KR20230071673A - Method, computer device, and computer program for building open-domain dialogue system using language model - Google Patents

Abstract

Disclosed are a method, computer device, and computer program for constructing an open-type domain conversation model using a language model. A prompt that becomes an input statement of the language model as a role specification of a chatbot and an example of a conversation between the chatbot and a person is configured, wherein the new conversation data of the role and conversation pattern included in the prompt can be generated as the data for learning the chatbot through the language model by inputting the prompt into the language model. Therefore, the present invention is capable of providing a conversation system with an improved response quality.

Description

Method, computer device, and computer program for constructing an open domain dialog model using a language model

The description below relates to techniques for building an open-domain dialogue model.

A question and answer (QA) system is a dialog system that provides answers by referring to appropriate documents when a user inputs a question.

As language models have recently developed, the performance of dialog systems is rapidly improving, and it is a trend that extends not only to natural languages but also to multi-modal domains such as images and voices.

For example, Korean Patent Publication No. 10-2002-0030545 (published on April 25, 2002) discloses a technology for providing answers to questions using artificial intelligence technology and natural language processing technology.

It is possible to build an open domain conversation model that can naturally communicate with people and maintain a consistent role.

Context few-shot learning of the language model allows us to build a dialog dataset that satisfies the role.

A method executed on a computer device, wherein the computer device includes at least one processor configured to execute computer readable instructions contained in a memory, and the method comprises, by the at least one processor, a role specification of a chatbot ( configuring a prompt that is an input sentence of a language model as an example of a role specification and a conversation between the chatbot and a person; and generating, by the at least one processor, new dialogue data of roles and dialogue patterns included in the prompt through the language model by inputting the prompt into the language model.

According to one aspect, the method further comprises building, by the at least one processor, an open-domain dialogue model for the chatbot by supervised learning on the conversation data. can include

According to another aspect, the method may further include filtering, by the at least one processor, the conversation data through annotations.

According to another aspect, the filtering may include classifying the annotated problem remarks among the speeches of the chatbot included in the conversation data as negative examples for learning the chatbot. there is.

According to another aspect, the method further comprises collecting, by the at least one processor, conversation data between the chatbot and a person as learning data of the chatbot through an interface in which the chatbot and a person directly communicate. can

According to another aspect, the collecting may include, when a modification request is received for some of the chatbot's remarks provided through the interface, generating an alternative speech through the language model to correct the partial speech. can include

According to another aspect, the collecting may include classifying utterances requested for correction as negative examples for learning the chatbot; and classifying the utterance corrected as the substitute utterance as a positive example for learning the chatbot.

According to another aspect, the constructing may include an Out-of-Bounds Detection Model (Out-of-Bounds Detection Model) and a retrieve-then-rerank approach after finding response candidates. In , a response selection model for filtering the response candidates based on the perplexity (PPL) of the language model, and maximum likelihood estimation are applied to positive examples of the dataset and unlikelihood The domain conversation model may be built using at least one of a Response Generation Model that applies learning to negative examples in a dataset.

A computer program stored in a computer readable recording medium to execute the method in the computer device is provided.

A computer device comprising at least one processor configured to execute computer readable instructions included in a memory, wherein the at least one processor includes a role specification of a chatbot and an input statement of a language model as an example of a conversation between the chatbot and a person. and generating new dialogue data of roles and dialogue patterns included in the prompt through the language model by inputting the prompt to the language model.

According to embodiments of the present invention, it is possible to provide a dialog system with improved response quality by establishing an open domain dialog model capable of maintaining a consistent role while having a natural conversation with a person.

According to embodiments of the present invention, a conversation dataset that satisfies a role can be quickly and efficiently created through a data collection framework that utilizes contextual snapshot learning of a language model.

1 is a diagram illustrating an example of a network environment according to an embodiment of the present invention.
2 is a block diagram illustrating an example of a computer device according to one embodiment of the present invention.
3 illustrates a framework for building a role-specified open domain dialog model in an embodiment of the present invention.
4 is a flowchart illustrating an example of a method for creating an open domain conversation performed by a computer device according to an embodiment of the present invention.
5 is an exemplary diagram for explaining a process of generating conversation data according to an embodiment of the present invention.
6 shows an example of filtering conversation data according to an embodiment of the present invention.
7 is an exemplary diagram for explaining a process of collecting conversation data between a chatbot and a person in one embodiment of the present invention.
8 illustrates an exemplary structure of an open domain dialogue model according to an embodiment of the present invention.

