KR20220065523A - Method and apparatus for expanding a task-oriented dialogue corpus - Google Patents

Abstract

The present invention provides a method for extending the destination conversation corpus includes the steps of: analyzing a conversation history between a user and a system by analyzing slot values of the slots stored for each utterance turn of the user and the system and a search result type of an entity database matching the slot values; selecting a candidate slot value based on the analyzed conversation history; and extending the destination conversation corpus by replacing the slot value with the candidate slot value.

Description

Method and apparatus for expanding a task-oriented dialogue corpus

The present invention relates to a conversation corpus processing technique required for learning of a destination-oriented conversation processing system.

With the recent development of neural network-based learning technology, interest in a destination-oriented conversation system based on a learning algorithm is increasing.

The destination-directed conversation system is a system for performing a specific task through conversation, and may be used, for example, for various information search, schedule management, product order, hotel or restaurant reservation, and the like.

Due to the nature of destination-oriented conversations in which conversations are conducted using an external database in a specific domain, a learning corpus is sparse, and it is necessary to create a new conversation corpus every time the external database is updated, which is a time-consuming and expensive corpus. Requires build work.

Since the performance of the learning algorithm-based system is determined by the amount of learning data, the performance of the destination-oriented dialog processing system with a small dialog corpus is not high.

Therefore, in order to improve the performance of the learning algorithm-based destination-oriented dialog processing system and at the same time reflect the updated database in the dialog system, a technology for automatically expanding the dialog corpus is required, but research on this is still insufficient.

SUMMARY OF THE INVENTION It is an object of the present invention to provide a method and apparatus for automatically expanding a dialogue corpus required for learning a dialogue system directed to a destination in order to improve the performance of the dialogue system directed to a destination.

The above and other objects, advantages and features of the present invention, and a method for achieving them will become apparent with reference to the embodiments described below in detail in conjunction with the accompanying drawings.

A method for expanding a conversation corpus directed to a destination according to the present invention comprises the steps of: analyzing an existing conversation corpus composed of an existing tagged utterance tagged with a slot and a slot value, and generating a conversation history for the existing tagged utterance; searching a database including entities belonging to the same domain as the domain of the existing conversation corpus, and selecting entities corresponding to the generated conversation history as entity candidates; generating a lexical utterance by removing the slot value from an existing tagging utterance; generating a similar derivational utterance similar to a sentence meaning of the generated divergence utterance from the generated divergence utterance by using a clustering technique; generating an extended demyelinating utterance from the pseudo-devomic utterance by filling the slots included in the pseudo-devomic utterance with a candidate slot value representing an attribute value of the entity selected as the entity candidate; and configuring the expanded lexical utterance into an expanded dialogue corpus.

The destination-directed conversation corpus generated according to the present invention has an advantage in that it can be built automatically without incurring a separate conversation creation cost. Therefore, when the DB is updated or has a small amount of conversation corpus compared to a large DB, the performance of the conversation system based on the learning algorithm can be improved.

1 is a block diagram illustrating an apparatus for expanding a destination-directed conversation corpus according to an embodiment of the present invention.
2 is a flowchart illustrating a method for expanding a dialogue corpus directed to a destination according to an embodiment of the present invention.
3 is a diagram illustrating an exemplary computing device to which the method for expanding the destination-directed conversation corpus of the present invention can be applied.

Specific structural or functional descriptions of the embodiments according to the concept of the present invention disclosed in this specification are only exemplified for the purpose of explaining the embodiments according to the concept of the present invention, and the embodiments according to the concept of the present invention are It may be implemented in various forms and is not limited to the embodiments described herein.

Since the embodiments according to the concept of the present invention may have various changes and may have various forms, the embodiments will be illustrated in the drawings and described in detail herein. However, this is not intended to limit the embodiments according to the concept of the present invention to specific disclosed forms, and includes changes, equivalents, or substitutes included in the spirit and scope of the present invention.

The terminology used herein is used only to describe specific embodiments, and is not intended to limit the present invention. The singular expression includes the plural expression unless the context clearly dictates otherwise. In the present specification, terms such as "comprise" or "have" are intended to designate that the described feature, number, step, operation, component, part, or combination thereof exists, and includes one or more other features or numbers, It should be understood that the existence or addition of steps, operations, components, parts or combinations thereof is not precluded in advance.

