one or more slots.

When the chatbot performs conversation processing, the natural language processor first classifies the intent from the sentence input by the user, extracts the slot, and then finds a matching block by examining the conditions listed in the block containing the classified intent. .

In the examples shown in FIGS. 18 and 19, when an intent is classified as "I4: Intent of inquiry", the blocks containing this intent are B1, B2, and B3, and the blocks matched according to the conditions are listed in the table below. Same as 1.

Condition block Custom condition check order $service == "subscribe" B1: Enrollment Guide One $service == "Withdrawal" B2: Withdrawal Information 2 $level == "Full Member" and $service == "Registration" B3: Regular member information 3

The order of checking conditions can be determined by the user designating the order of listing blocks. If, as shown in Table 1, processing is performed in the order listed by the user, a sentence such as regular member registration satisfies the conditions of $level == "regular member" and $service == "registration", but at the same time logically $service = = "subscription" condition is also satisfied, and block B1 having a higher condition check order may be executed, which may not be a desirable result.

In order to solve this problem, if priority is given again in consideration of the inclusion relationship of conditions, it can be determined as shown in Table 2 below.

Condition block Custom condition check order Reordering the ranking considering inclusion $level == "Full Member" and $service == "Registration" B3: Regular member information 3 One $service == "subscribe" B1: Enrollment Guide One 2 $service == "Withdrawal" B2: Withdrawal Information 2 3

At this time, the inclusion relationship of the condition can be determined by the inclusion relationship of the set of slots used in the conditional expression. The inclusive side (large set) is processed with priority over the included side (small set), and if there is no inclusion relationship with each other, Priorities designated by the user may be maintained. In the example of Table 2, the condition of B3, which inspects both level and service, has a higher order of condition inclusion than B1 and B2, which inspect only service, so B3 can be processed with priority over B1 or B2.

FIG. 20 is a diagram illustrating an example of prioritization according to inclusion relationships of conditions corresponding to the coffee ordering scenarios illustrated in FIGS. 9 to 14 .

Referring to FIG. 20, it can be seen that the condition corresponding to two coffee slots and the takeout slot has higher priority than the condition corresponding to one coffee slot.

Therefore, when the user says "I want to take Americano takeout", the slot coffee is filled with "Americano" and the slot takeout is filled with "Takeout", so the execution block coffee order takes precedence over the execution block takeout confirmation.

When the chatbot process is executed, it collects intents from connected blocks in the current user response waiting state, and can search for an appropriate intent among the intents collected through the natural language processor.

When searching for a block by intent, a suitable block can be searched by arranging in the order of the priority of the intent conditions of the blocks that contain the intent and the priority defined by the user.

If a suitable block is not found because the condition is not met, the above method can be repeated using the next suitable intent.

When a block that satisfies the conditions is found, it moves to the corresponding block, performs a predetermined action in the chatbot's response process, and enters the state of waiting for the next user response according to the condition.

21 is a block diagram illustrating an example of modules necessary for executing a chatbot.

Referring to FIG. 21, if the modules required when the chatbot executes the resources defined by the bot designer are largely divided around the conversation flow processor, the part that controls message input and output, the part that manages block status, the part that processes natural language, and the block It can be seen as a part that processes conditions and a part that executes blocks.

In the example of FIG. 21, the modules necessary for running the chatbot include a dialog flow processor, a block state manager, a message input gateway, a message output gateway, a response message generator, a block process executor, an NLU searcher, a block searcher, an intent classifier, a slot extractor, and a natural language processor. , NLU priority manager, condition resolver and block condition priority manager.

22 is a diagram illustrating a process of searching for intents and blocks while the chatbot is running.

Referring to FIG. 22, when a user's utterance is input while the chatbot is waiting for a user's input, blocks that can be executed at the current location are first searched. At this time, sub-blocks connected to the current (execution) block may be searched first, and then main blocks of the scenario may be searched. At this time, the main block means a list of blocks to be operated in the initial state of the conversation, and becomes the target blocks that the bot searches for when the user first speaks. That is, if the intent is identified only for the subblocks that are executable after the current block and the main blocks that are newly executable, efficient dialog flow processing is possible.

When a block is searched, the processing priority is checked by enumerating intents mapped to the searched block. At this time, intents defined in the searched blocks are listed, and matching scores of intents corresponding to the user's speech are determined using a natural language processor. The defined intent pattern threshold score list is sequentially searched, and the names of intents and matching scores are compared, and the intent with the highest matching score among the intents that satisfy the pattern and threshold score is selected first. extract

If there is one or more blocks to which the matched intent belongs, the processing priorities are sorted according to the defined intent mapping condition, and if the conditions have the same priority, the scenario and block priorities defined by the user are followed. At this time, the priority of the condition may be prioritized according to the inclusion relationship of the condition comparison phrase. That is, when the condition is constructed based on the slots extracted from the user's utterance, the comparison syntax constituting the condition is composed mainly of the slots mentioned by the user and the slots not mentioned are omitted, so a condition containing many slots It can be seen as a better reflection of the user's intention. That is, when there are several conditions that are true in the conditional statement, the condition that includes more constraints or slots may be considered higher in priority.