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings.

Embodiments of the present invention relate to techniques for building an open-domain dialogue model.

Embodiments including those specifically disclosed in this specification can create an open domain conversation model that can maintain a consistent role while naturally conversing with a person.

The open domain dialogue generating system according to embodiments of the present invention may be implemented by at least one computer device, and the open domain dialogue generating method according to embodiments of the present invention includes at least one of the open domain dialogue generating systems. It can be performed through a computer device of. At this time, a computer program according to an embodiment of the present invention may be installed and driven in the computer device, and the computer device may perform the open domain conversation creation method according to the embodiments of the present invention under the control of the driven computer program. can The above-described computer program may be combined with a computer device and stored in a computer readable recording medium in order to execute an open domain conversation generating method on a computer.

1 is a diagram illustrating an example of a network environment according to an embodiment of the present invention. The network environment of FIG. 1 shows an example including a plurality of electronic devices 110 , 120 , 130 , and 140 , a plurality of servers 150 and 160 , and a network 170 . 1 is an example for explanation of the invention, and the number of electronic devices or servers is not limited as shown in FIG. 1 . In addition, the network environment of FIG. 1 only describes one example of environments applicable to the present embodiments, and the environment applicable to the present embodiments is not limited to the network environment of FIG. 1 .

The plurality of electronic devices 110, 120, 130, and 140 may be fixed terminals implemented as computer devices or mobile terminals. Examples of the plurality of electronic devices 110, 120, 130, and 140 include a smart phone, a mobile phone, a navigation device, a computer, a laptop computer, a digital broadcast terminal, a personal digital assistant (PDA), and a portable multimedia player (PMP). ), and tablet PCs. As an example, FIG. 1 shows the shape of a smartphone as an example of the electronic device 110, but in the embodiments of the present invention, the electronic device 110 substantially uses a wireless or wired communication method to transmit other information via the network 170. It may refer to one of various physical computer devices capable of communicating with the electronic devices 120 , 130 , and 140 and/or the servers 150 and 160 .

The communication method is not limited, and short-distance wireless communication between devices as well as a communication method utilizing a communication network (eg, a mobile communication network, a wired Internet, a wireless Internet, and a broadcasting network) that the network 170 may include may also be included. For example, the network 170 may include a personal area network (PAN), a local area network (LAN), a campus area network (CAN), a metropolitan area network (MAN), a wide area network (WAN), and a broadband network (BBN). , one or more arbitrary networks such as the Internet. In addition, the network 170 may include any one or more of network topologies including a bus network, a star network, a ring network, a mesh network, a star-bus network, a tree or a hierarchical network, and the like. Not limited.

Each of the servers 150 and 160 communicates with the plurality of electronic devices 110, 120, 130, and 140 through the network 170 to provide commands, codes, files, contents, services, and the like, or a computer device or a plurality of computers. Can be implemented in devices. For example, the server 150 may be a system that provides a service (eg, a chat bot service) to the plurality of electronic devices 110, 120, 130, and 140 connected through the network 170.

2 is a block diagram illustrating an example of a computer device according to one embodiment of the present invention. Each of the plurality of electronic devices 110 , 120 , 130 , and 140 or each of the servers 150 and 160 described above may be implemented by the computer device 200 shown in FIG. 2 . For example, an open domain dialog creation system according to embodiments of the present invention may be implemented by the computer device 100 shown in FIG. 1 .

