KR20230000044A - Apparatus for automatically providing response and a method for controlling the same - Google Patents

Abstract

The present invention relates to a device for constructing an emotional dictionary of new words and emoticons, wherein the device includes: an emotional dictionary construction module; and a database unit which is connected to the emotional dictionary construction module and in which collected social data to be worked is stored. The emotional dictionary construction module is configured to partition the social data to be worked into sentences to create partitioned sentences, extract new words and emoticons from the partitioned sentences, determine, if at least one of the extracted new words and emoticons is the same as a determined new word or a determined emoticon, another new word or emoticon other than the same new word or the same emoticon as the determined new word or emoticon among the extracted new words and emoticons as a subordinate new word or subordinate emoticon, extract the partitioned sentences as emotional sentences, calculate the polarity strength of the emotional sentences using the total number of determined new words and the determined emoticons present in the emotional sentences and the polarity of the determined new words and the determined emoticons, and construct a new word and emoticon database by storing the subordinate new word, the subordinate emoticon, and the polarity strength in the database unit. According to the present invention, a response generation module can generate responses by considering the emotions of conversation participants.

Description

Automatic response providing device and control method thereof

The present invention relates to an automatic response providing device and a control method thereof, and more particularly, to determine the type of mood of a current conversation by using a coined word, emoticon, and sentimental word included in an input sentence, and to determine the type of the determined mood. An automatic response providing device for generating and outputting a response in consideration of the present invention and a control method thereof.

Due to the development of artificial intelligence (AI) technology, machine learning technology such as deep learning is developing day by day.

Accordingly, chatterbots, which are interactive messengers that answer questions, ask necessary questions, and conduct conversations with counterparts on behalf of humans, have been developed and used.

Due to the emergence of these chatbots, companies can save a lot of labor costs and provide the service to customers regardless of working hours.

These chatbots are divided into rule-based chatbots and machine learning-based chatbots according to their operation method, and can also be classified into one-time question-answering chatbots and continuous conversational chatbots. According to the method of generating answers, chatbots are further classified into search model chatbots and generative model chatbots.

A rule-based chatbot is a chatbot that operates according to predefined rules, and includes interpretation rules for interpreting user input, response rules for responding to inputs, and response rules.

The development of a rule-based chatbot does not require a large amount of data, is relatively easy to implement, and high-quality conversation service is possible if the rules are well defined. However, these rule-based chatbots have the disadvantage of requiring a lot of time and manpower to define rules for smooth conversation.

Machine learning-based chatbots can be implemented with less time and manpower compared to rule-based chatbots.

Since the machine learning-based chatbot uses natural language processing technology, it decomposes the natural language using various linguistic resources in the sentence input by the user, and then selects a predefined answer or creates a new answer using the decomposed natural language. .

However, since these conventional chatbots do not have a technique for determining emotions with users, they are used only functionally without emotional exchanges with users.

Therefore, since the conventional chatbot cannot grasp the user's emotional state, direct emotional exchange with the user is impossible.

In addition, since the user's emotional state is not taken into account, the accuracy of analysis for analyzing the user's utterance intention is also greatly reduced. Therefore, the user is not provided with a satisfactory service through the chatbot, and thus the service satisfaction level is lowered.

Republic of Korea Patent Registration No. 10-1894060 (Publication date: August 31, 2018, title of invention: Intelligent advertisement providing server using user analysis-based chatbot) Republic of Korea Patent Registration No. 10-1945297 (Public date: February 07, 2019, title of invention: production and provision server of messenger-linked chatbot)

The problem to be solved by the present invention is to increase the accuracy of the response by generating and outputting a response in consideration of the type of mood of the current conversation.

An automatic response providing apparatus according to one feature of the present invention for solving the above problem separates a current input sentence into word units and morpheme units, extracts a final keyword for the current input sentence using the separated morphemes, and An intent word is extracted from the final keyword, the intent type for the final keyword extracted as the intent word is determined, the number of intent types for the determined intent type is calculated, and the intent type having the largest number is selected as the current input sentence. A control module that determines the type of intention, extracts neologisms, emoticons, and sentimental words from the current input sentence connected to the control module and separated into word units and morpheme units from the control module, and extracted neologisms, emoticons, and sentimental words, respectively An emotion type extraction module that determines the emotion type of the current sentence by determining the emotion type for the current sentence and is connected to the emotion type extraction module, and the emotion type of the current sentence input from the emotion type extraction module and the emotion type of the immediately previous sentence are the same. After calculating different current sentence sentiment values according to the comparison result, the previous sentence sentiment value is reduced by the set size, and the current conversation is calculated by summing up the sentence sentiment values for each emotion type. and a mood type determination module for determining the mood type for .

The control module may extract a preliminary keyword for the current input sentence by extracting the subject, object, and predicate of the current input sentence using the separated morphemes, and similar words and synonyms stored corresponding to representative words. and if a preliminary keyword corresponding to a synonym or similar word exists among the preliminary keywords by comparing the preliminary keyword with a synonym database, a representative word corresponding to the corresponding synonym or similar word may be determined as a final keyword of the corresponding preliminary keyword, and the preliminary keyword Among them, preliminary keywords that do not correspond to synonyms or similar words may be determined as final keywords.

The emotion type extraction module compares each word with a new word and emoticon emotion dictionary database in which emotion types corresponding to each new word and emoticon are stored, and extracts the newly coined word and emoticon included in each word and extracts the emotion type. It is possible to determine the emotion type for each newly coined word and each emoticon, and compare each morpheme with a sentiment word dictionary database in which emotion words and emotion types corresponding to each emotion word are stored, and extract the emotion word from the morpheme. It is possible to determine the emotion type for each emotional word, calculate the total number of emotions for each emotion type from the emotion types for new words and emoticons and the emotion types for sentiment words, and select the emotion type with the largest number as the emotion type for the current sentence. can be judged by

The emotion type extraction module may remove stopwords included in each word before comparing the coined word and the emoticon emotion dictionary database.

The mood type determining module may determine, as the mood type for the current conversation, an emotion type corresponding to a sentence emotion value having the largest magnitude among total sentence emotion values equal to or greater than a predetermined value among sentence emotion values for each emotion type.

The mood type determination module may determine the mood type for the current conversation as stable when all sentence emotion values for each emotion type have values less than a predetermined value.

