KR102103273B1 - Method and apprartus for chatbots in customer service analyzing hierarchical user expression and generating responses - Google Patents

Abstract

The present invention relates to a method in which a chatbot device in which a conversation system is built in responds to a system when a user device utters a query through a communication system.
In the process of the present invention, (a) an object suitable for the purpose of conversation is extracted from the input text, (b) a feature vector necessary for machine learning is generated using the extracted object, and (c) the generated feature vector is used. Query understanding is performed by extracting hierarchically constructed query intention, conversation topics, and query status. In addition, (d) it creates answers by finding answers that match the hierarchical elements of query intention, conversation topic, and query situation.
With this configuration, you can understand user expressions hierarchically and generate accurate answers.

Description

{METHOD AND APPRARTUS FOR CHATBOTS IN CUSTOMER SERVICE ANALYZING HIERARCHICAL USER EXPRESSION AND GENERATING RESPONSES}

The present invention relates to a machine learning-based language analysis and processing technology, and in particular, to a chatbot technology that performs interactive customer consultation.

The technology of recognizing and responding to human voice by a machine has been applied to artificial intelligence systems in various fields in real life. These systems analyze the user's input sentence based on text mining technology to grasp the meaning and generate and output a response according to the intention. With the development of such information processing technology and machine learning technology, the use of a dialogue system or chatbot that can automate the answer to customer inquiries on behalf of the counselor is being tried here and there.

The current chatbot system can be largely divided into a destination-oriented conversation technology that creates a conversation for providing a specific service to a user, and an open domain conversation technology that is mainly used for providing a purposeless everyday conversation. . The destination-oriented conversation technology is to provide a most appropriate answer by analyzing a user's intention and query content, given a predetermined answer. Because it is a matter of choosing one of the limited answers, a relatively accurate answer is possible. However, it takes a lot of effort to prepare and refine the answers in advance. Open domain conversation technology creates answers to users' expressions by using machine learning or rules. This method has the advantage that it does not need to build a separate knowledge, and it is possible to implement a dialogue system only by securing a dialogue example. However, there is a disadvantage that the accuracy of the answer is poor and many unintended conversations occur.

The chatbot system for customer consultation on behalf of the counselor has a feature that the domain to be used is determined, and since the accuracy of the answer is important, the conversation technology toward the destination is mainly used. The chatbot system is required to have a technique for generating a system response suitable for a user's text input. In the past, a model was developed to extract the intention of a query by expressing the mapping relationship between questions and answers as a rule or learning by machine learning by using a pre-built data set composed of question-answers. When a user's question comes in, the pre-made model is used to classify what is the intention to best express the user's question. Then, it serves to provide answers that are categorized and intended.

Such a conventional intention classification model is included in a predefined question-answer data set or generates answers to similar questions, so it is possible to generate automated answers on behalf of the counselor, but it is complicated and has many questions. Definitely showed a problem when handling it.

The problem was when there were many intentions to classify. The prior art created problems in the accuracy of the answers. For example, if there are 10 intentions to be classified, relatively accurate classification of intentions may be possible because there are 10 candidates for answers to the questions. However, if the number of intentions to be classified increases to 1000, the target candidates increase, so it is necessary to find the correct intention among more candidates, and the accuracy is greatly reduced.

In addition, in the conventional method, when the content of the question is complicated or the question contains various information, it is difficult to find a suitable mapping between the question and the answer, so the accuracy of the answer is deteriorated.

The inventors of the present invention came to complete the present invention after participating in a national R & D project while researching and trying for a long time to improve the prior art.

The purpose of the present invention for solving the above problems is to understand the user's complex and various expressions more accurately in a conversation system, and to find a user's desired answer with high accuracy among many types of answers, and such a method. It is to provide an implemented dialogue system.

On the other hand, other objects not specified in the present invention will be additionally considered within a scope that can be easily inferred from the following detailed description and its effects.