Technical concepts and terms applicable to the present invention will be first introduced.

database structure

In the destination-directed conversation system, the information that the user can request is called a 'slot' (a field in the database) and, for example, the restaurant domain, slot information for one restaurant becomes one entity (record in the database).

business name area food price range Phone number Manseon Restaurant Gangnam episode 5 million won 1111-2222 bull giblets Jam-sil giblets 3 million won 3333-4444 Dongsoye Jong-ro Webfoot octopus 10,000 won 5555-6666

Assuming the database as shown in Table 1 above, the slots are domain search conditions such as <region>, <food>, and <price range> and user information such as <business name> and <phone number>. It can be composed of things you want.

The slot value is the attribute value of the object corresponding to each slot. For example, in the case of <Food=Bibimbap>, <Food> is the slot, and [Bibimbap] is the slot value.

Learning-based destination-oriented dialogue system

The goal of the destination-oriented conversation system to which the present invention can be applied is to search a database according to a condition desired by the user and to provide information desired by the user.

For each user/system utterance, a destination-directed conversation system is learned from a conversation corpus in which a slot/slot value is tagged (sometimes additionally tagged with the utterance intention).

Since tagging information for each utterance does not exist in an actual conversation situation, the system is trained to find the information desired by the user from the conversation history in the form of spontaneous utterances and to output an appropriate response. Here, finding information desired by the user from the conversation history in the form of spontaneous speech means 'inference of slots and slot values'.

Spontaneous utterances are utterances in which slots are removed from tagging utterances, and are used as input to a destination-oriented dialogue system.

tagging speech I am looking for a restaurant in <Region = Gangnam>. spontaneous combustion I'm looking for a restaurant in Gangnam. out-of-vocabulary speech I'm looking for a restaurant in <region>.

Table 2 above shows examples of tagging utterances, spontaneous utterances, and out-of-vocabulary utterances.

How to auto-expand the dialog corpus

The dialog corpus automatic expansion method referred to in the present invention aims to generate dialogs referring to various utterances and various objects in a database of the same domain based on an existing dialog corpus in order to improve the performance of a learning-based destination-oriented dialog system.

Hereinafter, embodiments will be described in detail with reference to the accompanying drawings. However, the scope of the patent application is not limited or limited by these embodiments. Like reference numerals in each figure indicate like elements.

An embodiment of the present invention automatically extends the corpus from the database (DB) of the same domain by utilizing the existing destination-directed conversation corpus. The destination-oriented conversation corpus includes ontology tagging information composed of a slot-slot value pair defined for a conversation between a user and a system and all utterances made to achieve a user's purpose in a specific domain. Here, the slot may be defined as a basic storage unit of information for conversation as essential information for achieving the user's purpose. Also, the destination-directed conversation corpus may include intention tagging information of all utterances.

Table 3 below is an example of a destination-oriented conversation corpus in a restaurant search domain that searches for a restaurant based on a user's desired condition.

Talker (turn) type of utterance utterance dialogue act User(1) spontaneous combustion I'm looking for a restaurant in Gangnam. Information (Region = Gangnam) out-of-vocabulary speech I'm looking for a restaurant in <region>. System(1) spontaneous combustion Do you have any food you want? Request (food) out-of-vocabulary speech Do you have any food you want? User(2) spontaneous combustion I want to eat bibimbap. Inform (food = bibimbap) out-of-vocabulary speech I want to eat <food>. system(2) spontaneous combustion There is restaurant A that sells bibimbap. Inform(Name=A) out-of-vocabulary speech There is a restaurant <name> that sells <food>.

The utterances tagged with the slot <region, food, etc.> defined for the utterances of the user/system are called delexicalized utterances. Each utterance may be expressed in units of dialogue act by using the intention of the additionally tagged utterance.

The destination-oriented dialogue system finds the condition (slot value) desired by the user through dialogue with the user, searches the DB, and outputs the result that meets the condition. Therefore, the flow of conversation may vary depending on the order in which slot values are mentioned and the DB search result corresponding to the slot value.

1 is a block diagram illustrating an apparatus for expanding a destination-directed conversation corpus according to an embodiment of the present invention.

Referring to FIG. 1 , an apparatus 100 for expanding a destination-directed conversation corpus according to an embodiment of the present invention may be a computing device.

A computing device includes, for example, a processor 101, a memory 103, a storage 105, a user input/output device (not shown), a communication device (not shown), and a system bus (not shown) connecting them. not), and the like.

