KR102480872B1 - Method for generating data for machine learning and method for scheduling machine learning using the same - Google Patents

Abstract

A method for generating training data for machine learning of an efficient chatbot and a method for scheduling machine learning are provided. A method for generating data for machine learning according to an embodiment of the present invention includes a data classification step of classifying data to be learned according to attributes of the data, and generating data for learning from data to be learned according to the attributes of the data to be learned. Include steps. Attributes of data may include a data generation cycle and/or a data validity period. In the machine learning scheduling method according to an embodiment of the present invention, the learning data generated using the above-described machine learning data generation method is loaded into the learning data DB at an appropriate time according to the properties of the learning data, and the machine learning of the chatbot engine A machine learning control step of controlling to execute is further provided.

Description

Method for generating data for machine learning and scheduling method for machine learning using the same {Method for generating data for machine learning and method for scheduling machine learning using the same}

The present invention relates to a method for generating data for machine learning and a scheduling method for machine learning using the same, and more particularly, to a method for generating training data for machine learning of an efficient chatbot and a method for scheduling machine learning.

An increasing number of companies are trying to apply a chatbot system that uses artificial intelligence to respond to customer consultations or inquiries through phone calls or messengers. The chatbot receives a voice or text message from the user, extracts the user's intent from it, and outputs a response corresponding to the extracted intent. The performance of a chatbot is influenced not only by the machine learning algorithm that trains the chatbot, but also by the quantity and quality of training data for machine learning (hereinafter referred to as 'data for machine learning').

With the development of artificial intelligence technology, simple responses using chatbots have become possible, but users' satisfaction is still low. This is due not only to the difficulty in extracting accurate intent from complex queries, but also to the fact that appropriate data are often not fetched even for relatively simple questions.

For example, Patent Registration No. 2119468 proposes a method of using counseling knowledge of a counselor as counseling knowledge of a chatbot by machine learning the counseling chatbot based on the counselor's counseling contents. However, in the modern situation where new information is quickly created and changed, the counselor's counseling content may be based on outdated information. Therefore, even if a chatbot is learned in this way, learning based on incorrect information may occur.

To improve these points, it is recommended to run machine learning as often as possible, but machine learning takes a lot of time when there is a lot of data, making it difficult to obtain the desired result.

The present invention has been made in view of these points, and an object of the present invention is to provide a method for generating data for machine learning that can provide appropriate data to a chatbot system and a method for scheduling machine learning using the same.

Another object of the present invention is to provide a method for generating data for machine learning that can provide appropriate data to a chatbot system while reducing the load associated with performing machine learning, and a machine learning scheduling method using the same.

A method for generating data for machine learning according to an embodiment of the present invention includes a data classification step of classifying data to be learned according to attributes of the data, and generating data for learning from data to be learned according to the attributes of the data to be learned. Include steps. Data attributes may include data generation cycle and/or data validity period.

The training data may have the same properties as the properties of the data to be learned used to generate the training data. The learning data may be stored in a database separate from learning data of other properties according to properties, or may be stored separately from each other, such as having a different field.

In one embodiment, a template may be used in the step of generating training data. The template is composed of a plurality of queries and corresponding responses, and the responses include parts that can be replaced according to data to be learned. Data for learning may be generated by substituting the replaceable part according to data to be learned.

The method for generating data for machine learning according to an embodiment of the present invention further includes deleting data for learning whose validity period has expired.

In the machine learning scheduling method according to an embodiment of the present invention, the learning data generated using the above-described machine learning data generation method is loaded into the learning data DB at an appropriate time according to the properties of the learning data, and the machine learning of the chatbot engine A machine learning control step of controlling to execute is further provided.

In one embodiment, in the machine learning control step, once machine learning is executed with learning data, control is performed so that learning is not executed again with the same data within a generation period of the corresponding learning data.

In one embodiment, the machine learning scheduling method of the present invention further includes controlling to execute machine learning with the training data at least every minimum training pause period regardless of the generation cycle attribute of the training data.

According to the present invention, since learning data capable of efficient training is generated from data to be learned according to the properties of the data to be learned, efficient machine learning of the chatbot is possible.

