KR102240910B1 - Korean Customer Service Associate Assist System based on Machine Learning - Google Patents

Abstract

A Korean customer service assistant assistance system based on machine learning is disclosed.
The system includes a server of a machine learning-based customer service assistant assistance system; A database storing question-response pairs of frequently asked questions (FAQs) of customers; And when a customer question is input, the question is transmitted to the server, the machine learning module searches the database and compares the customer question with the question in the predefined FAQ list, and breaks the Korean sentence of the customer question into phoneme units. And embedding each phoneme into a vector, using a BiGRU machine learning model to learn the similarity of the two sentences by using the sentence similarity model learning algorithm of the two sentences of the customer question and the FAQ question. It includes a sentence similarity model machine learning system that presents the highest five FAQs to the counselor in multiple choice, and automatically converts them to learning data and outputs answers to the corresponding questions when the counselor selects one of the FAQs.

Description

Korean Customer Service Associate Assist System based on Machine Learning}

The present invention relates to a machine learning-based Korean customer service assistant assistance system, and more particularly, to a natural language processing technology of artificial intelligence and a bidirectional gated recurrent unit (BIGRU) machine learning for sentence similarity measure of machine learning. A machine learning engine and a learning algorithm of a sentence similarity model are used, and the customer service system compares the questions of two sentences in a predefined FAQ question-response pair stored in the server's database for the customer's question. The top 5 FAQs with the highest value are presented to the agent in multiple choice, and the agent selects one of the matching question-answer pairs among the questions in the top 5 FAQs, and the machine learning sentence in the server's database The present invention relates to a machine learning-based Korean customer service assistant assistance system that automatically converts the counselor's choices into training data according to the similarity model and outputs a response to customer questions.

According to Digital Consumer Interaction Report 1, in a customer service system, in a situation where a customer needs to file a complaint or find a solution, an AI Chatbot that uses pre-learning of user question-and-answer patterns using Natural Language Processing (NLP). He says he prefers to talk with the employee rather than). However, only 48.5% of them were able to get the answer through a single query. In addition, according to a survey on Talkdesk2, 42% of customers chose "fast response rate" as their most important issue.

During natural language processing (NLP) processing, sentences detect language and are classified into morpheme units through morphological analysis according to the characteristics of the language, and the selected morpheme is a sentence vector (sentence). vectors). For example, a sentence vector is composed of attributes extracted from each question from FAQs classified into four query types. A similarity analysis is performed between the user query sentence vector and the previously collected query sentence vector.

Natural Language Processing (NLP) is used in automatic translation, speech recognition, search engines, and chatbots.

AI chatbots are conversations in which, for example, when you enter a question as if you are chatting on a messenger, an artificial intelligence (AI) module searches a database to analyze big data, communicates with people in everyday language, and outputs the result of answering the question. Used for brother messenger.

When answering a question, the chatbot consists of the steps of collecting question data-searching the question database-machine learning algorithm-and answering the question.

In addition, AI chatbots are used to increase customer satisfaction in customer service systems. According to recent research, chatbot technology is developing intensively [1], and chatbots can replace employees [2]. However, chatbots also have disadvantages such as unnatural conversations [3] and inconsistency with demands [4].

There was a study to estimate the intelligence of chatbots at the level of a 6-year-old child [5]. According to this study, there are obvious limitations to using AI chatbots directly for customer service.

In addition, the AI voice recognition chatbot receives voice query data, receives STT (Speech To Text) text data after voice recognition, searches the database, analyzes big data, and machine learning the answer result. Execute TTS (Text To Speech) to provide quick answers with response result texts and voice data.

Various sentence classification algorithms using AI indirectly have been disclosed [6-8].

[6] Yoon Kim's Traditional Neural Networks for Sentence Classification, Computation and Language, "Convolutional Neural Networks for Sentence Classification, Computation and Language," arXiv preprint arXiv:14085882, 2014.

[7] S Lai, L Xu, K Liu and J Zhao, Recursive-Convolutional Neural Networks for Text Classification (R-CNN), "Recurrent Convolutional Neural Networks for Text Classification," AAAI, pp 2267-2273, 2015.

[8] T Joachims, Text categorization with support vector machines: Learning with many relevant features," European conference on machine learning, 1998.

However, these methods have the disadvantage that the entire training process must be repeated when the class is modified. In addition, the Rule-based Chatbot grasps the user's intent through predefined keywords and patterns, and uses natural language processing (NLP) to pre-learn user query patterns and , A large amount of training data is essential to achieve the desired level of accuracy by analyzing the similarity with the learned queries, and generating pre-trained and formatted training data. It is not easy.