In addition, the computing device may be configured to further include modules controlled or executed by the processor 110 , and the modules are, for example, as shown in FIG. 1 , to an entity (DB entity) that is not in the existing conversation corpus. create a conversation about

To this end, the apparatus 100 for extending the destination-directed conversation corpus ('existing destination-directed conversation corpus' or 'existing conversation corpus') includes a conversation history analysis unit 110 , an object candidate selection unit 120 , and a utterance generation unit. (130).

Each of the components 110 , 120 , and 130 may be composed of a software module, a hardware module, and a combination thereof, and when each module is a software module, they reside in the memory 103 and are controlled by the processor 101 or the processor It may be an algorithm programmed to be executed according to an instruction input from 101 . When each module is a hardware module, they may be implemented with circuit logic built into the processor 101 .

Hereinafter, each configuration will be described in detail.

Conversation history analysis unit 110

The conversation history analysis unit 110 analyzes an existing conversation corpus composed of a slot and a tagged utterance tagged with a slot value (hereinafter, referred to as 'existing tagged utterance'), and analyzes data representing a conversation history for the existing tagged utterance. output as a result.

The data representing the conversation history includes data representing the slot combination (combination of slots) included in the existing tagged utterance and the type of DB search result retrieved from the database 105a using the slot combination as a keyword (hereinafter, 'DB search result') It contains data representing the type'). Here, as described above, the database 105a may have a structure in which a slot indicating a search condition of the same domain as the domain of the existing conversation corpus is used as a field, and an attribute value of an entity corresponding to the slot is recorded as a record. .

Table 4 below shows an example of the slot combinations constituting the existing tagged utterances and conversation history and the DB search result types corresponding thereto.

Talker (turn) conventional tagging utterances conversation history slot combinations DB search result type User (1) I'm looking for a <Food-Bibimbap> house in <Region=Gangnam>. (local, food) many system (1) There are a total of <object=4> places. user (2) Do you have a house with <price=under 10,000 won>? (region, food, price range) failure system (2) There is not. Please look for other conditions. user (3) <Price = 10,000 won> is also good. (local food, price range) only system (3) There is a restaurant <name=A> that meets the criteria you are looking for.

As shown in Table 4 above, the DB search result type may be classified into, for example, 'many', 'unique', 'failure', and the like. If the DB search result type is 'Failed', "There is no restaurant that satisfies the criteria you are looking for. Could you tell me about another criteria?" Conversation history can be analyzed from DB search result types such as

If the DB search result type is 'many', "There are N restaurants that match your criteria. Where is the best location?" You can analyze the flow of conversation (conversation history) from the DB search result type, such as.

Object candidate selection unit 120

The object candidate selector 120 selects, under the control of the processor 101, objects corresponding to slot values (slot value candidates) satisfying the slot combinations stored for each turn and the DB search result type as conditions in the database 105a. , and the searched objects are selected as object candidates.

Table 5 below shows examples of slot value candidates and entity candidates.

Condition slot value candidate object candidate note

다수Query (region)

many Area: {Wolpyeong, Dunsan} Dunsan Bun Restaurant, Dunsan Chinese Restaurant 1, Dunsan Chinese Restaurant 2 Excluded from candidate group because many restaurants were not found with "region = Wolpyeong" Query (food)

only Food: {Sweets, Korean, Chinese} Dunsan Bun Restaurant For the remaining candidate group, a large number of restaurants were searched for "food = Chinese food" and excluded

In the example of Table 5 above, when the existing tagging utterance is replaced and expanded to include the objects retrieved from the database 105a, it is shown that the substitutable object candidate is "Dunsan Bunsikdang".

Speech generator 130

The utterance generating unit 130 generates a new dialogue based on the candidate entity group selected by the entity candidate selection unit 120 , and configures the generated new dialogue into an expanded dialogue corpus.

In order to configure the expanded conversation corpus, the utterance generation unit 130 first performs a process of generating a de-vocabulary utterance from the existing tagging utterances included in the existing conversation corpus. Here, the process of generating a lexical utterance from the existing tagged utterance may be a process of removing a slot value from the existing tagged utterance tagged as a pair of a slot and a slot value.

Table 6 below shows examples of existing tagging utterances and de-vocabulary utterances generated therefrom.

conventional tagging utterances out-of-vocabulary speech I am looking for a restaurant in <Region = Gangnam>. I am looking for a restaurant in <region>. Do you have any food you want? Do you have any food you want? I want to eat <food = bibimbap>. I want to eat <food>. A place that sells <food = bibimbap> is <trade name = A>. A place that sells <food> is <trade name>.