In addition, since the timing of machine learning is controlled according to the data generation cycle, there is no need to train with all the training data, greatly reducing the time required for machine learning once.

1 is a block diagram showing an example in which the present invention is applied to a chatbot system.
2 is a functional block diagram of a machine learning scheduler according to a preferred embodiment of the present invention.
3 is a flowchart showing a process of generating data for machine learning and scheduling machine learning according to a preferred embodiment of the present invention.

The information that users inquire through chatbots includes information that is almost unchanged, while there is information that changes from moment to moment. For example, the name of an expressway such as 'Gyeongbu Expressway' and the length of a section hardly change, but the average speed or weather of a specific section of the Gyeongbu Expressway changes from moment to moment.

In addition, while there is data that changes periodically, there is data that changes irregularly. For example, while the communication fee paid by the user is renewed every month, the basic communication fee may be changed irregularly.

In the present invention, considering the unique variability and cycle of each data, data is provided to the machine learning device at a cycle suitable for each data, thereby reducing the load on the learning system and allowing appropriate machine learning to be performed.

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

1 is a block diagram showing an example in which the present invention is applied to a chatbot system.

In a general chatbot system, a user's conversation converted into text or a conversation entered as text by the user is input to the chatbot engine 10, and the chatbot engine 10 responds to the user's conversation according to the trained learning result. For example, a language model such as BERT (Bidirectional Encoder Representations from Transformers) may be used in the chatbot engine 10 . BERT is a machine translation neural network using a self-attention technique, and can improve the performance of artificial intelligence conversations in a specific field through pre-training embedding. By performing machine learning on the chatbot engine 10 periodically or frequently using the machine learning data DB 20, the performance of the chatbot engine 10 is improved and the latest information can be reflected.

In the present invention, the machine learning scheduler 100 processes raw data to be used to train the chatbot engine (hereinafter referred to as 'data to be learned'), and training data suitable for efficiently training the chatbot engine 10 (hereinafter referred to as 'learning data') ') is generated, and the chatbot engine 10 is machine-learned according to the properties of the data using the generated learning data.

Next, the configuration of the machine learning scheduler 100 according to a preferred embodiment of the present invention will be described with reference to FIG. 2 .

The data classification unit 110 classifies input data to be learned according to its properties. Attributes of the data may include a generation period of the data, a validity period of the data, and the like. The data to be learned may be automatically generated by a computer, a sensor device, or the like, or may be created by a person.

Classification of data according to attributes may be determined according to data input sources, or may be determined from words included in the title or content of a document including data. For example, data from a place that provides daily data may be determined to have a period of one day. For example, if the title of the document includes 'annual report', the period of data can be determined as annual. Alternatively, after learning the documents whose properties have been set by a person with an artificial intelligence algorithm, the properties of the input documents may be set by the learned artificial intelligence engine.

The data generation cycle represents a generation cycle of periodically generated data. For example, a company's business report is annual, and its quarterly report is quarterly. The daily number of deaths from traffic accidents, daily precipitation, and daily traffic volume, etc., which are counted every day, have a daily generation cycle.

The validity period of the data indicates the validity period of the generated data. For example, when tolls are free during the Chuseok holiday period, toll information for the corresponding period is valid only during the Chuseok holiday period. If traffic is restricted due to construction in some road sections, the information is valid only during the construction period. Some data do not have a fixed expiration date. For example, since the validity period of a certain company's corporate registration number, address, etc. is not determined, the validity period may be set as undetermined.

The data classification unit 110 classifies input attributes of data to be learned and stores them in different databases 120a, .., 120n (or files) according to the attributes, or sets attribute fields of data to be learned. 2 illustrates a periodic database, that is, a real-time data DB 120a generated in real time, a daily data DB 120b generated every day, a weekly data DB 120c generated weekly, and an annual data DB 120n generated annually. showing hostility.

The learning data processing unit 130 converts the classified data to be learned into learning data according to attributes. The converted learning data may also have the same properties as the data to be learned. The converted training data may be stored separately from learning data of other attributes according to attributes. In addition, the learning data processing unit 130 may delete the learning data for which the validity period has expired so that they are not used in the next learning.