In the first aspect of the present invention, when a user device utters a query through a communication system, the chatbot device in which the conversation system is built in responds to the system, so that the hierarchical user expression performed by the chatbot device is understood and answered. It's about how to implement the interactive consultation chatbot you create.

The method of the present invention:

(a) extract the object suitable for the purpose of the conversation from the entered text,

(b) Using the extracted object, a feature vector necessary for machine learning is generated,

(c) Using the generated feature vector, hierarchically constructed query intention, dialogue subject, and query status are extracted to perform query understanding,

(d) After performing a step of finding an answer that matches the hierarchical elements of the query intention, conversation topic, and query situation, and generating an answer,

(e) combining the generated answer with a voice and transmitting a response to the user device.

In addition, in any preferred embodiment of the present invention, the chatbot device may be built in a hospital system.

In addition, in a preferred embodiment of the present invention, step (a) is an object recognition step, and the objects extracted in this step are N (N is an integer of 1 or more) expressed as (object type, representative value). Can be extracted with

In addition, the present invention relates to an interactive consultation chatbot device in which a chatbot device in which a conversation system is built performs a system response when the user device utters a query through a communication system. This device:

Text preprocessor that performs pre-processing for input text

An object extraction unit for extracting an object capable of analyzing text preprocessed by the text preprocessor,

A feature vector generator that generates a numerical vector using the analyzed objects and predefined words

A query intention understanding unit, a conversation topic understanding unit, and a query situation understanding unit configured to understand the generated feature vector into three categories: query intention, conversation topic, and query situation. The conversation subject understanding unit and the query situation understanding unit each include a plurality of hierarchical classifiers.

Through the above-described problem solving means, the present invention has an advantage in that accurate response data can be extracted from the user's question in consideration of the user's intention as well as the subject and situation of the desired answer.

In addition, the present invention can improve the accuracy of answers by hierarchically classifying various and complicated questions of users by using a hierarchical classification method.

If it is utilized in the chatbot counseling system of the present invention that provides an answer to a user's counseling question, inquiry, conversation subject, and conversation situation, effective machine counseling can be implemented in various fields. For example, a lifestyle-based interactive medical questionnaire for chronic patients can be effectively implemented.

On the other hand, even if the effects are not explicitly mentioned herein, it is noted that the effects described in the following specification expected by the technical features of the present invention and the potential effects thereof are handled as described in the specification of the present invention.

1 is a diagram conceptually showing the overall system configuration of the present invention.
2 shows a schematic system and process configuration of a conversation system 100 according to any preferred embodiment of the present invention.
3 shows a configuration of a machine learning procedure for hierarchical classification according to an embodiment of the present invention.
4 shows an example of a conversation subject hierarchy in a system scenario according to the present invention.
5 shows an example of a query intention hierarchy in a system scenario according to the present invention.
6 shows an example of a query context hierarchy in a system scenario according to the present invention.
7 is a diagram showing an example of a label structure of answer data according to the present invention.
8 schematically shows a configuration of a procedure for calculating the similarity between a question and an answer candidate according to an embodiment of the present invention.
9 schematically shows the entire process of an implementation method of an interactive consultation chatbot according to an embodiment of the present invention.
※ The accompanying drawings indicate that they are exemplified by reference for understanding the technical idea of the present invention, and the scope of the present invention is not limited thereby.

Hereinafter, a configuration of the present invention guided by various embodiments of the present invention and effects resulting from the configuration will be described with reference to the drawings. In the description of the present invention, when it is determined that the subject matter of the present invention may unnecessarily obscure the subject matter of the present invention, the detailed description thereof will be omitted.

1 conceptually shows the system configuration of the present invention. When the user device 1 utters a query through a communication system, the chatbot device 10 built in the customer consultation center delivers the optimized response to the query to the user device 1 in real time.

The user device 1 means a terminal that can access the customer service center through a communication system. For example, various types of telephones performing wired / wireless communication.