In the example of Table 6 above, in the case of the existing tagging utterance "I am looking for a restaurant in <region=Gangnam>", by removing <Gangnam> corresponding to the slot value, "I am looking for a restaurant in <region>." A lexical utterance is generated.

In the other example of Table 6 above, in the case of the existing tagging utterances such as “There is a place that sells <food = bibimbap>, there is <trade name = A>.” By removing it, a de-vocabulary utterance is created, "There is a <trade name> for a place that sells <food>."

As such, when the evasive utterance is generated, the utterance generation unit 130 extracts (generates) similar evasive utterances from the generated divergence utterance by using a vector-based clustering technique. Here, the vector-based clustering technique is a well-known technique widely used to measure the similarity between sentences in one specific sentence by clustering various sentences. A paraphrase method and the like may be exemplified, and detailed descriptions thereof are replaced with known techniques.

Table 7 below shows examples of lexical utterances and similar lexical utterances similar to the sentence meanings of these lexical utterances.

out-of-vocabulary speech pseudo-lexical utterances I am looking for a restaurant in <region>. Are there any restaurants in <region>? Do you have any food you want? What food are you looking for? I want to eat <food>. I wish there was a place that sells <food>. A place that sells <food> is <trade name>. For <food> restaurants, there is <name>.

As shown in Table 7 above, analogous lexical utterances extracted from lexical utterances according to the vector-based clustering technique are tagged in the same slot as those tagged in lexical utterances, and are defined as having similar meanings to lexical utterances at the same time. can do.

As such, when the lexical derivational utterance is extracted from the lexical lexical utterance, the utterance generating unit 130 uses an attribute value (or a slot value) of an entity arbitrarily extracted from among the entities selected by the entity candidate selection unit 120 to generate the lexical lexical verbatim utterance. An extended out-of-vocabulary utterance is generated through a tagging process that fills the slot of

Table 8 below shows examples of pseudo-devomic utterances and extended out-of-vocabulary utterances generated from these pseudo-devomic utterances.

pseudo-lexical utterances extended lexical utterances Are there any restaurants in <region>? Are there any restaurants in <Region=Dunsan>? What food are you looking for? What food are you looking for? I wish there was a place that sells <food>. I wish there was a place that sells <food = snack>. For <food> restaurants, there is <name>. <Food=Bunsik> There is <Name=Dunsan Bunsikdang> as a restaurant.

2 is a flowchart illustrating a method for expanding a destination-directed conversation corpus according to an embodiment of the present invention.

Referring to FIG. 2 , first, in step S210, the conversation history analysis unit 110 analyzes the existing conversation corpus composed of the existing tagged utterances tagged with slots and slot values to obtain the conversation history for the existing tagged utterances. The creation process is carried out. Here, the conversation history includes a slot combination and a DB search result type.

Next, in step S220, a process of selecting, as object candidates, objects corresponding to the conversation history retrieved from a database consisting of objects belonging to the same domain as the domain of the existing conversation corpus by the object candidate selection unit 120 is performed do.

Next, in step S230 , a process of generating a lexical utterance by removing the slot value from the existing tagged utterance is performed by the utterance generating unit 130 .

Next, in step S240 , a process of generating, by the utterance generation unit 130 , a similar derivational utterance similar to the sentence meaning of the generated devlexical utterance from the generated devlexical utterance by using the clustering technique is performed.

Next, in step S250, a new slot value indicating the attribute value of the entity selected as the entity candidate (candidate slot indicating the attribute of the entity candidate) in the slot included in the pseudo-de-vocabulary utterance by the utterance generating unit 130 value), a process of generating an extended out-of-vocabulary utterance from the pseudo-devomic utterance is performed.

Next, in step S260 , a process of configuring the expanded lexical utterance into an expanded dialogue corpus is performed by the utterance generator 130 .

3 is a diagram illustrating an exemplary computing device to which the method for expanding the destination-directed conversation corpus of the present invention can be applied.

Referring to FIG. 3 , a computing device 500 to which the method for expanding the destination-directed conversation corpus of the present invention may be applied may be used to perform various procedures as discussed herein.

Computing device 500 may be any of a variety of computing devices, such as desktop computers, notebook computers, server computers, portable computers, tablet computers, and the like.

Computing device 500 includes one or more processors (CPU, GPU) 402 , one or more memory device(s) 404 , one or more interface(s) 406 , one or more mass storage device(s) 408 . , one or more input/output (I/O) device(s) 410 , and a display device 430 , all connected to bus 412 .