As prior art 1 related to this, in Patent Registration No. 10-1130444, a "similar sentence identification system using machine translation technique" registered by Microsoft Corporation is disclosed.

The identification of similar sentences in text is in many applications of natural language processing (NLP) systems. An example of identification and generation of similar sentences is a question answering system.

Prior art 1 obtains a set of text segments from a cluster of different articles written for the same event. The set of text segments then performs a text alignment technique to identify similar sentences from the text segments in the text. Also, it can be used to generate similar sentences.

As a method of training a paraphrase processing system,

Accessing a plurality of documents;

Identifying, from the plurality of documents, a cluster of related texts written by different authors on a common subject-the cluster of related texts is also identified as being from different news agencies and related to a common event. Become -;

Receiving the cluster of related texts;

Selecting a set of text segments from the cluster, wherein the selecting comprises grouping required text segments among the related texts into a set of related text segments; And

Identifying a similar sentence relationship between texts in text segments included in the set of related text segments using textual alignment,

The text alignment includes aligning words in text segments in the set of related text segments using statistical textual alignment, and identifying the pseudo-sentence relationship based on the aligned words. do.

As prior art 2 related to this, in Patent Registration No. 10-1991486, "an apparatus and method for expanding a multi-word database based on sentence similarity" is disclosed.

The multilingual database expansion device based on sentence similarity

As a device that expands the multilingual knowledge database based on sentence similarity,

Selecting an expansion target word, which is a word desired to be expanded by adding a usage example for the meaning of a multi-word included in the multi-language knowledge database, collecting at least one document containing the expansion target word, and the collected document A document collection unit that extracts only the expansion target sentences including the expansion target words;

Selecting a word with multiple meanings (hereinafter referred to as'multilingual analysis target word') used for sentence similarity calculation to add usage examples for each meaning of the expansion target word to the multilingual knowledge database, and A morpheme analysis is performed on the usage examples by meaning and the at least one extension target sentence, and a similarity is calculated by comparing the meaning by meaning of the multiword analysis target word with the extension target sentence according to the performed morpheme analysis result,

A multi-word recognition unit for classifying a multi-word meaning according to the calculation result;

Calculate the reliability of the classified multi-term meaning, and compare the calculated reliability with a preset threshold to verify whether the expansion target sentence can be used as an example of the multi-term meaning, and the expansion target sentence can be used as an example. A multi-word verification unit to determine the meaning of a multi-word; And

And a database expansion unit for adding the expanded target sentence to a multi-word database as an example for the determined multi-word meaning.

However, the customer service system requires the introduction of an artificial intelligence system for customer centers and call centers, and the question-answer problem of a customer in response to a customer's question is a problem that determines whether the two sentences are similar and takes a lot of response time. . In addition, it did not provide a system that effectively shortens the response time by matching customer inquiries with pre-defined FAQs using machine learning technology.

Patent registration number 10-1130444 (registration date March 19, 2012), "Similar sentence identification system using machine translation technique", Microsoft Corporation Patent Registration No. 10-1991486 (Registration Date June 14, 2019), "Multilingual Database Expansion Device Based on Sentence Similarity and Its Method", Electronics and Telecommunications Research Institute

Zumstein, Dand S Hundertmark, "Chatbots-An Inter-active Technology for Personalized Communication, Transactions and Services", IADIS International Journal on WWW/Internet, 2017 Carl Benedikt Frey and Michael A Osborne, "The future of employment: How susceptible are jobs to computerisation?," Technological Forecasting and Social Change, 2017. Will Knight, "AI's Language Problem," MIT Tech-nology Review, 2016. Coniam, "The linguistic accuracy of chatbots: usability from an ESL perspective," Text & Talk, 2014. Feng Liu, Yong Shi and Ying Liu, "Intelligence quotient and intelligence grade of artificial intell-igence" Annals of Data Science, Volume 4, Issue 2, pp. 179-191, June 2017. Yoon Kim, "Convolutional Neural Networks for Sentence Classification, Computation and Language," arXiv preprint arXiv:14085882, 2014. S Lai, L Xu, K Liu and J Zhao, "Recurrent Convolutional Neural Networks for Text Classification," AAAI, pp 2267-2273, 2015. T Joachims, "Text categorization with support vector machines: Learning with many relevant features," European conference on machine learning, 1998. Kyunghyun Cho, Bart Van Merrienboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio, "Learning phrase reprentations using rnn encoder-decoder for statistical machine translation," arXiv preprint arXiv:14061078, 2014