The automatic response providing apparatus according to the above feature may further include a response generation module connected to the control module, wherein the control module additionally receives a conversation category, the conversation category, the intention type, and the current input sentence. may be output to the response generation module, and the response generation module may generate a response to the current input sentence by machine learning using the inputted dialogue category, the intention type, the current input sentence, and the mood type. can

The emotion type may be one of joy, sadness, trust, anger, fear, expectation, surprise, and displeasure.

A control method of an automatic response providing apparatus according to another aspect of the present invention includes the step of separating, by a control module, a current input sentence into word units and morpheme units, wherein the control module uses the separated morphemes to provide a final keyword for the current input sentence. extracting, by the control module, extracting an intent word from the extracted final keyword and determining an intent type for the final keyword extracted as the intended word; Calculating the number and determining the intention type having the largest number as the intention type of the current input sentence, the emotion type extraction module receives the current input sentence divided into word units and morpheme units from the control module, and creates new words extracting emoticons and emotional words, the emotion type extraction module determining the emotion types of the current sentence by determining the emotion types for each of the extracted neologisms, emoticons, and emotional words; the emotion type determination module extracting the emotion types Comparing whether the emotion type of the current sentence and the emotion type of the immediately previous sentence input from a module are the same, wherein the emotion type determination module determines different emotion values of the current sentence according to a comparison result between the emotion type of the current sentence and the emotion type of the immediately previous sentence. calculating, by the mood type determining module, calculating the current sentence emotion value and then reducing the previous sentence emotion value by a set size; and the mood type determining module calculates the sentence for each emotion type calculated by summing the emotion values of the sentence for each emotion type and determining a mood type for the current conversation by using the magnitude of the emotion value.

In the step of extracting the final keyword, the control module extracts a preliminary keyword for the current input sentence by extracting the subject, object, and predicate of the current input sentence using separated morphemes, and the control module extracts a similar word and a predicate. comparing the preliminary keyword with a synonym database in which synonyms are stored corresponding to representative words, and determining whether a preliminary keyword corresponding to a similar word or synonym exists among the preliminary keywords, wherein the control module selects among the preliminary keywords If a preliminary keyword corresponding to a similar word or synonym exists, determining a representative word corresponding to the corresponding synonym or synonym as a final keyword of the corresponding preliminary keyword, and the control module, among the preliminary keywords, a preliminary keyword that does not correspond to the synonym or similar word It may include determining as the final keyword.

The step of extracting the newly coined words, emoticons, and emotional words may include comparing each word with a new coined word and emoticon emotion dictionary database storing new words, emoticons, and emotion types corresponding to each newly coined word and emoticon, by the emotion type extraction module; The emotion type extraction module extracts the newly coined word or emoticon included in each word if there is a new word or emoticon stored in the new word and emoticon emotion dictionary database in each word, and the emotion type extraction module performs the step of extracting the newly coined word and emoticon emotion and determining the type of emotion for each newly coined word and each emoticon extracted from the dictionary database.

The step of extracting the newly coined word, emoticon, and emotional word may further include removing stopwords included in each word before comparing each word with a new coined word and an emoticon emotion dictionary database.

The step of determining the emotion type of the current sentence may include comparing, by the emotion type extraction module, each morpheme with a sentiment word dictionary database in which emotional words and emotion types corresponding to each emotion word are stored; If there is a sentiment word stored in the sentiment word dictionary data, extracting the sentiment word included in the morpheme. Determining, by the emotion type extraction module, the emotion type for the emotional word extracted from the sentiment word dictionary database, wherein the emotion type extraction module is the total number of emotions for each emotion type among the emotion types for newly coined words and emoticons and the emotion types for emotional words. and determining, by the emotion type extraction module, an emotion type having the largest number among the total number of calculated emotion types as the emotion type of the current sentence.

The step of determining as the mood type for the current conversation is determining whether a total sentence emotional value of a predetermined value or more exists among sentence emotional values for each emotion type, and if a total sentence emotional value of a predetermined value or more exists among the sentence emotional values for each emotion type, Determining the emotion type corresponding to the sentence emotion value having the largest magnitude among the total sentence emotion values for each emotion type having a predetermined value or more as the mood type for the current conversation, and the total sentence emotion value of each emotion type having a total sentence emotion value greater than or equal to the predetermined value exists If not, determining the mood type for the current conversation as stable.

The control method of the automatic response providing device according to the above feature may further include outputting, by the control module, the conversation category, the intention type, and the current input sentence to a response generating module.

The emotion type may be one of joy, sadness, trust, anger, fear, expectation, surprise, and displeasure.

According to this feature, since not only the current input sentence, but also the type of intention for the current input sentence and the type of mood for the current conversation are input to the response generating module, the response generating module matches the mood for the current input sentence and the intention of the conversation participant. response can be generated.

Also, since the mood type for the current conversation is determined using the emotion type for the current input sentence, the response generation module may generate a response in consideration of the conversation participant's emotion.

At this time, since the emotion type is finally determined using not only emotional words included in the current input sentence but also neologisms and emoticons, the accuracy of the determined emotion type can be greatly improved.

1 is a schematic block diagram of an automatic response providing system according to an embodiment of the present invention.
Figure 2a is a schematic block diagram of the user terminal shown in Figure 1;
FIG. 2B is a schematic block diagram of the automatic response providing device and database unit shown in FIG. 1 .
3 is an operation flowchart of a user terminal of an automatic response providing system according to an embodiment of the present invention.
4 to 6 are operational flowcharts of the automatic response providing device of the automatic response providing system according to an embodiment of the present invention, FIG. 4 is an operational flowchart of a control module, and FIG. 5 is an operational flowchart of an emotion type extraction module. 6 is an operation flowchart of the mood type determination module.

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings. In describing the present invention, if it is determined that adding a detailed description of a technology or configuration already known in the related field may obscure the gist of the present invention, some of them will be omitted from the detailed description. In addition, the terms used in this specification are terms used to properly express the embodiments of the present invention, which may vary depending on people or customs related to the field. Therefore, definitions of these terms will have to be made based on the content throughout this specification.