As shown in FIG. 2 , the computer device 200 may include a memory 210, a processor 220, a communication interface 230, and an input/output interface 240. The memory 210 is a computer-readable recording medium and may include a random access memory (RAM), a read only memory (ROM), and a permanent mass storage device such as a disk drive. Here, a non-perishable mass storage device such as a ROM and a disk drive may be included in the computer device 200 as a separate permanent storage device distinct from the memory 210 . Also, an operating system and at least one program code may be stored in the memory 210 . These software components may be loaded into the memory 210 from a computer-readable recording medium separate from the memory 210 . The separate computer-readable recording medium may include a computer-readable recording medium such as a floppy drive, a disk, a tape, a DVD/CD-ROM drive, and a memory card. In another embodiment, software components may be loaded into the memory 210 through the communication interface 230 rather than a computer-readable recording medium. For example, software components may be loaded into memory 210 of computer device 200 based on a computer program installed by files received over network 170 .

The processor 220 may be configured to process commands of a computer program by performing basic arithmetic, logic, and input/output operations. Instructions may be provided to processor 220 by memory 210 or communication interface 230 . For example, processor 220 may be configured to execute received instructions according to program codes stored in a recording device such as memory 210 .

The communication interface 230 may provide a function for the computer device 200 to communicate with other devices (eg, storage devices described above) through the network 170 . For example, a request, command, data, file, etc. generated according to a program code stored in a recording device such as the memory 210 by the processor 220 of the computer device 200 is controlled by the communication interface 230 to the network ( 170) to other devices. Conversely, signals, commands, data, files, etc. from other devices may be received by the computer device 200 through the communication interface 230 of the computer device 200 via the network 170 . Signals, commands, data, etc. received through the communication interface 230 may be transferred to the processor 220 or the memory 210, and files, etc. may be stored as storage media that the computer device 200 may further include (described above). permanent storage).

The input/output interface 240 may be a means for interface with the input/output device 250 . For example, the input device may include a device such as a microphone, keyboard, or mouse, and the output device may include a device such as a display or speaker. As another example, the input/output interface 240 may be a means for interface with a device in which functions for input and output are integrated into one, such as a touch screen. The input/output device 250 and the computer device 200 may be configured as one device.

Also, in other embodiments, computer device 200 may include fewer or more elements than those of FIG. 2 . However, there is no need to clearly show most of the prior art components. For example, the computer device 200 may be implemented to include at least some of the aforementioned input/output devices 250 or may further include other components such as a transceiver and a database.

Hereinafter, specific embodiments of a method and apparatus for constructing an open domain dialogue model using a language model will be described.

The large-scale language model used in the present invention is a language model learned from massive data and refers to a natural language generation model that can appropriately perform a corresponding task given only a few-shot samples.

In other words, a large-scale language model is an autoregressive model, which can refer to a language model that can be inferred without fine-tuning using a method such as one-shot learning. You can have more than twice as many parameters (eg, more than 100 billion parameters, etc.). For example, large-scale language models such as Generative Pre-trained Transformer 3 (GPT-3) or HyperClova are good short-shot learners that can be controlled through natural prompts, allowing them to understand patterns with only a small amount of data through NLP. In-context learning of the ability to solve problems is possible.

Large pretrained language models are being utilized to handle open domain conversations.

After pre-training a model on large-scale social comment chain data to primarily model conversational behavior, we fine-tune a variety of target datasets to improve human engagement and humanity. To avoid unwanted behavior of the model including toxicity and bias in human-to-human conversations, automatic filtering by predefined criteria can be used to exclude parts of the training data.

Synthetic Dialogue Creation There are many approaches to synthetic dialog creation to reduce the cost of dialog aggregation, and they typically define task schemas, rules, and templates to simulate specific scenarios in task-oriented dialogue. However, to transfer an invisible target domain, training data from the source domain is required.

In a task-oriented dialog system, the system side plays a functional role that utilizes the explicit knowledge base of a particular domain. For example, agents act as reservation helpers or information providers in various domains such as restaurants and hotels. Each conversational agent can be assigned an explicit persona to promote more specific and consistent responses in an open domain conversational environment. However, it is only based on the persona given in every conversation session, and moreover, the persona given by a few natural language sentences is not sufficient to represent a specific role in a real scenario.

Conversational systems, such as chatbots, may have system policies on whether or not to allow certain types of speech, even if no goals or tasks are specified other than actively participating in the conversation.

In this embodiment, by building an open domain dialogue model in which roles are assigned, a dialogue model in which a system can naturally communicate with a human on an open domain topic while maintaining consistency in a specific role can be built.