An object of the present invention for solving the above problem is a machine learning engine of bidirectional gated recurrent unit (BIGRU) machine learning for measuring sentence similarity of natural language processing (NLP) technology of artificial intelligence and machine learning (ML). It uses the learning algorithm of the sentence similarity model and, in the customer service system, compares the questions of two sentences in the predefined FAQ question-response pair stored in the database of the server for the customer question, and has the highest similarity. If the top 5 FAQs are presented to the agent in multiple choice, the agent selects one of the matching question-answer pairs among the top 5 FAQ questions, and the machine learning sentence similarity model is displayed in the server's database. Accordingly, it provides a machine learning-based Korean customer service assistant assistance system that automatically converts the counselor's choices into training data and outputs responses to customer questions.

This study proposed a new customer service assistant assistance system so that customer service representatives can effectively shorten response time by matching customer inquiries with predefined FAQs, and simultaneously generate training data from employee actions for system update. .

)과 Y = (y1,y2,y3,...

)이 주어질 때, 두 점 사이의 유클리드 거리는 d(X,Y)는

계산된다. In order to achieve the object of the present invention, a machine learning-based Korean customer service assistant assistance system includes: a server of a machine learning-based customer service assistant assistance system; A database storing question-response pairs of frequently asked questions (FAQs) of customers; And when a customer question is input, the question is transmitted to the server, a machine learning module searches the database and compares the customer question with a question in a predefined FAQ list, and the Korean sentence of the customer question is converted for each character. It decomposes into phoneme units, embedding each phoneme into a vector, and uses a machine learning model to learn the similarity of two sentences using a sentence similarity model learning algorithm for the two sentences of the customer question and the FAQ question. It includes a sentence similarity model machine learning system that automatically converts the top k FAQs with high similarity to the agent in multiple choice and when one of the FAQs is selected and outputs the answers to the corresponding questions. ,
The machine learning model uses a machine learning engine of bidirectional gated recurrent unit (BIGRU) machine learning for sentence similarity measure of machine learning and a sentence similarity model learning algorithm,
The sentence similarity model is
After pre-processing of tokenization of customer questions, each sequence of tokens is embedded as a random vector through each embedding layer, and the embedding size is set to 256. , An arbitrary vector is calculated by bidirectional gated recurrent unit (BIGRU) machine learning used for sentence similarity measure,
The similarity of the sentence of the customer question and the question-response pair of the FAQ selected from the database, and the similarity of the two sentences is calculated by the sigmoid function using the full connect layer and the euclidian distance,
The similarity of the two sentences (sentence1, sentence2) is between 0 and 1, and the closer the value is to 1, the more similar the two sentences. Using the distance d(X,Y), two points in N-dimensional space X = (x1,x2,x3,...

) And Y = (y1,y2,y3,...

Given ), the Euclidean distance between two points is d(X,Y)

Is calculated.

The machine learning-based Korean customer service assistant assistance system of the present invention is a machine of bidirectional gated recurrent unit (BIGRU) machine learning for sentence similarity measure of natural language processing (NLP) technology of artificial intelligence and machine learning (ML). Using the learning engine and the learning algorithm of the sentence similarity model, the customer service system compares the questions of two sentences in a predefined FAQ question-response pair stored in the server's database for the customer's question. If the highest five FAQs are presented to the agent in multiple choice, and the agent selects one of the matching question-answer pairs among the top five FAQ questions, the machine learning's sentence similarity in the server's database Depending on the model, the counselor's choices are automatically converted into training data, which has the effect of outputting answers to customer questions.

In this study, a customer service representative can effectively shorten the response time to customer questions by storing customer questions in the server database and matching them with predefined FAQs, and simultaneously generate learning data from the actions of agents (employees) to update the system. A new customer service assistant assistance system has been proposed so that it can be done.

In customer service, the question-and-answer problem is simplified to the problem of determining whether two sentences are similar. In the proposed machine learning-based Korean customer service assistant assistance system, a question-answer pair of one of the top five FAQs is selected from a predefined FAQ for customer questions. In other words, the agent simply selects one of the five FAQ questions. Then, the agent's choices are automatically converted into training data, and the answers to the questions are output.

1 is a diagram showing a learning model structure of bidirectional gated recurrent unit (BIGRU) machine learning of each sentence.
2 is a diagram showing the overall service configuration of a Korean customer service assistant assistance system based on machine learning.
3 is a diagram showing accuracy with only baseline data.
4 is a diagram showing the accuracy of automatically generated data (accuracy with only automatically generated data).
5 is a diagram showing accuracy with the first re-training method.
6 is a diagram showing accuracy with the second re-training method.
7 is a diagram showing an accuracy change over repeated re-trainings of the first method.
8 is a diagram showing an accuracy change over repeated re-trainings of the second method.