The terminology used herein is intended only to refer to specific embodiments and is not intended to limit the present invention. As used herein, the singular forms also include the plural forms unless the phrases clearly indicate the opposite. As used herein, the meaning of 'comprising' specifies specific characteristics, regions, integers, steps, operations, elements and/or components, and other specific characteristics, regions, integers, steps, operations, elements, components and/or groups. does not exclude the presence or addition of

Hereinafter, an automatic response providing apparatus 20 and method according to an embodiment of the present invention will be described with reference to the accompanying drawings.

First, with reference to FIG. 1 , an automatic response providing system including the automatic response providing apparatus 20 according to the present example will be described.

As shown in FIG. 1, the automatic response providing system of this example includes a user terminal 10, an automatic response providing device 20 communicating with the user terminal 10, and a database unit connected to the automatic response providing device 20 ( 30) may be provided.

The user terminal 10 may be a terminal of a participant who accesses the automatic response providing device 20 and conducts a conversation with the automatic response providing device 20 .

Therefore, the user terminal 10 communicates with the automatic response providing device 20, inputs a sentence through a chat window and transmits it to the automatic response providing device 20, and inputs the sentence transmitted from the automatic response providing device 20. (Hereinafter, a sentence input to the user terminal 10 is referred to as an 'input sentence') and may receive and output a response.

The user terminal 10 may be a fixed computer such as a desktop computer or a portable computer such as a smart phone.

As shown in FIG. 2A, the user terminal 10 includes a user input unit 11, a communication unit 12, an operation control unit 13 connected to the user input unit 11 and the communication unit 12, and an operation control unit 13. ) It may be the storage unit 14 and the information output unit 15 connected to.

At this time, the components 11 to 15 shown in FIG. 1 are not essential, so the user terminal 10 additionally includes at least one other component other than these components 11 to 15 or among them. Some may be omitted.

The user input unit 11 may generate signals related to commands or data necessary for controlling the operation of the user terminal 10 .

Accordingly, the participant may input the topic of the conversation and input sentences using the user input unit 11 .

The user input unit 11 may include a key pad, a touch pad, a jog switch, or a mouse.

The communication unit 12 is for wired or wireless communication with an external device such as the automatic response providing device 20, and may include at least one communication module enabling communication between networks in which the external device is located.

For example, the communication unit 12 may include an internet module.

The Internet module refers to a module for wireless or wired Internet access, and Internet technologies such as WLAN (Wireless LAN) (WiFi) or Wibro (Wireless broadband) may be used.

However, in an alternative example, the communication unit 12 may include at least one other communication module, such as a short-distance communication module or a mobile communication module, as well as an Internet module.

The operation controller 13 is a control module 231 that controls overall operations of the user terminal 10 and may be a processor.

Therefore, when the operation control unit 13 is connected to the automatic response providing device 20 according to the signal applied to the user input unit 11, a category selection screen for selecting a category of conversation and a conversation with the automatic response providing device 20 Controls the output of the dialog window for this, outputs an input sentence currently input through the user input unit 11 (hereinafter, the currently input input sentence is referred to as 'current input sentence') to the dialog window, and also controls the user input unit 11 The selected category and the current input sentence input through the automatic response providing device 20 may be transmitted.

In addition, the operation control unit 13 allows a response corresponding to the current input sentence to be output on the chat window when it is transmitted from the automatic response providing device 20 .

The storage unit 14 is a storage medium for storing data necessary for the operation of the user terminal 10 or data generated during operation, and may be a memory such as a hard disk or a ROM.

The information output unit 15 is for generating an output related to time under the control of the operation controller 13 and may include a display module.

Therefore, since the display module can display the corresponding image on the screen according to the operation of the operation controller 13, it can output a category selection screen and a chat window.

Such a display module may include at least one display device among a liquid crystal display, an organic light emitting diode display, a flexible display, and a 3D display.

In an alternative example, the information output unit 15 may additionally generate an output related to at least one of auditory and tactile as well as visual output.

Therefore, when the information output unit 15 generates an output related to hearing, the information output unit 15 may additionally include a microphone and a speaker, and when generating an output related to the tactile sense, the information output unit 15 ) may further include a vibration motor.

The automatic response providing device 20 automatically generates and transmits a response corresponding to the current input sentence transmitted from the user terminal 10 to the user terminal 10, and can perform a chatbot function.

The automatic response providing device 20 may be a server or a computer.

As shown in FIG. 2B, the automatic response providing device 20 of this example includes a communication unit 21, a response control unit 23 connected to the communication unit 21, and a storage unit 24 connected to the response control unit 23. can be provided.

These components 21-23 are also not essential, so the automatic response providing device 20 may additionally include at least one other component other than these components 21-23 or omit some of them. can

Similar to the communication unit 12 of the user terminal 10, the communication unit 21 is for wired or wireless communication with external devices such as the user terminal 10 or the database unit 30, and between networks where the external devices are located. It may include at least one communication module enabling communication of

For example, the communication unit 21 of the automatic response providing device 20 may include an Internet module, but may additionally include other types of communication modules.

The response control unit 23 also controls overall operations of the automatic response providing device 20 and may include a processor.

As shown in FIG. 2B, the response controller 23 includes a control module 231, an emotion type extraction module 232, a mood type determination module 233, as shown in FIG. ) and a response generation module 234.

The control module 231 may divide the current input sentence input from the user input unit 11 into word units and morpheme units, and extract a final keyword for the current input sentence using the separated morphemes.

In addition, the control module 231 may extract the intended word from the extracted final keyword, determine the type of intent for the final keyword extracted as the intended word, calculate the number of types for each type of intent for the determined type of intent, and calculate the largest The number of intention types may be determined as the intention type of the current input sentence.

The emotion type extraction module 232 connected to the control module 231 may receive the current input sentence divided into word units and morpheme units from the control module 231 .

Therefore, the emotion type extraction module 232 extracts new words and emoticons from each separated word, extracts emotional words from the separated morphemes, and determines the emotion type for each of the extracted new words, emoticons, and sentiment words, and determines the current sentence. Emotion types can be determined.

The mood type determination module 233 connected to the emotion type extraction module 232 compares whether the emotion type of the current sentence input from the emotion type extraction module 232 is the same as the emotion type of the immediately previous sentence, and determines whether or not the emotion type of the immediately previous sentence is the same. The current sentence emotion value can be calculated.