The chatbot device 10 is an electronic device composed of one or more hardware and software installed in a customer consultation center of a company (institution). The query of the user device 1 is transmitted to the chatbot device 10 by voice, and the chatbot device 10 converts the extracted response to voice according to the process of the present invention and transmits it to the user device 1 through a speaker. .

The chatbot device 10 includes a conversation system 100 and a memory 200, particularly in implementing the method of the present invention. The conversation system 100 is a hardware / software device that performs the language processing of the present invention, and includes a control device that controls the driving of the software and the input / output and functions of the device. The memory 200 serves to store data in a pre-built knowledge base and to store various databases related to machine learning. The chatbot device 10 includes the conversation system 100 and the memory 200 equipment, and a detailed configuration example is illustrated in FIG. 2.

2 shows a schematic system and process configuration of a conversation system 100 according to any preferred embodiment of the present invention. A method of understanding hierarchical user expressions and generating answers according to the present invention is illustrated.

The user voice expressed through the user device is converted into input text and enters the conversation system. The text pre-processor 101 performs pre-processing, including checking and correcting typos in the input text, line break processing, and special symbol processing.

The object extraction unit 103 extracts the pre-processed text in the form of information that can be analyzed. That is, the entity is extracted from the input text according to the purpose of the conversation. The object recognition performed by the object extraction unit 103 uses resources stored in a language DB such as a synonym dictionary 210, a semantic feature dictionary 211, and a pattern dictionary 213. Morphological analysis of the input text of the user is performed, and the synonym dictionary 210 is used to divide it into semantic units, and refined through natural language processing such as stemming and headword conversion. The first analyzed word is whether a word matching the synonym dictionary 210 exists. The matched synonyms are analyzed by using the semantic dictionary 211. Proper nouns or numeric data that are not included in the dictionary are recognized through pattern analysis or a separate dictionary of individual names.

The format of the data output through the object extraction unit 103 is the output of the object extraction unit 103 in N arrays represented by the object type and the representative value if the number of recognizable objects is N (N is an integer of 1 or more). This is fixed. It looks like this:

(Object type 1, Representative value 1), (Object type 2, Representative value 2), ... , (Object type N, representative value N)

Let's look at this meaning through an example input statement. Assume that the following sentence is entered through a user query.

"Tomorrow morning After 11:00  I want to know the time of the KTX train from Seoul Station to Daejeon Station. ”

Then, the object extraction unit 103 of the present invention can be analyzed as follows.

(Date, Tomorrow), (Time, 11:00 AM), (Departure, Seoul Station), (Destination, Daejeon Station), (Transportation, KTX), (Key Words, Time), (Key Words, Inquiries)

The feature vector generator 110 generates a numerical vector using the analyzed objects and predefined words. For the method of generating a digitized vector, for example, known methods such as Bag of Word, TF / IDF, and word embedding may be used.

The generated feature vector is used for classification based on machine learning. In the present invention, classification is performed in three aspects: query intention, conversation topic, and query situation. In addition, labels used in machine learning classification are organized in a hierarchical structure. The intention of inquiry and the topic of conversation select the intention and subject selected from the first-class classifier first, and if the lower classifier exists under each selected subject, classify the lower class.

The query intention understanding unit 120 may include a plurality of classifiers, such as the first-class classifier 121, the second-class classifier 122, and the third-class classifier 123, as shown, in performing classification of the classifier. The intention classification learning models corresponding to each class are used (221, 222, 223).

The conversation topic understanding unit 130 may include a plurality of classifiers, such as the first-class classifier 131, the second-class classifier 132, and the third-class classifier 133, as illustrated, hierarchies in classifier performance Each subject class learning model (231, 232, 233) is used.

The query situation understanding unit 120 performs hierarchical classification only for the second and third layers.