To this end, certain conditions of persona, style, safety, and system policies must be met, and a framework capable of data extensibility can be provided.

In this embodiment, it is possible to provide a human-AI collaborative data construction method for creating an extensible supervisory dataset for an open domain conversation that satisfies a role. In particular, a large-scale language model can be leveraged to create an entire conversational session between a user and a system with a contextual snapshot learning approach.

3 illustrates a framework for building a role-specified open domain dialog model in an embodiment of the present invention.

Referring to FIG. 3 , a framework for collecting supervision data for constructing a role-specified open domain dialogue model is shown.

Referring to FIG. 3 , (1) a chatbot developer provides role specifications of a desired chatbot and several examples of conversations.

(2) The large-scale language model creates the entire conversation dataset, and chatbot developers or crowd workers filter the utterances included in the conversation.

(3) The conversation model corresponding to the chatbot agent is learned through supervised learning on the conversation dataset.

(4) A chatbot developer or crowd worker can provide additional feedback on the chatbot's remarks while having a one-on-one chat with the chatbot.

Inputs to the framework are role specifications specified by chatbot developers, which define the constraints of conversational interactions on the system. We assume that the human-bot conversation dataset for that specification may not be available due to the lack of a conversation system to interact with the specific condition and the lack of public conversation data. Because the range of conversations is very wide and diverse due to the nature of open domain conversations, it is also impossible to write all conversation cases manually.

Therefore, we focus on constructing an entire conversational dataset using minimal conversational examples by leveraging raw shot learning in the context of a large-scale language model.

4 is a flowchart illustrating an example of a method for creating an open domain conversation performed by a computer device according to an embodiment of the present invention.

The open domain conversation creation method according to the present embodiment may be performed by the computer device 100 described above. In this case, the processor 120 of the computer device 100 may be implemented to execute a control instruction according to an operating system code or at least one program code included in the memory 110 . Here, the processor 120 causes the computer device 100 to perform the steps (S410 to S430) included in the open domain dialog generation method of FIG. 4 according to a control command provided by a code stored in the computer device 100. The device 100 can be controlled.

Referring to FIG. 4 , in step S410, the processor 120 may generate an entire conversation dataset using a prompt that is an input sentence of a language model. In particular, the processor 120 may generate conversation data of a specific nature through the language model by inputting a prompt including a brief description of chatbot properties and an example of a conversation exchanged between the chatbot and a human to the language model.

The processor 120 may attach several dialog examples that satisfy given specifications to prompts serving as input statements of the language model, and an appropriate system description for constructing inputs to the language model.

5 is an exemplary diagram for explaining a data configuration process according to an embodiment of the present invention.

In FIG. 5 (a) shows an example of the input prompt 500. The prompt 500 may include a brief description of chatbot properties 510 and an example of a conversation 520 exchanged between the chatbot and a person.

The processor 120 may input the prompt 500 to the language model to generate natural language dialogue data from the language model. In other words, the processor 120 may obtain dialogue data as a result of language generation through a language model generation or completion function after inputting the prompt 500 input sentence into the language model. The processor 120 may input the prompt 500 as a language model, analyze the natural language pattern of the dialogue example included in the prompt 500 through the language model, and obtain new dialogue data having the corresponding pattern.

That is, the processor 120 may provide both the system specified from the pattern of the prompt 500 and the entire conversation session of the user role through the language model. In FIG. 5, (b) shows an example of a new conversation generated with a pattern of prompts (a) through a language model.

Accordingly, the processor 120 may generate conversation data of a natural pattern for an open domain topic while maintaining consistency in a designated role using a language model. Conversation data generated through the language model can be used as data for learning an open domain dialogue model.

Referring back to FIG. 4 , in step S420, the processor 120 may filter the conversation data generated in step S410 through annotations of chatbot developers (or crowd workers). It is difficult to include all the details of a specification in a prompt and reflect it in data generation. Accordingly, it is possible to filter conversation data generated through a language model using human annotations. The processor 120 may provide each chat session generated through the language model to an annotator corresponding to a chatbot developer, and request labeling of problem statements out of the scope of the chat using the annotation.