In addition, the mood type determination module 233 calculates the current sentence emotion value, reduces the previous sentence emotion value by a set value, and then sums the sentence emotion values for each emotion type for all input sentences up to the current conversation to obtain the largest sentence emotion value. The emotion type having the emotion value may determine the mood type for the current conversation.

The response generating module 234 may be connected to the control module 231 and generate a response to the current input sentence by machine learning using the conversation category, intention type, current input sentence, and mood type applied from the control module 231. can create

Then, the response generation module 234 outputs the generated response to the control module 231 so that the response generated by the control module 231 to the user terminal can be transmitted.

The storage unit 24 of the automatic response providing device 20 is also a storage medium for storing data necessary for the operation of the automatic response providing device 20 or data generated during operation, such as a hard disk or a ROM. It may be a memory.

The database unit 30 may be connected to the automatic response providing device 20 by wire or wirelessly, and may include a plurality of databases for storing and managing data necessary for the operation of the automatic response providing device 20, respectively.

As shown in FIG. b, the database unit 30 of this example includes a new coined word and emoticon sentiment dictionary database, a morpheme dictionary database 32, a sentiment word dictionary database 33 storing sentiment words, and storing synonyms and synonyms for words. similar and thesaurus database, and an intention type database 35 storing the types of intentions for words.

A new word and emoticon emotion dictionary database may include items in which new words and emoticons and types of emotions (ie, types of emotions) for each new word and each emoticon are stored in correspondence with each other.

A newly coined word is a non-standard word, and may be an abbreviation of an existing word or sentence, a newly created word, or a letter composed of at least one consonant vowel.

), 이미지(예,

), 4바이트 유니코드 문자(예,

) 등을 포함할 수 있다.An emoticon is a symbol that expresses an emotion, and is an ASCII character (e.g.,

), images (eg,

), 4-byte Unicode characters (e.g.,

) and the like.

Emotion types may include various human emotions, such as 'joy', 'sadness', 'trust', 'anger', 'fear', 'expectation', 'surprise' and 'unpleasure'.

Therefore, the emotion type extraction module 232 extracts the newly coined word and emoticon included in the current input sentence through a comparison operation with the new coined word and emoticon emotion dictionary database, and determines the emotion type for each extracted new word and emoticon. there is.

The morpheme dictionary database 32 is a database in which a morpheme dictionary for morpheme analysis used for segmentation of a current input sentence is stored, and the morpheme dictionary may include parts of speech and usage information for each morpheme.

In the emotional word dictionary database 33 , emotional words that are words representing emotions and emotions and emotion types corresponding to each emotional word may be stored in correspondence with each other.

Accordingly, the emotion type extraction module 232 may extract a sentiment word from each morpheme of the current input sentence through a comparison operation with the sentiment word dictionary database 33 and determine the emotion type for each extracted sentiment word. there is.

The synonym and thesaurus database 34 (34) may store representative words and synonyms and synonyms of each representative word corresponding to each other. Accordingly, when a word extracted as a preliminary keyword in the current input sentence is the same as a similar word or a synonym, the control module 231 may replace the corresponding word with a representative word and determine the replaced representative word as a final keyword.

The intent type database 35 may be used to determine the type of intent (ie, type of intent) of the current input sentence.

In this example, intention types may be classified into “request”, “command”, “warning”, “congratulations”, and “regards”, but these types of intentions may be added or subtracted as needed.

Accordingly, the intention type database 35 may store words related to the participant's intention (hereinafter, a word related to the participant's intention is referred to as 'intention word') and the intention type of each intention word corresponding to each other.

The operation of the automatic response providing system having such a structure will be described.

First, with reference to FIG. 3, the operation of the user terminal 10 will be described.

The operation of the user terminal 10 starts, and as an example, a chatbot operation provided by the automatic response providing device 20 may be performed as an operation of an application installed in the user terminal 10 .

Therefore, when the operation of the application starts, the dynamic control unit of the user terminal 10 displays a conversation category selection screen capable of selecting a conversation category using the video data stored in the storage unit 14 to the information output unit 15. It can be output (S11).

The conversation category selection screen may include a plurality of conversation categories that the user can select. For example, the conversation category may be divided into 'depression', 'anxiety', 'family', and 'sex'.

In this way, when the conversation category selection screen is output to the information output unit 15, the user can select one desired category from among a plurality of conversation categories using the user input unit 11.

Accordingly, the operation control unit 13 may determine whether a conversation category selected by the user from among a plurality of conversation categories exists using a signal applied from the user input unit 11 (S12).

When a corresponding signal is input from the user input unit 11 by a user, that is, a participant's selection operation, the operation control unit 13 reads the signal applied from the user input unit 11 and determines the conversation category selected by the user. (S13), the determined conversation category may be transmitted to the automatic response providing device 20 through the communication unit 12 (S14).

Then, the operation control unit 13 uses the image data stored in the storage unit 14 to output the information output unit 15 to the information output unit 15 so that a dialog window for performing an input for a conversation can be output. Do (S15). Accordingly, a participant may enter a desired sentence or the like into a chat window output to the information output unit 15 to conduct a conversation with the automatic response providing device 20, that is, chatting.

When an input sentence is input to the chat window through the user input unit 11 (S16), the operation control unit 13 may determine the input sentence input to the user input unit 11 and output it to the information output unit 15 (S17). ). Accordingly, the participant can check the input sentence inputted by the participant through the information output unit 15 .

When input of the current sentence to the participant is finally completed, the operation controller 13 may transmit the current input sentence to the automatic response providing device 20 using the communication unit 12 . At this time, the operation control unit 13 may determine that input of the current input sentence is completed when input of a feature key (eg, line change key (enter key)) of the user input unit 11 is determined. . The current input sentence may be stored in the storage unit 14 .

In this way, when the current input sentence is transmitted to the automatic response providing device 20, the operation controller 13 of the user terminal 10 determines whether a response is transmitted from the automatic response providing device 20 through the communication unit 12. It can (S19).

Accordingly, when a response to the current input sentence is transmitted from the automatic response providing device 20, the operation control unit 13 can store the transmitted response in the storage unit 14 and also output it to the information output unit 15. (S110).

Through this operation, during the chatbot operation with the automatic response providing device 20, the participant can input a desired input sentence through the chat window and receive a response from the automatic response providing device 20.