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

The present invention proposes a new customer service assistant assistance system that helps customer service personnel answer customer questions. The algorithm of the proposed customer service assistant assistance system is as follows.

First, when a customer accesses a specific customer service website, a messenger system, or a social community site such as Facebook and a customer inputs a question to an agent (employee), the question is transmitted to a machine learning-based customer service assistant assistance system. Then, the machine learning-based customer service assistant assistance system compares the received question with the questions in a predefined frequently asked questions (FAQ) list. The customer service assistant assistance system then provides the top five question-answer pairs from the predefined FAQ, and finally the agent selects one question from the five recommendations. After that, the selected question automatically delivers learning data and outputs the corresponding answer.

As a result, additional training data is automatically generated by the selection of the counselor (employee) and used for relearning of the system. Experimental results show that the automatically generated data of the proposed customer service assistant assistance system recursively enhances the accuracy of the system.

This task is structured as follows. Sections 2 and 3 detail the learning model and the proposed methodology, respectively. Section 4 presents experimental results and Section 5 presents conclusions.

II. Learning model

The learning model of the customer service assistant assistance system consists of three processes.

process; i) preprocess, ii) embedding, and iii) bidirectional gated recurrent unit (BIGRU) machine learning

1) The Korean sentence of the customer's question is decomposed into phoneme units for each character, 2) each phoneme is embedding into a vector (embedding size is set to 256), and 3) two sentences are used using bidirectional gated recurrent unit (BIGRU) machine learning. Learn the degree of similarity.

4) In the use of counselors to respond to customers, customer question-response pairs stored in advance in the server database are prepared,

5) Compare each in the form of (customer question-one of the FAQs previously stored in the server's database),

6) Compare the customer question with the question similarity of one selected FAQ, and present the top 5 FAQs with the highest similarity between the two sentences in multiple choice.

7) If there is an answer to the actual customer's question among these, the agent selects one of the 5 FAQs.

8) If the answer to the actual customer's question does not exist, select the'Does not exist' item.

9) When the agent selects an item, it is automatically converted to new learning data and saved in the database immediately.

10) After that, the model is periodically retrained with newly collected training data and updated in the database.

A. Preprocess

Korean'Hangul' consists of 19 initial consonants, 21 vowels, and 27 final consonants.

For example, the letter'한' is composed of (initial consonant'ㅎ') + (vowel'ㅏ') + (final consonant'ㄴ'), and the pronunciation is pronounced like'han'.

As shown in Table I, for each element a token is provided as follows.

Customer Service Assistant Assistance System 700

A server 710 of a customer service system;

A database 720 in which a question-response pair of frequently asked questions (FAQ) is stored when an agent responds to a customer; And

When a question is input as a customer chats, the question is transmitted to the server 710, and the machine learning module searches the database and compares the question-response pair data of a preset number of FAQs to define the customer's question in advance. It compares with the questions in the FAQ list, decomposes the Korean sentence of the customer's question into phoneme units for each character, embedding each phoneme into a vector, and learns the similarity of the two sentences using the BiGRU machine learning model to answer the question. It includes a sentence similarity model (sentence similarity model) machine learning system 730 to output the answer,

)과 Y = (y1,y2,y3,...

)이 주어질 때, 두 점 사이의 유클리드 거리는 d(X,Y)는

계산된다. The machine learning model uses a machine learning engine of bidirectional gated recurrent unit (BIGRU) machine learning for sentence similarity measure of machine learning and a sentence similarity model learning algorithm,
The sentence similarity model is
After pre-processing of tokenization of customer questions, each sequence of tokens is embedded as a random vector through each embedding layer, and the embedding size is set to 256. , An arbitrary vector is calculated by bidirectional gated recurrent unit (BIGRU) machine learning used for sentence similarity measure,
The similarity of the sentence of the customer question and the question-response pair of the FAQ selected from the database, and the similarity of the two sentences is calculated by the sigmoid function using the full connect layer and the euclidian distance,
The similarity of the two sentences (sentence1, sentence2) is between 0 and 1, and the closer the value is to 1, the more similar the two sentences. Using the distance d(X,Y), two points in N-dimensional space X = (x1,x2,x3,...

) And Y = (y1,y2,y3,...

Given ), the Euclidean distance between two points is d(X,Y)

Is calculated.