6 shows an example of filtering conversation data generated by a language model. Chatbot developers can specify separate labels through annotations for some of the chatbot utterances that are evaluated as problem utterances in the conversation data generated through the language model. At this time, the processor 120 uses the annotated problem statement as the negative example 601 in reflecting the filtering result on the training data of the open domain conversation model, while the problem statement corresponding to the negative example 601 Based on , at least one utterance prior to the problem utterance may be used as a positive example 602 .

In the dialog data generated by the language model, the context after the problem statement annotated by the chatbot developer is highly likely to contain damaged content, so it is dropped from being used for learning the open domain dialog model.

Referring back to FIG. 4, in step S430, the processor 120 transfers conversation data exchanged between the chatbot developer and the chatbot in such a way that the chatbot developer (or crowd worker) directly participates in a conversation with the chatbot in the open domain conversation model. It can be collected as learning data. The processor 120 may engage an open domain conversation model in a multi-turn conversation with a person in the step of constructing a dataset. A chatbot developer, as a system user, can communicate with the chatbot in turn, and can correct inappropriate problem statements among the statements provided as responses from the chatbot during the conversation process.

For example, the processor 120 may provide an interface through which a chatbot developer can directly correct a problem statement of a chatbot during a conversation between the chatbot developer and the chatbot.

As another example, if the processor 120 receives a correction request from the chatbot developer for a problem statement of the chatbot during a conversation between the chatbot developer and the chatbot, the processor 120 calls a language model to generate an alternative utterance capable of correcting the problem statement through the language model. can The processor 120 may correct the chatbot's problem statement into an alternative statement according to the chatbot developer's request, and then proceed with the next conversation, and when a correction request is received from the chatbot developer again, the processor 120 may regenerate the alternative statement.

7 is an exemplary diagram for explaining a process of collecting conversation data between a chatbot and a person in one embodiment of the present invention.

7 shows a web-based conversation collection interface screen 700 .

As shown in FIG. 7 , the processor 120 may provide a dialog collection interface screen 700 to collect learning data of an open domain dialog model. A chatbot developer can exchange and exchange a conversation with the chatbot in turn through the conversation collection interface screen 700 .

The conversation collection interface screen 700 may include a 'modify' button 701 for requesting correction of remarks regarding problematic remarks outside the dialogue range during a conversation with the chatbot. A chatbot developer may select a problem type and input a 'modify' button 701 when the chatbot's remarks do not match the chatbot specifications or roles.

When the chatbot developer selects the 'modify' button 701 for the problem statement among the chatbot's speech, the processor 120 may call a language model to generate an alternative speech capable of correcting the problem statement through the language model. The chatbot developer may proceed with the next conversation if the alternative remark is appropriate, and may repeatedly request correction of the remark through the 'modify' button 701 if the alternative remark is not appropriate.

The conversation collection interface screen 700 may further include a 'save conversation' button 702 for saving conversation data between the chatbot developer and the chatbot. The chatbot developer may save the entire chat session through the 'save conversation' button 702 when there are no utterances out of the conversation scope or the conversation is ended with all of them corrected.

The processor 120 may use the entire conversation session exchanged between the chatbot developer and the chatbot as training data of the open domain conversation model when some utterances are corrected and the conversation is finished. At this time, a utterance corrected as an alternative utterance during a conversation can be used as a positive example, and a utterance before correction, that is, a problem utterance requested for correction through the 'correction' button 701 can be used as a negative example.

Furthermore, the processor 120 may utilize a process of correcting the chatbot's problem statement as an evaluation index for the chatbot. Since the input of the 'modify' button 701 means that an inappropriate utterance is returned, it can be used for the system error rate, that is, the corrected response rate among all responses.

The architecture of the role-assigned open domain dialog model is as follows.

Various architectures can be applied to model a dialog system that satisfies the role specification using the previously created dialog dataset.

The open domain conversation model according to the present invention may include an Out-of-Bounds Detection Model.