Next, an operation of the automatic response providing apparatus 20 generating a response to the current input sentence and transmitting the generated response to the user terminal 10 will be described using FIGS. 4 to 6 .

First, an example of an operation of the control module 231 will be described with reference to FIG. 4 .

When the operation of the automatic response providing apparatus 20 starts, the control module 231 may receive the conversation category transmitted from the user terminal 10 through the communication unit 21 and store it in the storage unit 24 (S21). ).

Then, when the current input sentence is transmitted from the user terminal 10, the control module 231 may store the current input sentence in the storage unit (S22).

In this example, the storage unit 24 may include a user input table in which a current input sentence is stored, and this user input table may be divided into a token part, a sentence part, and an emotion part. there is.

The token unit stores an identification number for identifying a participant in a current conversation, that is, an identification number of the user terminal 10. For example, the identification number may be an IP address.

The sentence part is a part where the transmitted input sentence itself is stored, and the emotion type extracted for the current input sentence may be stored in the emotion part.

As an example, the input sentence transmitted from the user terminal 10 is 'I'm having a hard time because my boss talks too much. ㅠㅠ.', the control module 231 writes 'I'm having a hard time because my boss is talking too much. It is possible to store the entire current input sentence of '.

Since data in the user input table is accumulated whenever a current input sentence is input, the number of user input tables is independent of the number of input sentences, but may be the same.

Next, the control module 231 may perform an intention determination operation for determining the type (ie, intention type) of the intention (ie, the participant's thoughts or plans) for the current input sentence using the current input sentence.

To this end, the control module 231 may divide the current input sentence stored in the storage unit 24 into word units and then store them in the storage unit 24 (S23). At this time, the control module 231 may separate words using spaces included in the current input sentence, that is, blank spaces.

Therefore, if the current input sentence 'I'm having a hard time because my boss is talking too much' is separated into word units using spaces, "I", "workplace", "boss", "talk", "too much" ”, “severely”, “because”, “hard”, and “ㅠㅠㅠ” can be separated respectively.

Next, the control module 231 outputs each separated word to the emotion type extraction module 232 so that the emotion type for the current input sentence can be extracted (S24). Therefore, when a separate word for the current input sentence is input from the control module 231, the emotion type extraction module 232 starts an operation to extract new words and emoticons from the input word to determine the emotion type for the current input sentence. can do.

In this example, the emotion type extraction module 232 omits stopwords such as periods (.), exclamation marks (!), question marks (?), and investigations from each input word, and then extracts new words and emoticons. can

Then, the control module 231 separates each separated word into morpheme units and stores them in the storage unit 24 (S25), and then outputs each separated morpheme to the mood type determination module 233. (S26).

In this case, the control module 231 compares the morphemes stored in the morpheme dictionary database 32, separates the corresponding word into morpheme units again, tags the corresponding parts of speech, and stores them in the storage unit 24.

Accordingly, when a separated morpheme for the current input sentence is input from the control module 231, the emotion type extraction module 232 starts an operation to extract a sentiment word from the input morpheme to determine the emotion type for the current input sentence. can

As an example, through word-by-word separation, “I”, “work”, “boss”, “talk”, “too”, “severely”, “because”, “tough”, and “ㅠㅠㅠ” The entered input sentences are again “me/np”, “a/jx”, “work/nng”, “boss/nng”, “go/jks”, “say/nng”, “to/jko”, “too/mag ”, “Shimha/VA”, “Ge/EC”, “Ha/VV”, “Arthur/EC”, “Hard/VA” and “Uh/EC” can be separated into morpheme units.

Since the newly coined word or emoticon is not registered in the corresponding database 32 as a morpheme, the newly coined word “ㅠ” can be deleted without being separated into a morpheme when the morpheme is separated.

In this morpheme analysis, parts of speech may be assigned to each separated morpheme, as described above.

Here, NP is a pronoun, JX is an auxiliary, NNG is a common noun, JKS is a nominative case, JKO is an objective case, MAG is a common adverb, VA is an adjective, EC is a connecting ending, VV is a verb, and EC is a connecting ending. It can be a part of speech symbol.

After the morpheme analysis of the current input sentence is completed, the control module 231 may analyze the syntax using each separated morpheme to extract a preliminary keyword for the current input sentence.

In this example, preliminary keywords may be subject, object, and predicate.

Since the subject is a word with a nominative particle, the object is a word with an object particle, and the predicate can be a verb or an adjective, the control module 231 uses the part-of-speech symbol assigned through morpheme analysis in the current input sentence. A subject, an object, and a predicate may be determined and extracted as preliminary keywords (S27).

In this case, in the case of adjectives, adjectives located immediately before the noun may be excluded from the predicate. However, even if an adjective is located in front of a noun, if the adjective is a VP (verb phrase) after parsing, the adjective may be regarded as a predicate.

In this example, since the syntax analysis is to analyze the conversation intention of the participant who inputs the input sentence, the control module 231 may determine the conversation intention by extracting the subject, object, and predicate of the input sentence.

Thus, “me/NP”, “a/JX”, “work/NNG”, “boss/NNG”, “go/JKS”, “say/NNG”, “to/JKO”, “too/MAG”, “ The input sentences separated by morpheme units into “shimha/VA”, “crab/EC”, “ha/VV”, “Arthur/EC”, “hard/VA” and “uh/EC” are, for example, I/ NP/JX/NP_SBJ”, “work/NNG/NP”, “boss/NNG/JKS/NP_SBJ”, “say/NNG/JKO/NP_OBJ”, “too/MAG/AP”, “SG/VA It can be parsed and stored in the storage unit 24 in the form of tagging sentence components, such as "Ge/EC/VP", "Ha/VVAs/EC/VP", and "Hard/VA/EC/VP". At this time, syntax analysis can be done using the syntax analysis function of the Konlpy Python package.

Therefore, in the current input sentence 'I'm having a hard time because my boss is talking too much ㅠㅠ', the control module 231 responds with "I/NP is/JX/NP_SBJ", "boss/NNG is/JKS/NP_SBJ", " Say/NNG/JKO/NP_OBJ”, “Severe/VAG/EC/VP”, “Ha/VVAS/EC/VP”, and “HARD/VA/EC/VP” can be extracted as preliminary keywords. .