Processor(s) 402 includes one or more processors or controllers that execute instructions stored in memory device(s) 404 and/or mass storage device(s) 408 . Processor(s) 402 may also include various types of computer-readable media, such as cache memory.

Memory device(s) 404 may include volatile memory (eg, random access memory (RAM) 414 ) and/or non-volatile memory (eg, read-only memory (ROM) 416 ). a variety of computer-readable media. Memory device(s) 404 may also include rewritable ROM, such as flash memory.

Mass storage device(s) 408 include various computer-readable media such as magnetic tape, magnetic disks, optical disks, solid-state memory (eg, flash memory), and the like. As shown in FIG. 4 , a particular mass storage device is a hard disk drive 424 . Various drives may also be included in mass storage device(s) 408 to enable reading from and/or writing to various computer readable media. Mass storage device(s) 408 include removable storage 426 and/or non-removable media.

Input/output (I/O) device(s) 410 include various devices that allow data and/or other information to be input to or retrieved from computing device 400 . Exemplary I/O device(s) 410 include a cursor control device, keyboard, keypad, microphone, monitor or other display device, speaker, printer, network interface card, modem, lens, CCD or other image capture device, etc. include

Display device 430 includes any type of device capable of displaying information to one or more users of computing device 400 . Examples of display devices 430 include monitors, display terminals, video projection devices, and the like.

A graphics-processing unit (GPU) 432 may be coupled to the processor(s) 402 and/or the display device 430 . The GPU may be operable to render computer-generated images and perform other graphics processing.

A GPU may include some or all of the functionality of a general purpose processor, such as processor(s) 402 . The GPU may also include additional functions specific to graphics processing. The GPU may include hard-coded and/or hard-wired graphics functions related to coordinate transformations, shading, texturing, rasterization, and other functions useful for rendering computer-generated images.

Interface(s) 406 include various interfaces that enable computing device 400 to interact with other systems, other devices, or other computing environments.

Exemplary interface(s) 406 include any number of different network interfaces 420 , such as interfaces to local area networks (LANs), wide area networks (WANs), wireless networks, and the Internet.

Other interface(s) include user interface 418 and peripheral device interface 422 .

Interface(s) 406 may also include one or more user interface elements 418 . Interface(s) 406 may also include one or more peripheral interfaces, such as interfaces for printers, pointing devices (mouse, trackpad, etc.), keyboards, and the like.

Bus 412 includes processor(s) 402 , memory device(s) 404 , interface(s) 406 , mass storage device(s) 408 , and I/O device(s) 410 . ) to communicate with each other, as well as other devices or components connected to the bus 412 .

Bus 412 represents one or more of several types of bus structures, such as a system bus, a PCI bus, an IEEE 1394 bus, a USB bus, and the like.

For purposes of illustration, although programs and other executable program components are shown as separate blocks herein, such programs and components may reside at various times in different storage components of computing device 400 , and the processor(s) ) 402 . Alternatively, the systems and procedures described herein may be implemented in hardware, or hardware, or a combination of hardware, software and/or firmware. For example, one or more application specific integrated circuits (ASICs) may be programmed to perform one or more systems and procedures described herein.

So far, the present invention has been looked at focusing on examples. Those of ordinary skill in the art to which the present invention pertains will understand that the present invention can be implemented in variously changed or modified forms without departing from the essential characteristics of the present invention. Therefore, the disclosed embodiments are to be considered in terms of illustrative rather than restrictive views. The scope of the present invention is indicated in the claims rather than the foregoing description, and all differences within the scope equivalent thereto should be construed as being included in the present invention.

Claims (1)

analyzing an existing conversation corpus composed of a slot and an existing tagged utterance tagged with a slot value, and generating a conversation history for the existing tagged utterance;
searching a database including entities belonging to the same domain as the domain of the existing conversation corpus, and selecting entities corresponding to the generated conversation history as entity candidates;
generating a lexical utterance by removing the slot value from an existing tagging utterance;
generating a similar derivational utterance similar to a sentence meaning of the generated divergence utterance from the generated divergence utterance by using a clustering technique;
generating an extended evasive utterance from the lexical derivational utterance in such a way that the slot included in the lexical lexical utterance is filled with a candidate slot value indicating an attribute value of the entity selected as the entity candidate; and
Constructing the extended lexical utterance into an extended dialogue corpus
A method for expanding a destination-directed conversation corpus comprising:

Images