When a customer question is input to the server 710, the customer service assistant assistance system 700 provides a first consonant, a vowel, and a final consonant for each character in a word divided into sentences/tabs. Disintegrates. Decomposed elements are replaced by corresponding tokens given in Table 1. One question is converted into a sequence of tokens.

The length of the maximum sequence of tokens is fixed at 150. If it is longer than 150, the excess is cut off. The remaining sections whose sequence is shorter than 150 are filled with zeros and tokenized. Consequently, the questions are tokenized as shown in Table 2.

Question "Can I cancel my reservation?"
("Can/May I cancel the reservation?) Decomposition [ㅇ,ㅖ.ㅇ,ㅑ,ㄱ,ㅊ, ㅜ,l,ㅅ,ㅗ,ㄷ,
ㅗ,l,ㄴ,ㅏ,ㅇ,ㅛ,?] Tokenization [12,27,12,22,41,15,36,10,28,100,4,
31,3,20,12,32,131,0,0, ...0]
TABLE II VERTORIZATION OF KOREAN QUESTION

B. Sentence similarity model

After the tokenization preprocessing of the question, each sequence of tokens is embedded as a random vector through each embedding layer. In this study, the embedding size was set to 256. Any vector is computed for each BiGRU. The bidirectional gated recurrent unit (BIGRU) is well known for its sentence similarity measure [9].

The similarity of the sentence of the customer question and the question-response pair of the FAQ selected from the database, and the similarity of the two sentences is calculated by the sigmoid function using the full connect (FC) layer and the euclidian distance. .

)과 Y = (y1,y2,y3,...

)이 주어질 때, 두 점 사이의 유클리드 거리는 d(X,Y)는 다음과 같이 계산된다. The similarity of two sentences (sentence1, sentence2) is between 0 and 1. The closer the value is to 1, the more similar the two sentences are. The Euclidean distance d(X,Y) is used to measure the distance between the query and the training data. Two points in N-dimensional space X = (x1,x2,x3,...

) And Y = (y1,y2,y3,...

Given ), the Euclidean distance between two points is calculated as d(X,Y)

; 및The sigmoid function is a nonlinear function, a binary mode sigmoid function having a value between 0 and 1 where τ1(x) is nonlinear,

; And

로 구분된다. Bipolar mode sigmoid function with a value between -1 and 1 where τ2(x) is nonlinear

It is divided into.

과 같이 표현되며, α는 경사도를 결정하며, α가 클수록 경사가 급하다. The binary mode sigmoid function is

It is expressed as, α determines the slope, and the larger α, the steeper the slope.

, 이진 모드 sigmoid 함수

, Binary mode sigmoid function

, 양극 모드 sigmoid 함수

, Bipolar mode sigmoid function

III. Proposed method

A. Baseline model training

Initially, 60,000 pairs of questions are collected through the Web.

Data collected from various Korean portal sites are modified in a promised format and used for training. There are two sentences file and a label file used for training.

First, in a sentence file, each line includes a pair of sentences separated by tabs. Second, each line in the label file has a value of '1' or '0'. If the sentence pair in the sentence file is similar, the label is '1'; otherwise, the label is '0'. As with the baseline data set, 30000 pairs of sentences and corresponding labels are created. The ratio of the label is '1' to '0' is 1:1.

B. Service-on-the-loop data acquisition

In the proposed customer service assistant assistance system, training data can be automatically generated during customer service. When a customer question is entered, the customer question is transmitted to the customer service assistant assistance system. The question-response pairs of the top five FAQs that match the FAQs stored in the server's database are then displayed to the service staff (agent), and the service staff selects one of them. As a result, new data is written to the database. This procedure can be summarized in 5 steps as shown in FIG. 2 and below.

1) Input question: Customer asks a question.

2) transfer the question to the server

The server transmits the input customer question and FAQ list to a machine learning model using a machine learning engine of machine learning and a sentence similarity model learning algorithm.

3) Recommendations

The sentence similarity learning model displays the top 5 FAQs with the highest sentence similarity corresponding to customer questions to the service staff.

4) Show recommended FAQs

The customer service assistant assistance system displays recommended FAQs to the customer service staff (agent).

5) data acquisition

The customer service staff selects one of the recommended FAQs. As a result, new customer question data is automatically created and stored in the database. Data with FAQs selected by the customer service staff stored in the server database for customer questions is displayed as '1', and data without customer questions in the FAQ (the rest) are displayed as '0'.