Table 1 shows an example of converting annual data with a generation cycle of 1 year into training data.

annual data data for training 67 public holidays (2020) question:
How many holidays are there this year?
How many holidays are there this year?
How many public holidays are there in 2020?
What are the number of public holidays this year?
How many holidays are there this year?
What are the number of public holidays in 2020?
How many holidays are there this year?
How many holidays are there this year?
...
A: There are 67 public holidays this year.

In addition, when new annual data is generated, existing annual data can be converted as shown in Table 2. That is, since information corresponding to this year has become information corresponding to last year over the years, training data is generated accordingly.

annual data data for training 66 public holidays (2019) question:
How many holidays were there last year?
How many holidays were there last year?
How many public holidays are there in 2019?
What was the number of public holidays last year?
What was the number of public holidays last year?
What are the number of public holidays in 2019S?
How many holidays were there last year?
How many holidays were there last year?
...
Answer: The number of public holidays last year was 66.

Table 3 shows an example of processing daily data.

daily data data for training 5 deaths in traffic accidents (November 30, 2020) question:
How many people died in traffic accidents yesterday?
How many people died in traffic accidents yesterday?
How many car accidents died yesterday?
Please tell me the number of deaths in traffic accidents yesterday.
...
Answer: Yesterday's traffic accident death toll was 5.

Similarly, when new daily data is generated, existing daily data can be converted as shown in Table 4.

daily data data for training 3 deaths in traffic accidents (November 29, 2020) question:
How many people died in traffic accidents?
How many people died in a car accident the other day?
How many car accident deaths occurred the day before yesterday?
Please tell us the number of deaths in traffic accidents the day before yesterday.
...
Answer: The number of deaths in traffic accidents the day before yesterday was three.

In one embodiment, the data processing unit 130 for learning may generate data to be learned using a template.

For example, a query and response set of data for annual learning may be generated using a template as illustrated in Table 5. A template consists of a query and a corresponding response, and the response includes parts that can be replaced according to data. For example, if data to be learned indicating that the number of holidays this year is 67 is input, data for annual learning can be generated by replacing the '#' part of Table 5 with data of '67'.

vaginal answer How many holidays are there this year? The number of public holidays this year is #day. How many holidays are there this year? The number of public holidays this year is #day. How many public holidays are there in 2020? The number of public holidays in 2020 is #day. What are the number of public holidays this year? The number of public holidays this year is #day. How many holidays are there this year? The number of public holidays this year is #day. What are the number of public holidays in 2020? The number of public holidays in 2020 is #day. How many holidays are there this year? The number of public holidays this year is #day. How many holidays are there this year? The number of public holidays this year is #day. ... ...

Templates can be created using artificial intelligence or manually based on existing actual Q&A data, or can be automatically additionally created based on already prepared templates using synonym dictionaries and verb usage patterns.

The scheduler 140 loads the learning data created in the learning data processing unit 130 into the learning data DB 20 at an appropriate time according to the properties including the annual, daily, etc. generation cycle of each learning data, and the chatbot engine 10 run the machine learning of For example, for annual learning data, once data is created and learning is performed, learning is not performed with the data for one year. For monthly learning data, learning is not performed with the same data for one month after learning is executed. Data for daily learning runs daily learning. Data for real-time learning is included every time machine learning is performed and learning is executed.

Depending on the embodiment, a minimum training suspension period may be set to prevent being excluded from training for too long. In this case, training is performed every minimum training pause period regardless of the attributes of the learning data. For example, if the minimum training suspension period is 6 months, training may be performed on all learning data at least every 6 months.

According to the present invention, since the machine learning timing is controlled according to the data generation cycle, it is not necessary to train with all training data, so the time required for one machine learning is greatly reduced. In addition, for data having an attribute of the expiration date, the data can be deleted when the expiration date has elapsed, thereby preventing unnecessary data from being used for machine learning.

Next, referring to FIG. 3, a process of generating data for machine learning and scheduling machine learning according to a preferred embodiment of the present invention will be described.