The sorted conditions are inspected in order of priority, and the block to which the condition satisfied in the first order belongs is treated as an execution block. If a block that satisfies the condition is not found, the blocks to which the intent belongs are searched again with the intent of the next priority, and the block is found by repeating the process of determining the priority of the condition and checking the condition.

23 is an operational flowchart illustrating a method of providing a chatbot service according to another embodiment of the present invention.

Referring to FIG. 23 , a method for providing a chatbot service according to another embodiment of the present invention determines an intent corresponding to a user conversation (S2310).

In this case, an intent corresponding to the user dialogue may be selected from one or more candidate blocks connected to the current block through a graph or from among intent candidates connected to the main block.

In this case, the intent corresponding to the user dialogue may be determined by considering the first intent candidate linked to the candidate block with priority over the second intent candidate linked to the main block.

In addition, in the chatbot service providing method according to another embodiment of the present invention, an execution block corresponding to the intent is selected using an intent mapping list (S2320).

In this case, the intent mapping list is provided for each of the execution blocks and may correspond to one or more intents.

In this case, the intent mapping list may include an intent identifier and a slot condition for each intent.

In this case, the slot condition may correspond to conditional statements related to one or more slots.

At this time, the slot condition may include a null condition.

In this case, the intent mapping list may include intent identifiers and slot conditions corresponding to a plurality of intents (N:1), and the intent is that different execution blocks have the same intent identifier. In this case, if the slot conditions are different, they may be connected (1:N) to different execution blocks.

In this case, the slots corresponding to the execution block may include slots of intents connected to the execution block through the intent mapping list.

At this time, the execution block corresponding to the intent is a first execution block having an inclusion relationship of the slot condition corresponding to the large set when the intent corresponding to the user dialogue is connected to a plurality of execution blocks. It can be selected with priority over the second execution block having an inclusion relationship of slot conditions corresponding to a small set included in . In this case, the inclusion relationship of the slot condition may be the inclusion relationship of slots included in the slot condition. For example, an execution block corresponding to $level == "regular member" and $service == "subscription" may be selected with priority over an execution block corresponding to $service == "subscription".

In addition, the chatbot service providing method according to another embodiment of the present invention performs conversation processing corresponding to the intent based on the execution block (S2330).

According to the above configuration, the chatbot service providing method according to an embodiment of the present invention has N:N (N is a natural number) mapping between intents classified as suitable for a natural language processor and execution blocks processing the intents. This allows several different utterances to perform one function, and conversely, similar utterances to perform completely different functions.

In addition, in the chatbot service providing method according to an embodiment of the present invention, it is possible to define various intent-execution block mapping relationships by using an intent mapping list including an intent identifier and a slot condition for each intent.

In addition, in the method for providing a chatbot service according to an embodiment of the present invention, when an intent corresponding to a user conversation is identified, candidate blocks that can be executed after the current execution block are identified, and only intents connected to the candidate blocks are identified. Fast and accurate intent identification is possible by targeting the identification.

24 is a graph showing performance of a natural language processor according to an increase in the number of speech patterns.

Referring to FIG. 24, when intents are classified according to functions provided by a chatbot, when various spoken sentences are added, similar speech patterns often appear in several intents, and natural language processor (NLU) intents as shown in A It can be seen that the classification performance is rapidly degraded.

In order to solve this problem, by defining an intent for each similar speech pattern in a form suitable for processing by the natural language processor, even if the speech pattern increases, the performance of the natural language processor can be maintained high as in B.

In addition, by mapping different speech patterns to a unit called 'block' so that the bot can properly process them according to conditions, various speech patterns can be efficiently processed.

Blocks are connected in the form of a graph, and several response waiting states can be defined according to conditions, so that work processes with various processing flows or steps can be visually defined with a small number of blocks.

Implementations of the functional operations and subject matter described herein may be implemented in digital electronic circuitry, or in computer software, firmware, or hardware, including the structures disclosed herein and their structural equivalents, or in a combination of one or more of the foregoing. can be implemented Implementations of the subject matter described herein are one or more computer program products, that is, one or more modules of computer program instructions encoded on a tangible program storage medium for execution by or for controlling the operation of a processing system. can be implemented

A computer readable medium may be a machine readable storage device, a machine readable storage substrate, a memory device, a composition of matter that affects a machine readable propagating signal, or a combination of one or more of these.

In this specification, the term 'system' or 'apparatus' encompasses all devices, devices, and machines for processing data, including, for example, a programmable processor, computer, or multiple processors or computers. A processing system may include, in addition to hardware, code that forms an execution environment for computer programs on demand, such as, for example, code constituting processor firmware, a protocol stack, a database management system, an operating system, or a combination of one or more of these. .