In an embodiment, the customer service assistant assistance system 700 presents the top five FAQs having high sentence similarity with respect to the customer question to the service staff in multiple choice, and the service staff selects one of the recommended FAQs, The response to the corresponding customer question is displayed,

C. Re-training

Customer question data that is not in the FAQ selected by the customer service staff is displayed as '0', and the new customer question data is re-constructed using a learning model and stored in the server database to store the machine learning model. Can be used to re-train using. Relearning can be done in two ways.

The first method is to maintain the weights of the machine learning model (baseline model) using the machine learning engine of the machine learning and the sentence similarity model learning algorithm, and start training again with new data.

The second method integrates the basic data of the FAQ (question-response pair) and the new customer question data that is input, and starts learning again by storing the question-response pair in the server database from the initial initial stage. Using this kind of circular structure, it is possible to continuously and automatically improve the accuracy of customer question-and-answer in customer service systems.

Rather than a model trained with general sentences (Naver intellectual data crawling), if the customer question data actually gathered in the customer service assistant assistance system is re-learned with the training data using the proposed method, the sentence of the customer question in the corresponding customer service area Similarity was predicted more accurately.

The most representative advantages or features of the present invention are,

1) When starting a new customer service, there is no need to collect a lot of customer question data, and customer question data stored in the database of the server automatically accumulates during customer service, reducing the response time with a more accurate customer question-response model. The quality of customer service is improved.

2) Customer service staff (counselors) are not replaced, but supported, and the quality of customer service can be leveled upwards regardless of the individual's ability by changing the subject matter to multiple choice (this means that customers' complaints are usually low-educated. Considering that the agent does not understand the question well, it is a different advantage than the existing chatbot service).

III. Experiment result

A. Experimental environment

The Baseline model learned data by web crawling on several Korean portal websites. The number of basic training data is 30,000, and the ratio of labels '1' to '0' is 1:1. As a list of FAQs, 13 FAQs were randomly selected from the Facebook FAQ web page, which is completely different from the portal website for the default FAQ.

Because it is not easy to get real customer questions, customer questions were artificially collected in the database of the server through web crawling by the Facebook customer community. The batch size of the question data was 128, and the embedding size of the vector for the question data was 256, and a sentence similarity model training experiment using machine learning was performed. The execution rate (learning rate) was set to 0.001, and the maximum length of sequence of tokens was set to 150.

We used the Adam optimizer, and binary cross entropy was used as the loss function.

Binary cross entropy uses machine learning's decision tree learning algorithm to convert all data from the root node of the decision tree to the leaf node for classification into a rule. When a child node is created and learned based on a splitting attribute, each of the training data has a value between 0 and 1 for prediction accuracy.

Entropy I is used as a measure to measure the degree of homogeneousness of a data set.

Accuracy refers to how accurate the prediction of a classification model in machine learning in binary classification and multi-class classification is.

The final output of the machine learning model (basic model) using the machine learning engine of the machine learning and the sentence similarity model learning algorithm is a value between 0 and 1, which means similarity. A value of 1 for similarity means that the two sentences are the same. Accuracy is calculated through a pre-defined test set. The test set consists of pairs of sentences labeled '1' that are not used for training. This pair of sentences consists of an entered customer question and a list of FAQs containing the corresponding customer question-response pairs stored in the server database. The correct output for each customer question input means that the correct answer is included in the top k recommendations of the customer service assistant assistance system. All experiments were performed for k = 1, 2 and 5.

B. Effectiveness of the proposed system

This experiment measured the effect of domain specific data on the accuracy of the model. 3 and 4 show the accuracy of only the customer question data automatically generated and the baseline data of the FAQ of the FAQ by the customer service assistant assistance system, respectively. Since the generated customer question data is more domain-specific than the underlying data, the accuracy of the generated customer question data is much higher with smaller data sizes. However, response data for new customer questions is continuously generated and stored in a database of the server, and basic data (question-response pairs) of the FAQ stored in the server database are also re-learned and used essential for customer service.

5 and 6 show the accuracy of the first and second relearning methods, respectively. As shown in the figure, the second method is more effective than the relearning method of the first method. The reason is that in the first method, the weights of the network are in a converged state. On the other hand, the second method cancels the converged state.

7 and 8 illustrate accuracy change over repeated re-trainings by the first and the second re-training method, respectively. As shown, both methods are sufficiently effective when applied repeatedly. This is the most important advantage of the customer service assistant assistance system.