Next, the control module 231 compares each preliminary keyword extracted from the current input sentence with similar words and synonyms stored in the synonym and synonym dictionary database 34 to extract a representative word for the corresponding preliminary keyword, and then extracts the extracted representative word. The keyword may be stored in the storage unit 24 as a final keyword (S28).

Accordingly, the control module 231 may determine whether a preliminary keyword corresponding to a synonym or similar word exists among the extracted preliminary keywords through a comparison operation with the synonym and thesaurus dictionary database 34 .

If there is a preliminary keyword identical to a synonym or synonym stored in the synonym and synonym dictionary database 34 among the extracted preliminary keywords, the control module 231 converts a representative word corresponding to the synonym or synonym to the final keyword of the corresponding preliminary keyword. can be determined by

However, among preliminary keywords, preliminary keywords that do not correspond to synonyms or synonyms, that is, preliminary keywords that are not stored in the synonym and synonym dictionary database 34 may be in a state in which no analog or synonym corresponding thereto exists. Accordingly, in this case, the control module may determine the corresponding similar keyword as the final keyword.

As already described, the synonym/thesaurus database stores representative words and similar words and synonyms corresponding to each representative word.

Accordingly, the control module 231 extracts a representative word corresponding to the corresponding preliminary keyword when each preliminary keyword is stored in the corresponding database 34 as a similar word or a synonym, and the extracted representative word It can be determined by the final keyword.

For example, when the representative word is 'difficult', synonyms and synonyms of 'difficult' may be 'difficult', 'difficult', 'difficult', 'difficult', and the like.

When the representative word is 'severe', similar and synonyms of 'severe' can be 'severe' and 'severely', and when the representative word is 'hada', similarities and synonyms of 'hada' can be 'severely' and 'severely'. It can be 'so', 'so' and 'so'.

Therefore, the representative word of "severely" as a preliminary keyword may be "severe", the representative word of 'haseo' may be 'do', and the representative word of 'severe' may be 'difficult'.

However, when the preliminary keyword is not stored as a synonym or synonym in the database 34, the control module 231 may determine that the preliminary keyword itself is a representative word rather than a synonym or synonym corresponding to the representative word. there is. Accordingly, when the preliminary keyword is not stored as a synonym or synonym in the corresponding database 34, the control module 231 may determine the corresponding preliminary keyword as the final keyword.

Therefore, since representative words to be replaced do not exist for the preliminary keywords “I”, “boss”, and “word”, the control module 231 may store “I”, “boss”, and “word” as final keywords. there is.

Therefore, final keywords for the current input sentence "I'm having a hard time because my boss is talking too hard ㅠㅠ" may be 'me', 'boss' and 'talk', 'severe', 'do', and 'difficult'.

In this way, when the final keyword for the current input sentence is determined, the control module 231 compares each of the determined final keywords with the intention type database 35 to determine the intention type for the current input sentence (S29) and generates a response. It can be output to the module 234 (S210). At this time, the control module 231 may output the current input sentence stored in the storage unit 24 and the selected dialogue category along with the type of intention to the response generation module 234 (S210).

In this example, intention types may be classified into “request”, “command”, “warning”, “congratulations”, and “regards”, but these types of intentions may be added or subtracted as needed.

As already described, the intent type database 35 may store intent words and intent types corresponding to each intent word.

Accordingly, since final keywords that are not stored as intended words in the intention type database 35 may exist among final keywords, intended words respectively corresponding to all final keywords may not exist in the intention type database 35 .

In this way, among the final keywords 'me', 'boss' and 'talk', 'severe', 'do' and 'difficulty' for 'I'm having a hard time because my boss talked too much', the intention type database (35 ) may be 'severe' and 'difficult', and the intention types of these 'severe' and 'difficult' may be 'regards', respectively.

In this way, if the intent type for the final keyword, that is, the final keyword stored as an intent word in the intent type database 35 is determined, the control module 231 sums up the number of final keywords having the same intent type, The number of each determined intention type may be calculated.

Then, the control module 231 may determine the intention type having the largest number among the calculated numbers for each intention type as the intention type for the current input sentence.

Accordingly, the intention words stored in the intention type database 35 may be 'serious' and 'difficult', and the intention types of these 'severe' and 'difficult' may be 'regards', respectively.

Because of this, in the current input sentence 'I'm having a hard time because my boss is talking too much ㅠㅠ', the number of intention types for request, command, warning, and congratulation are all '0', whereas the number of intention types for regards Since the number is '2', the control module 231 determines the intention type of the current input sentence for 'I'm having a hard time because my boss is talking too much ㅠㅠ' as 'regards' and outputs it to the response generation module 234. can

In this way, the control module 231 divides the current input sentence into word units and outputs them to the emotion type extraction module 232, separates the current input sentence into word units and outputs it to the mood type determination module 233 by dividing the current input sentence into morpheme units, ) determines the mood type (ie, mood type) for the current conversation so that the determined mood type can be output to the response generation module 234.

Accordingly, the response generation module 234 controls the control module 231 by generating a response to the current input sentence using the intention type, the current input sentence, the selected dialogue category, and the mood type transmitted from the control module 231. It can be output to module 231.

In this way, when a response to the current input sentence is input due to the operation of the response generation module 234, the control module 231 stores the input response in the storage unit 24 and also uses the communication unit 21 to It can be output to the user terminal 10 (S211).

Accordingly, the user terminal 10 may output the response transmitted from the automatic response providing device 20 to the chat window as a response to the current input sentence.

Next, referring to FIG. 6 , the operation of the emotion type extraction module 232 will be described.

As described above, the emotion type extraction module 232 is for determining the emotion type for the current input sentence, and can determine the emotion type using the newly coined word and emoticon included in each separated word and each morpheme. .

In this example, the determined emotion type may be 'joy', 'sadness', 'trust', 'anger', 'fear', 'expectation', 'surprise', and 'unpleasure', but is not limited thereto.

Therefore, when separate words and morphemes for the current input sentence are input from the control module 231, the emotion type extraction module 232 determines the emotion types included in the current input sentence using the separated words and performs a control operation. The control operation (S32) may be performed by determining the type of emotion included in the current input sentence using the morpheme separated from (S31).

At this time, the control operations S31 and S32 may be performed in parallel as shown in FIG. 5 . However, unlike this, these control operations (S31 and S32) may be performed serially in series. For example, a control operation (S31) using a word may be performed first, and then a control operation (S32) using a morpheme may be performed.