A computer program (also known as a program, software, software application, script or code) may be written in any form of programming language, including compiled or interpreted language, a priori or procedural language, and may be a stand-alone program or module; It may be deployed in any form including components, subroutines, or other units suitable for use in a computer environment. A computer program does not necessarily correspond to a file in a file system. A program may be in a single file provided to the requested program, or in multiple interacting files (e.g., one or more modules, subprograms, or files that store portions of code), or parts of files that hold other programs or data. (eg, one or more scripts stored within a markup language document). A computer program may be deployed to be executed on a single computer or multiple computers located at one site or distributed across multiple sites and interconnected by a communication network.

On the other hand, computer-readable media suitable for storing computer program instructions and data include, for example, semiconductor memory devices such as EPROM, EEPROM and flash memory devices, magnetic disks such as internal hard disks or external disks, magneto-optical disks and CDs. - may include all forms of non-volatile memory, media and memory devices, including ROM and DVD-ROM disks. The processor and memory may be supplemented by, or incorporated into, special purpose logic circuitry.

Implementations of the subject matter described herein may include back-end components, such as, for example, data servers, or may include middleware components, such as, for example, application servers, or may include, for example, web browsers or graphical users through which users may interact with implementations of the subject matter described herein. It may also be implemented in a computing system that includes a front-end component, such as a client computer having an interface, or any combination of one or more of such back-ends, middleware, or front-end components. The components of the system are interconnectable by any form or medium of digital data communication, such as, for example, a communication network.

Although this specification contains many specific implementation details, they should not be construed as limiting on the scope of any invention or what is claimed, but rather as a description of features that may be unique to a particular embodiment of a particular invention. It should be understood. Likewise, certain features that are described in this specification in the context of separate embodiments may also be implemented in combination in a single embodiment. Conversely, various features that are described in the context of a single embodiment can also be implemented in multiple embodiments individually or in any suitable subcombination. Further, while features may operate in particular combinations and are initially depicted as such claimed, one or more features from a claimed combination may in some cases be excluded from that combination, and the claimed combination is a subcombination. or sub-combination variations.

In addition, although operations are depicted in the drawings in a specific order in this specification, it is not to be understood that such operations must be performed in the specific order shown or sequential order or that all illustrated operations must be performed in order to obtain desired results. Can not be done. In certain cases, multitasking and parallel processing can be advantageous. Further, the separation of various system components in the embodiments described above should not be understood as requiring such separation in all embodiments, and the program components and systems described may generally be integrated together into a single software product or packaged into multiple software products. It should be understood that you can

As such, this specification is not intended to limit the invention to the specific terms presented. Therefore, although the present invention has been described in detail with reference to the above-described examples, those skilled in the art may make alterations, changes, and modifications to the present examples without departing from the scope of the present invention. The scope of the present invention is indicated by the following claims rather than the detailed description above, and all changes or modifications derived from the meaning and scope of the claims and their equivalent concepts should be interpreted as being included in the scope of the present invention. do.

As described above, the slot filling-based chatbot service providing method and apparatus according to the present invention are not limited to the configuration and method of the embodiments described above, but the embodiments can be modified in various ways All or part of each embodiment may be configured by selectively combining them.

According to the present invention, a method and apparatus for providing a chatbot service based on slot filling are disclosed. A method for providing a chatbot service according to an embodiment of the present invention determines an intent corresponding to a user conversation, selects an execution block corresponding to the intent using an intent mapping list, and fills slots corresponding to the execution block. , performs dialog processing corresponding to an intent based on slot filling, and an intent mapping list is provided for each of the execution blocks and corresponds to at least one intent. Accordingly, the chatbot can effectively respond to various user conversation patterns, making it possible to provide an efficient chatbot service.

610: communication unit 620: processor
630: memory 710: chatbot
720: natural language processor 730: chatbot resource DB
740: chatbot builder

Claims (5)

A slot filling based service providing device,
determining an intent corresponding to the user conversation;
selecting an execution block corresponding to the intent using an intent mapping list;
performing slot filling corresponding to the execution block;
performing conversation processing corresponding to the intent based on the slot filling;
A slot filling-based chatbot service providing method, characterized in that one slot can be mapped with a plurality of entities. The method of claim 1,
A method for providing a chatbot service based on slot filling, characterized in that one entity has a plurality of representative words. The method of claim 1,
The slot filling-based chatbot service providing device,
Slot filling-based chatbot service providing method further comprising the step of performing a slot reuse setting capable of maintaining a slot context by using a previously input slot as it is in consideration of a scenario including the execution block. The method of claim 3,
The method of providing a chatbot service based on slot filling, characterized in that the slot values of the slots of intents connected to the execution block in which the slot reuse setting is performed are not deleted even when the slots are re-executed after the scenario ends. The method of claim 1,
The intent mapping list is
A method for providing a chatbot service based on slot filling, which is provided for each of the execution blocks, and corresponds to at least one intent, but includes an intent identifier and a slot condition for each intent.

Images