IV. conclusion

The machine learning-based Korean customer service assistant assistance system of the present invention is a machine learning engine of bidirectional gated recurrent unit (BIGRU) machine learning and a sentence similarity model for sentence similarity measure of artificial intelligence natural language processing technology and machine learning. Using the learning algorithm of (sentence similarity model), in the customer service system, when a customer question is entered, the top 5 FAQs with the highest similarity between the two sentences are compared to the questions of the FAQ predefined in the server database in multiple choice. And the agent selects one of the matching question-answer pairs among the top five FAQs, the agent's choices are automatically trained according to the sentence similarity model of machine learning in the server's database. It has the effect of outputting a response to customer questions by converting it into training data.

In this study, a customer service representative saves customer questions (queries) in server data and compares them with predefined FAQ questions, effectively shortens the response time to the customer questions of the matching FAQ, and takes the staff's actions to update the system. A new customer service assistant assistance system was proposed to simultaneously generate learning data from

In customer service, the question-and-answer problem is simplified to the problem of determining whether two sentences are similar. In the proposed machine learning-based Korean customer service assistant assistance system, when a service employee selects a question-answer pair among the top 5 FAQs from a predefined FAQ for customer questions, that is, the agent Just select a question-response pair from the FAQ for one of your questions. Then, the service staff's choices are automatically converted into training data, and responses to customer questions are output.

This study proposed a new machine learning-based customer service assistant assistance system that allows customer service staff to answer questions. The machine learning-based customer service assistant assistance system can automatically accumulate new customer question data in the server database while the customer service staff is servicing, so the question-response pair is re-established and retrained to solve the data shortage problem in various customer service fields Can be avoided. In addition, according to the experimental results, the more customer question data the system collects, the higher the accuracy. This means that the accuracy of the system improves over time by customer question data automatically accumulated in the machine learning-based customer service assistant assistance system.

First of all, the machine learning-based customer service assistant assistance system converts short-choice problems into multiple-choice problems, and in customer service, employees (agents) can make customers more satisfied as well as reduce the time to answer customer questions.

In addition, the customer service assistant assistance system can be applied not only to Korean but also to other languages such as English and French.

Embodiments according to the present invention may be implemented in the form of program instructions that can be executed through various computer means and recorded in a computer-readable recording medium. The computer-readable recording medium may include program instructions, data files, and data structures alone or in combination. Computer-readable recording media include storage, magnetic media such as hard disks, floppy disks and magnetic tapes, optical media such as CD-ROMs and DVDs, and magnetic media such as floptical disks. -A hardware device specially configured to store and execute program instructions in a magneto-optical media, and a storage medium such as ROM, RAM, flash memory, etc. may be included. Examples of program instructions may include those produced by a compiler, as well as high-level language codes that can be executed by a computer using an interpreter as well as machine language codes. The hardware device may be configured to operate as one or more software modules to perform the operation of the present invention.

As described above, the method of the present invention is implemented as a program and can be read using software of a computer, such as a recording medium (CD-ROM, RAM, ROM, memory card, hard disk, magneto-optical disk, storage device, etc.). ) Can be stored.

Although it has been described with reference to the preferred embodiments of the present invention, various modifications or variations of the present invention within the scope not departing from the spirit and scope of the present invention described in the following claims by those of ordinary skill in the relevant technical field It will be understood that this can be done.

700: machine learning-based customer service assistant assistance system
710: server 720: database
730: sentence similarity model

Claims (9)

A server of a machine learning-based customer service assistant assistance system;
A database storing question-response pairs of frequently asked questions (FAQs) of customers; And
When a customer question is entered, the question is transmitted to the server, the machine learning module searches the database and compares the customer question with the question in a predefined FAQ list, and the Korean sentence of the customer question is phoneme for each character. It decomposes into units, embedding each phoneme into a vector, and uses a machine learning model to learn the similarity of the two sentences using a sentence similarity model learning algorithm for the two sentences of the customer question and the FAQ question. It includes a sentence similarity model machine learning system that automatically converts the top k FAQs with high sentence similarity to the counselor in multiple choice and when one of the FAQs is selected, automatically converts them into training data and outputs the answers to the corresponding questions.
The machine learning model uses a machine learning engine of bidirectional gated recurrent unit (BIGRU) machine learning for sentence similarity measure of machine learning and a sentence similarity model learning algorithm,
The sentence similarity model is
After pre-processing of tokenization of customer questions, each sequence of tokens is embedded as a random vector through each embedding layer, and the embedding size is set to 256. , An arbitrary vector is calculated by bidirectional gated recurrent unit (BIGRU) machine learning used for sentence similarity measure,
The similarity of the sentence of the customer question and the question-response pair of the FAQ selected from the database, and the similarity of the two sentences is calculated by the sigmoid function using the full connect layer and the euclidian distance,
The similarity of the two sentences (sentence1, sentence2) is between 0 and 1, and the closer the value is to 1, the more similar the two sentences. Using the distance d(X,Y), two points in N-dimensional space X = (x1,x2,x3,...