The simplest way to limit utterances based on a chatbot's role specifications is to detect and discard utterances that fall outside of category boundaries. As an example, a BERT (Bidirectional Encoder Representations form Transformer) based binary classifier can be used to classify positive and negative examples in a conversation dataset. Since the classifier cannot carry out a conversation alone, a two-stage model is assumed, and the response prediction model returns a response that is censored by the classifier. When out-of-category utterances are detected, one of the predefined questions on other, similar topics can be selected and returned, using the language model to select the question with the lowest perplexity (PPL).

An open domain dialog model according to the present invention may include a response selection model.

One of the ways to restrict speech according to the chatbot's role specification is to pre-filter the response candidates for the response selection model. As an example, we can use a two-step approach to retrieve-then-rerank with a response selection model. The retriever of the poly-encoder architecture finds the top k responses from the response candidates and then reranks the top k responses by the reranker of the cross-encoder architecture.

Since the filtering of response candidates has limitations, it is important to determine the context in which an answer cannot be answered together with the response candidates. One effective way to predict unanswerable contexts is to take advantage of the model's uncertainty, for example a similar approach using thresholds. If all retrieved response scores are lower than a certain threshold, we can predict an unresponsive context. Another example is another approach using the PPL of the language model, which connects the dialog context and the searched response to input into the language model, measures the PPL of the response, and finally determines the response based on a threshold value.

An open domain dialog model according to the present invention may include a response generation model.

Language model fine-tuning for target data is effective in learning the characteristics of tasks. The fine-tuned language model can be considered as a response generation model using Maximum Likelihood Estimation. On the other hand, unlikelihood learning is effective in mitigating undesirable characteristics of generative models (eg token repetitions, logical inconsistencies, etc.).

In this embodiment, maximum likelihood estimation is applied to the positive examples in the dataset so that the chatbot generates utterances of desirable characteristics, while non-likelihood learning can be applied to negative examples so that the chatbot does not generate utterances of undesirable characteristics. . The above two types of learning can be performed simultaneously.

The open domain dialog model according to the present invention may include a retrieve-fail-generate model.

In constructing an open domain dialogue model, a search-failure-generation pipeline combining the above-described response selection model and response generation model may be constructed.

8 illustrates a search-failure-generation model structure in one embodiment of the present invention.

Referring to FIG. 8 , a search-failure-generate model 800 is a response selection model 810 that uses a two-step approach of retrieve-then-rerank, and a response selection model 810 that generates appropriate utterances while simultaneously generating appropriate utterances. It can be configured as a response generation model 820 that has been trained not to generate unintentional utterances.

At this time, the response selection model 810 tries to select an appropriate response, and if the response selected by the unresponsive context prediction model is ignored, the response generation model 820 returns a response for the given context. It is relatively easy to control the response selection model 810 by managing response candidates. Accordingly, the response selection model 810 is responsible for most responses, and the response generation model 820 is used only when response selection fails.

The present embodiments can create a conversation dataset that satisfies the role through a data collection framework that utilizes contextual snapshot learning of the language model, and expands the conversation dataset through data filtering using annotations and direct conversation between humans and chatbots. can do it We can use these dialog datasets as training data to model dialog systems that meet role specifications.

As described above, according to embodiments of the present invention, a dialog system with improved response quality can be provided by establishing an open domain dialog model capable of maintaining a consistent role while having a natural conversation with a person.

The devices described above may be implemented as hardware components, software components, and/or a combination of hardware components and software components. For example, devices and components described in the embodiments include a processor, a controller, an arithmetic logic unit (ALU), a digital signal processor, a microcomputer, a field programmable gate array (FPGA), and a programmable PLU (programmable logic unit). logic unit), microprocessor, or any other device capable of executing and responding to instructions. The processing device may run an operating system (OS) and one or more software applications running on the operating system. A processing device may also access, store, manipulate, process, and generate data in response to execution of software. For convenience of understanding, there are cases in which one processing device is used, but those skilled in the art will understand that the processing device includes a plurality of processing elements and/or a plurality of types of processing elements. It can be seen that it can include. For example, a processing device may include a plurality of processors or a processor and a controller. Other processing configurations are also possible, such as parallel processors.