The user expression understanding unit 150 performs query understanding by hierarchically extracting a query intention, a conversation topic, and a query situation of the input sentence hierarchically according to the three classifications. Then, the answer candidate generation unit 160 generates an answer candidate using the answer knowledge 260, and then the system response generation unit 170 generates the final answer and combines it with the voice to output it.

Meanwhile, a machine learning procedure for hierarchical classification is illustrated in FIG. 3.

The question data 50 and the hierarchical labels 51, 53, and 55 are each composed of a pair. As described above, the question data 50 is expressed as a numerical vector expressing the question through text pre-processing, object extraction, and feature vector generation procedures, and is transmitted to the automatic classification learner 180.

The subject label (51), the intent label (53), and the context label (55) are composed of hierarchical structures with labels related to conversation topics, question intentions, and question situations, respectively. 180).

That is, the vector expressing the question and the vector expressing the label are used as learning data of the automatic classification learner 180. In addition, the automatic classification learner 180 classifies the first, second, and third layers respectively according to the label hierarchy, and finally generates classification learning models 281, 282, and 283 for each layer. It is utilized in the structure shown in.

Then, let's take a look at the meaning of the hierarchical label for dialogue topic, query intention, and query situation as a hierarchical structure used for classifying machine learning through an illustrated scenario. 4 to 6 are scenarios for inquiring about chronic diseases in a hospital customer center system in which the chatbot device of the present invention is built.

As shown in the dialogue topic hierarchy of FIG. 4, in the case of a dialogue topic, it serves to find content on a topic included in a question. The first-tier classification is organized according to the contents of the dialogues such as <Health Management>, <Diseases and Symptoms>, and <Hospital Care>, and the first-tier classification can be divided into subdivided classifications as shown according to the details.

5 relates to a query intention hierarchy. The intention of inquiry mainly indicates the purpose of the question rather than the content of the question. <Inquiry>, <Reservation>, and <Inquiry> are illustrated. <Contact> is a one-tier classification that asks for a specific item, and two-tier classification consisting of <method inquiry>, <result inquiry>, <concept inquiry>, <prognosis inquiry>, etc., <prevention inquiry> in the case of <method inquiry> Method>, <Corrective Action>, etc., can be layered into three-layer classification. Whether it is for a hospital reservation or a simple information inquiry, it is classified into <reservation> and <information inquiry> into one-tier classification, respectively.

6 relates to the query situation hierarchy. The inquiry situation serves to find the situation in which the questioner is facing. Unlike the other two classifications, all elements constituting the query situation are extracted without performing the first-class classification. In the example, the 1st level elements that express the query status are <Subject Inquiry>, <Patient Status>, <Gender>, and <Age Group>. For the 1st level component, it is not classified so that all items can be selected. Classification.

Now, let's understand the model that generates answer data using the hierarchical label of this hierarchy through a more specific scenario. It is assumed that such a query is input from the user device.

“My grandmother was hospitalized with diabetes last month and was discharged after a simple operation. It is okay to drink water, but if you take medicine or eat, you will vomit immediately. What should I do?"

In response to the above-mentioned question as described above, the label for the intention of inquiry, the topic of conversation, and the situation of inquiry may be hierarchically determined. Query analysis according to the present invention will be made as follows.

Topic:

Tier 1: disease and symptoms,

Layer 2: chronic disease,

Layer 3: Diabetes

Inquiry:

Tier 1: Inquiries,

Layer 2: Method inquiry,

Layer 3: how to deal

Query Status:

1st layer: Inquiries,

Layer 2: Family

Tier 1: Patients,

Layer 2: Chronic disease

Tier 1: Gender,

Layer 2: Female

Tier 1: Age Group,

Tier 2: 60+

The candidate candidate knowledge is assigned a label to a previously prepared answer based on the above defined query intention, conversation topic, and question status. The label structure of the answer data is shown in, for example, FIG. 7.

The extraction of answer candidates to fit a user's question can be followed by a procedure of extracting candidates with a score above a certain level as candidates after scoring and sorting the degree of matching with three items (topic, intention, situation) and hierarchy. .