) And Y = (y1,y2,y3,...

Given ), the Euclidean distance between two points is d(X,Y)

Computed, machine learning-based Korean customer service assistant assistance system. delete The method of claim 1,
When a question is entered into the customer service assistant assistance system, the question is divided into initial consonants, vowels, and final consonants for each letter in sentences/tabs, and the decomposed elements are replaced by corresponding tokens. , One question is converted to a sequence of tokens,
The maximum sequence length of the token is fixed at 150, if it is longer than 150, the excess part is truncated, and the remaining sections shorter than 150 are filled with zeros and tokenized, a machine learning-based Korean customer service assistant assistant system. delete The method of claim 1,
The sentence similarity model machine learning system
1) The Korean sentence of the customer's question is decomposed into phoneme units for each character, 2) each phoneme is embedding into a vector (embedding size is set to 256), and 3) two sentences are used using bidirectional gated recurrent unit (BIGRU) machine learning. Learning the similarity of (sentence 1, sentence2),
4) A customer question-response pair is prepared for use in the customer response of the counselor,
5) The customer question is compared with one of the FAQs previously stored in the server database, and 6) the customer question is compared with the question similarity of the selected one, and the top 5 FAQs with the highest similarity between the two sentences are presented to the agent. Presented in multiple choice
7) If there is an answer to the actual customer's question among them, the agent selects one of the 5 FAQs. 8) If the answer to the actual customer's question does not exist, the agent selects the'Does not exist' item.
9) When the agent selects an item, it is automatically converted into new learning data and stored in the database.
10) A machine learning-based Korean customer service assistant assistance system that periodically retrains the newly collected training data by the machine learning model and updates the database. The method of claim 1,
Customer questions are artificially collected by the customer community through web crawling and stored in the database of the server. The batch size of the question data is 128, and the embedding size of the vector for the question data is A sentence similarity model training experiment using machine learning was performed at 256, the execution speed (learning rate) was set to 0.001, the maximum length of sequence of tokens was set to 150, and
Adam optimizer was used, binary cross entropy was used as the loss function,
The binary cross entropy rules all data from a root node of a decision tree for classification to a leaf node using a decision tree learning algorithm of machine learning. When creating and learning a child node based on the splitting attribute, each of the training data has a value between 0 and 1 for prediction accuracy,
Entropy I is used as a measure of the homogeneous degree of a data set,
Accuracy refers to how accurate the prediction of the classification model of machine learning in binary classification and multi-class classification is,
The final output of the machine learning model using the machine learning engine of the machine learning and the sentence similarity model learning algorithm is a value between 0 and 1, which means similarity, and the value of similarity 1 is the same for both sentences. Means,
Accuracy is calculated through a predefined test set, which consists of pairs of sentences labeled '1' that are not used for training, which pairs of sentences are input to the customer question and the corresponding customer. A machine learning-based Korean customer service assistant assistance system consisting of a list of FAQs including question-response pairs. The method of claim 1,
The customer service assistant assistance system displays recommended FAQs having k question-response pairs to the service staff in response to the entered customer question,
Data with FAQs selected by the customer service staff stored in the server database for customer questions is displayed as '1', and data without questions in the FAQ is displayed as '0'.
A machine learning-based Korean customer service assistant assistance system in which service staff selects one of the recommended FAQs and answers to corresponding customer questions are output. The method of claim 7,
The customer service assistant assistance system presents the top five FAQs with high sentence similarity to the customer question in multiple choice, and the service employee selects one of the recommended FAQs, and responds to the corresponding customer question. A machine learning-based Korean customer service assistant assistance system that outputs the response. The method of claim 7,
Customer question data that is not in the FAQ selected by the customer service staff is displayed as '0', and the new customer question data is used for re-training by reconstructing question-response pairs using a learning model. And
Relearning is carried out in two ways,
The first method is to maintain the weights of the machine learning model (baseline model) using the machine learning engine of the machine learning and the sentence similarity model learning algorithm and start training again with new data,
The second method resumes learning by integrating the basic data of the FAQ (question-response pair) with the new customer question data, and storing and accumulating the question-response pairs from the initial initial stage again.
Machine learning-based Korean customer service assistant assistance system that continuously and automatically improves the accuracy of the system's customer question-and-answer.

Images