When data to be learned is input, the machine learning scheduler 100 classifies the data to be learned according to the properties of the data. Attributes of the data may include a generation period of the data, a validity period of the data, and the like. Classification of data according to attributes may be determined according to data input sources, or may be determined from words included in the title or content of a document including data. Alternatively, after learning the documents whose properties have been set by a person with an artificial intelligence algorithm, the properties of the input documents may be set by the learned artificial intelligence engine. Classified data is stored in different databases (or files) according to attributes, or attribute fields are set.

The machine learning scheduler 100 generates learning data from data to be learned according to the properties of the data to be learned. Data for learning may be generated through the procedure described with reference to Tables 1 to 5, for example.

The machine learning scheduler 100 loads the learning data into the learning data DB 20 at an appropriate time according to the attributes including the annual, daily, etc. period of each learning data and controls the chatbot engine 10 to execute machine learning do. For example, for annual learning data, once data is created and learning is performed, learning is not performed with the data for one year. For monthly learning data, learning is not performed with the same data for one month after learning is executed. Data for daily learning runs daily learning. Depending on the embodiment, a minimum training suspension period may be set to prevent being excluded from training for too long. In this case, training is performed for each minimum training pause period regardless of the attributes of learning data.

In the above, the present invention has been described with several examples, but even if all components constituting an embodiment of the present invention are combined or described as operating in combination, the present invention is not necessarily limited to these embodiments. That is, within the scope of the object of the present invention, all of the components may be selectively combined with one or more to operate. In addition, although all of the components may be implemented as a single independent piece of hardware, some or all of the components are selectively combined to perform some or all of the combined functions in one or a plurality of hardware. It may be implemented as a computer program having. Codes and code segments constituting the computer program may be easily inferred by a person skilled in the art. Such a computer program may implement an embodiment of the present invention by being stored in a computer or processor-readable storage medium, read and executed by a computer or processor.

Terms such as "comprise", "comprise" or "having" described above mean that the corresponding component may be present unless otherwise stated, and therefore do not exclude other components. It should be construed that it may further include other components.

The above description is merely an example of the technical idea of the present invention, and various modifications and variations can be made to those skilled in the art without departing from the essential characteristics of the present invention. Therefore, the embodiments disclosed in the present invention are not intended to limit the technical idea of the present invention, but to explain, and the scope of the technical idea of the present invention is not limited by these embodiments. The protection scope of the present invention should be construed according to the claims below, and all technical ideas within the equivalent range should be construed as being included in the scope of the present invention.

10 chatbot engine,
20 Data DB for learning,
100 Machine Learning Scheduler,
110 data classification unit;
120a,..120n periodic database,
130 learning data processing unit;
140 Scheduler.

Claims (10)

delete delete delete delete delete A machine learning scheduling method in a machine learning scheduler for machine learning a chatbot engine of a chatbot system, the method comprising:
A data classification step of classifying data to be learned by the machine learning scheduler according to data properties including data generation cycle and data validity period;
a learning data generation step in which a machine learning scheduler generates learning data having the same properties as the properties of the data to be learned from the data to be learned according to the properties of the data to be learned;
Machine learning control step in which the machine learning scheduler loads the learning data into the learning data DB at the time corresponding to the generation cycle property of each learning data and controls the machine learning of the chatbot engine to be executed
Including,
The machine learning scheduler controls the learning data for which learning has been executed once in the machine learning control step so that learning is not executed again with the corresponding data within a data generation cycle according to the data generation cycle attribute of the learning data. Controls to execute machine learning with training data at least every minimum training pause period regardless of the data generation cycle attribute of , deletes learning data whose validity period has expired according to the validity period attribute,
The data generation period property includes a real-time property, and the learning data of the real-time property is included every time machine learning is performed and controlled to execute the machine learning scheduling method.
delete delete According to claim 6,
In the generating data for learning step, data for learning is created from data to be learned using a template;
The template is composed of a plurality of queries and corresponding responses, and the responses include parts that can be replaced according to data to be learned.
A machine learning scheduling method for generating learning data by substituting the replaceable part according to data to be learned.

delete

Images