The procedure for calculating the similarity between question and answer candidates is shown in FIG. 8. After classifying the label (51, 53, 55) for the conversation topic, query intention, and query situation from the question data, the conversation topic, query intention, and query situation label (510, 520, 530) defined in the answer Calculate matching goodness of fit. The appropriateness between the question and the answer is calculated by summing each matching suitability, and P answers with high suitability (P is an integer of 1 or more) can be presented as a candidate for answering the question.

FIG. 9 shows the configuration of the entire process of the method for implementing the interactive consultation chatbot of the present invention as an example, and shows the structure of the present invention once again.

The user will query something through a user device such as a telephone to the consultation system in which the chatbot device of the present invention is built (S100).

Then, the object recognition step is performed. Text pre-processing is performed through a text pre-processor, and the object extraction unit 103 extracts a word suitable for a conversation purpose from the text (S110).

Next, feature vectors are extracted. That is, the feature vector generation unit generates features necessary for machine learning (S120).

Three elements, that is, a query intention understanding unit, a conversation subject understanding unit, and a query situation understanding unit, perform logic to understand a user expression (query) by a predefined hierarchical casting (S130).

Then, an answer is generated (S140), and as described above, the hierarchical structure of the three elements of the candidate candidate and the hierarchical structure of the three elements of the question data are matched, and the matching suitability is calculated to select the answer with the high score as a candidate Extract (S140). Then, it responds to the user device in combination with voice (S150).

For reference, an implementation method of an interactive counseling chatbot according to an embodiment of the present invention may be implemented in a form of program instructions that can be executed through various computer means and recorded in a computer-readable medium. The computer-readable medium may include program instructions, data files, data structures, or the like alone or in combination. The program instructions recorded on the medium may be specially designed and configured for the present invention, or may be known and available to those skilled in computer software.

Examples of computer-readable media include magnetic media such as hard disks, floppy disks and magnetic tapes, optical recording media such as CD-ROMs, DVDs, magnetic-optical media such as floptical disks, and ROM, RAM, Hardware devices specifically configured to store and execute program instructions, such as flash memory, may be included. Examples of program instructions include high-level language codes that can be executed by a computer using an interpreter, as well as machine codes such as those produced by a compiler. The above-described hardware device may be configured to operate as one or more software modules to perform the operation of the present invention, and vice versa.

The scope of protection of the present invention is not limited to the description and expression of the embodiments explicitly described above. In addition, it is pointed out once again that the scope of protection of the present invention may not be limited by obvious changes or substitutions in the technical field to which the present invention pertains.

Claims (4)

When a user device connects to a customer center conversation system of a hospital that supports interactive medical consultation through a communication system and speaks a query, the chatbot device built in the customer center conversation system responds to the system, so that the chatbot device :
(a) extracting an object suitable for the purpose of the conversation from the input text, but outputting and analyzing the data in an array of the form (object type, representative value of the object name),
(b) generating feature vectors for machine learning using extracted and analyzed objects and predefined words,
(c) For the feature vectors used for machine learning-based classification, the query intention understanding unit including a plurality of hierarchical classifiers classifies the query intention using an intention classification learning model corresponding to each layer. Performing hierarchically, the conversation subject understanding unit including a plurality of hierarchical classifiers hierarchically classifies the conversation subject using a subject classification learning model corresponding to each layer, and comprising a plurality of hierarchical classifiers The query situation understanding unit classifies the query situation hierarchically by using the situation classification learning module corresponding to each class, but the query situation is not classified by the first layer so that all items can be selected. The step of performing classification from above step,
(d) After performing a step of finding an answer that matches the hierarchical elements of the query intention, conversation topic, and query situation, and generating an answer,
(e) a method of implementing an interactive counseling chatbot, characterized by hierarchically understanding a user expression and generating an answer by including a step of transmitting a response to the user device by combining the generated answer with a voice.
delete delete delete

Images