For this reason, when each phrase separated from the current input sentence is input from the control module 231, the emotion type extraction model stores each input word in the storage unit 24 (S311) and is included in each word New words or emoticons included in the current input sentence may be extracted by extracting newly coined words or emoticons (S312).

Therefore, the emotion type extraction module 232 reads each separated word of the current input sentence stored in the storage unit 24, deletes the stopwords included in each word, and then each word from which the stopwords are deleted and the newly coined word. and compared with new words and emoticons stored in the emoticon emotion dictionary database, it is possible to determine whether new words and emoticons exist in each word.

Depending on the word, there may be a case in which both a coined word and an emoticon are not extracted, a case in which both a coined word and an emoticon are extracted, or a case in which only one of a coined word and an emoticon is extracted.

In this example, the reason why newly coined words and emoticons are extracted by separating the current input sentence into word units rather than morphemes is that new words and emoticons are not registered in Korean dictionaries in many cases. That is, new words and emoticons are not separated into morphemes because they are not recognized as morphemes during morpheme analysis, and thus new words and emoticons are not separated into morpheme units. Therefore, in this example, the emotion type extraction module 232 compares the newly coined word and emoticon stored in the new word and emoticon emotion dictionary database 31 with each separated phrase, and extracts the newly coined word and emoticon included in each phrase. do.

In addition, when the separated words include stopwords, it is difficult or impossible to extract new words and emoticons, so as described above, the stopwords included in each word are removed before the comparison operation with the new words and emoticon sentiment dictionary database 31 action can be made.

Through this comparison operation, the emotion type extraction module 232, when at least one of the newly coined word and emoticon exists in each word, extracts the newly coined word and emoticon existing in the corresponding word, and then the emotion type corresponding to the extracted new word and emoticon, respectively. can be determined and stored.

As already described, the newly coined word and emoticon emotion dictionary database stores not only newly coined words and emoticons but also emotion types corresponding to each new word and emoticon.

As an example, it is assumed that “ㅠ” is stored as a newly coined word in the new word and emoticon emotion dictionary database 31, and the emotion type corresponding to “ㅠ” is “sad”.

In this case, each word separated by the word separation operation of the control module 231 “I”, “work”, “boss”, “talk”, “too”, “seriously”, “because”, “ If "It's hard" or "ㅠㅠㅠ" is input, the emotion type extraction module 232 may extract the newly coined word 'ㅠㅠㅠ' from the word of "ㅠㅠㅠ".

Then, the emotion type extraction module 232 may determine the emotion type for the extracted neologism “ㅠ” as “sad”.

In this way, the emotion type extraction module 232 compares each word with the newly coined word and emoticon stored in the new word and emoticon emotion dictionary database 31, extracts the newly coined word and emoticon included in the current input sentence, and extracts each extracted emoticon and emoticon. The emotion type for the newly coined word may be determined and stored in the storage unit 24 . At this time, the determined emotion type may be stored in correspondence with the corresponding neologism and emoticon.

In addition, when the current input sentence separated by morphemes from the control module 231, that is, the morphemes, is input, the emotion type extraction module 232 compares the corresponding morphemes with the emotion words stored in the emotion word dictionary database 33 to correspond to the corresponding morphemes. It may be determined whether the morpheme is a word stored in the emotional word dictionary database 33 as a sentimental word (S322).

Accordingly, the emotion type extraction module 232 may determine whether the input morpheme exists in the emotion word dictionary database 33 as a emotion word.

If the input morpheme exists as a sentiment word in the emotional word dictionary database 33, the emotion type extraction module 232 may determine the input morpheme as an emotion or a sentiment word representing emotion and extract the corresponding morpheme as a sentiment word. .

Then, the emotion type extraction module 232 may determine the emotion type stored to correspond to the extracted sentiment word, and then store the determined emotion type corresponding to the corresponding morpheme in the storage unit 24 (S323). .

In this way, the emotion type extraction module 232 compares all morphemes of the current input sentence with the emotion word dictionary database 33, extracts a morpheme corresponding to the emotion word from among the input morphemes, and then determines the emotion type corresponding thereto. You can do it.

In this way, when the emotion type for the newly coined word and emoticon included in the current input sentence and the emotion type for the emotion word included in the current input sentence are determined, the emotion type extraction module 232 includes the emotion type included in the current input sentence. The emotion type for the current input sentence (hereinafter, the emotion type for the current input sentence will be referred to as 'current sentence emotion type') can be determined using the emotion type.

At this time, there may be cases in which the current input sentence does not determine any type of emotion among joy, sadness, trust, anger, fear, fear, surprise, and displeasure. . In this case, the emotion type extraction module 232 may determine 'no emotion' as the emotion type of the current sentence.

Accordingly, the emotion type extraction module 232 may sum the emotion types by type among all emotion types for the current input sentence determined using the newly coined word, emoticon, and morpheme (S33).

For example, if the input sentence is 'I'm having a hard time because my boss is talking too much', the type of emotion determined using the new word and emoticon may be 'sadness' corresponding to one new word 'ㅠㅠㅠ' .

In addition, the type of emotion determined using the morpheme may be 'unpleasant' corresponding to 'severe' and 'sadness' corresponding to 'difficulty'.

Therefore, the types of total judged emotions in the input sentence for 'I'm having a hard time because my boss talked too harshly ㅠㅠ' may be 'sadness' and 'displeasure', and the emotion of 'sadness' determined in the current input sentence The total number of types may be '2', and the total number of emotion types 'unpleasant' may be '1'. The total number of each of the different emotion types (ie, joy, trust, anger, fear, expectation, and surprise) may be '0'.

In this way, when the total number of emotions for each type determined in the current input sentence is calculated, the emotion type extraction module 232 determines the emotion type having the largest value among the calculated total number of emotions in the current sentence, which is the emotion type for the current input sentence. After determining as a type (S34), it is output to the mood type determination module 233 and can be stored in the 'emotion unit' of the user input table for the current input sentence generated in the storage unit 24 (S35).