Software may include a computer program, code, instructions, or a combination of one or more of the foregoing, which configures a processing device to operate as desired or processes independently or collectively. You can command the device. The software and/or data may be embodied in any tangible machine, component, physical device, computer storage medium or device to be interpreted by or to provide instructions or data to a processing device. there is. Software may be distributed on networked computer systems and stored or executed in a distributed manner. Software and data may be stored on one or more computer readable media.

The method according to the embodiment may be implemented in the form of program instructions that can be executed through various computer means and recorded on a computer readable medium. In this case, the medium may continuously store a program executable by a computer or temporarily store the program for execution or download. In addition, the medium may be various recording means or storage means in the form of a single or combined hardware, but is not limited to a medium directly connected to a certain computer system, and may be distributed on a network. Examples of the medium include magnetic media such as hard disks, floppy disks and magnetic tapes, optical recording media such as CD-ROM and DVD, magneto-optical media such as floptical disks, and ROM, RAM, flash memory, etc. configured to store program instructions. In addition, examples of other media include recording media or storage media managed by an app store that distributes applications, a site that supplies or distributes various other software, and a server.

As described above, although the embodiments have been described with limited examples and drawings, those skilled in the art can make various modifications and variations from the above description. For example, the described techniques may be performed in an order different from the method described, and/or components of the described system, structure, device, circuit, etc. may be combined or combined in a different form than the method described, or other components may be used. Or even if it is replaced or substituted by equivalents, appropriate results can be achieved.

Therefore, other implementations, other embodiments, and equivalents of the claims are within the scope of the following claims.

Claims (13)

In a method executed on a computer device,
The computer device includes at least one processor configured to execute computer readable instructions contained in a memory;
The method,
configuring, by the at least one processor, a prompt that is an input sentence of a language model with a role specification of the chatbot and an example of a conversation between the chatbot and a person; and
generating, by the at least one processor, new dialogue data of roles and dialogue patterns included in the prompt through the language model by inputting the prompt into the language model;
How to include. According to claim 1,
The method,
Building, by the at least one processor, an open-domain dialogue model for the chatbot by supervised learning on the dialogue data.
How to include more. According to claim 1,
The method,
filtering, by the at least one processor, the dialogue data through annotations;
How to include more. According to claim 3,
The filtering step is
Classifying the annotated problem speech among the speech of the chatbot included in the conversation data as a negative example for learning the chatbot.
How to include. According to claim 1,
The method,
Collecting, by the at least one processor, conversation data between the chatbot and a person as learning data of the chatbot through an interface in which the chatbot and a person directly communicate.
How to include more. According to claim 5,
The collecting step is
modifying some of the utterances by generating replacement utterances through the language model when a modification request is received for some of the utterances of the chatbot provided through the interface;
How to include. According to claim 6,
The collecting step is
Classifying the utterance requested for correction as a negative example for learning the chatbot; and
Classifying the utterance modified as the replacement utterance as a positive example for learning the chatbot.
How to include more. According to claim 2,
The building step is
Based on the perplexity (PPL) of the language model in an Out-of-Bounds Detection Model (Out-of-Bounds Detection Model) that detects utterances outside the scope of the role specification and a retrieve-then-rerank approach after finding response candidates A response selection model for filtering the response candidates, and response generation in which maximum likelihood estimation is applied to positive examples in the dataset and unlikelihood learning is applied to negative examples in the dataset. Building the domain conversation model using at least one of the models (Response Generation Model)
A method characterized by. A computer program stored in a computer readable recording medium to execute the method of any one of claims 1 to 8 in the computer device. In a computer device,
at least one processor configured to execute computer readable instructions contained in memory;
including,
The at least one processor,
Construct a prompt that is an input sentence of a language model with a role specification of the chatbot and an example of a conversation between the chatbot and a person,
Inputting the prompt to the language model to generate new conversation data of roles and conversation patterns included in the prompt through the language model.
Characterized by a computer device. According to claim 10,
The at least one processor,
Building an open domain conversation model for the chatbot by supervised learning on the conversation data
Characterized by a computer device. According to claim 10,
The at least one processor,
filtering the conversation data through annotations;
Characterized by a computer device. According to claim 10,
The at least one processor,
Collecting conversation data between the chatbot and a person as learning data of the chatbot through an interface in which the chatbot and a person directly communicate
Characterized by a computer device.

Images