Therefore, since the current sentence emotion type for the current input sentence 'I'm having a hard time because my boss talked too much' can be 'sadness' with a total number of two, the emotion type extraction module 232 determines the current sentence emotion type. Type can be judged as 'sad'

At this time, if there are two or more types of emotions determined for the current input sentence and there are two or more emotions having the same maximum value of the total number calculated for each emotion type, the emotion type extraction module 232 performs the current input The emotion for the sentence may be determined as 'no emotion' and transmitted as the mood type determination module 233, and may be stored as 'no emotion' in the emotion unit of the user input table.

In this way, when the current sentence emotion type is input from the emotion type extraction module 232, the mood type determination module 233 determines the mood type (ie, mood type) for the current conversation using the input current sentence emotion type. can judge

At this time, the determined mood type may be one of joy, sadness, trust, anger, fear, expectation, surprise, and displeasure, or stability.

To this end, first, the mood type determination module may compare the emotion type of the current input sentence and the emotion type of the immediately previous input sentence stored in the storage 24 (ie, the emotion type of the immediately previous sentence) with each other (S41 ).

Therefore, if the current sentence emotion type and the immediately previous sentence emotion type are the same (S42), that is, if the emotion types of the sentences input twice in succession are the same, the value for the current sentence emotion type (hereinafter, the current sentence emotion type) A numerical value for the type (referred to as 'current sentence emotion value') may be set to '2' (S43).

However, if the emotion type of the current sentence and the emotion type of the immediately previous sentence are different from each other (S42), that is, if the emotion types of the sentences input twice in succession are different from each other, the mood type determination module 233 determines the emotion value of the current sentence ' It may be set to 1' (S44).

Accordingly, the current sentence sentiment value may have a value of '2' or '1'.

Then, the mood type determination module 233 may adjust sentiment values of all sentences for previous input sentences except for the current sentence sentiment values. Accordingly, the mood type determination module 233 may reduce the current sentiment value of each sentence for all previous input sentences by a set size (eg, 0.1) and store it as the current sentence sentiment value for the previous input sentence (S45). In this case, the emotion type for each previous input sentence may have one of 'joy', 'sadness', 'trust', 'anger', 'fear', 'expectation', 'surprise', and 'unpleasure'.

For this reason, the sentence sentiment value may have a minimum value of '0' by sequentially decreasing from a maximum of 2 by 0.1 each time a new input sentence is transmitted from the user terminal 10 . Therefore, in this example, the sentence emotion value of each sentence input to the chat window may have a value in the range of 0 to 2, so that the minimum value of the sentence emotion value is '0' and the maximum value of the sentence emotion value is '2'. can

Next, the emotion determination module may calculate a total emotion value for each emotion type by summing sentence emotion values calculated for each emotion type, targeting all input sentences input to the current chat window (S46).

Next, the emotion determination module may determine the intensity of emotion (ie, intensity of emotion) for each emotion type by using the size of the total emotion value for each emotion type (S47).

In this example, the mood type determination module 233 may determine the emotional intensity of an emotion type (eg, 'sadness') having a total emotion value equal to or greater than a first set value (eg, 3.5 or greater) as 'strong'. there is.

In addition, the mood type determination module 233 determines the emotion intensity for the emotion type (eg, 'anger') in which the sum of total emotion values is less than the first set value and greater than or equal to the second set value (eg, greater than or equal to 0.4 and less than 3.5). can be judged as 'weak'. Here, the second set value may be a value smaller than the first set value.

In this way, if the emotional intensity for each emotion type is determined, the mood type determination module 233 may determine whether an emotion type having 'strong' emotional intensity exists (S48).

If there is an emotion type having 'strong' emotional intensity (S48), the mood type determining module 233 may determine whether there is one emotion type having 'strong' emotional intensity (S49).

When there is only one emotion type having an emotion intensity of 'strong' (S48), the mood type determination module 233 may determine the emotion type with an emotion intensity of 'strong' as the mood type for the current conversation (S49). ).

When there is one emotion type having an emotion intensity of 'strong', the mood type determination module 233 determines the emotion type having an emotion intensity of 'strong' as a mood type for the current conversation and stores it in the storage unit 24. It can (S410).

However, if there are a plurality of emotion types with 'strong', the mood type determination module 233 determines the emotion type with the largest emotion value among the emotion types with 'strong' as the mood type for the current conversation and stores it. It can (S411).

However, in step S48, when the emotion type determination module 233 does not exist the emotion type having 'strong' emotion intensity, the mood type determination module 233 determines the emotion type having 'weak' emotion intensity. It can be determined whether there exists (S412).

If emotion types having 'weak' emotion intensity exist, the mood type determination module 233 determines the emotion type having the largest total emotion value among the emotion types having 'weak' as the mood type for the current conversation. After that, it can be stored (S413).

However, when there is no emotion type having 'weak' emotion intensity, that is, when the total emotion values for each emotion type calculated are less than the second set value of 0.4 (S412), the mood type determination module 233 determines the current The mood type for the conversation may be determined to be 'stable' and then stored (S414).

Mood Type Determination Module 233

In this way, if the mood type for the current conversation is determined using the magnitude of the total emotion value for each emotion type, the mood type determination module 233 may output the determined mood type to the control module 231 and also calculate the A numerical value (ie, an emotional value) may also be output (S415).

For this reason, the mood type determination module 233 determines whether there is an emotion type having a total emotion value equal to or greater than a predetermined value (eg, the second set value), and if an emotion type having a total emotion value equal to or greater than the predetermined value exists, the most emotional type exists. An emotion type having a large value may be determined as a mood type.

However, when there is no emotion type having a total emotional value greater than or equal to a predetermined value and the total emotional value of all emotion types is less than the predetermined value, the mood type determination module 233 may determine the mood type as 'stable'.

In this case, when mood conversation is determined using emotion intensities (eg, 'strong' and 'weak'), data comparison time is greatly reduced, and thus processing time of the mood type determination module 233 can be greatly reduced.

For example, if the order of the current input sentence is 5th, the emotion type of the 5th input sentence is 'sadness' and the sentence emotion value is 2, the emotion type of the 4th input sentence is 'sadness' and the sentence emotion value is 1.9, the emotion type of the third input sentence is 'sadness' and the sentence emotion value is 0.8, the emotion type of the second input sentence is 'anger', the sentence emotion value is 0.7, and the first input sentence Assume that the emotion type of is 'fear' and the sentence emotion value is 0.6.

In this case, the total emotion value for 'sadness' up to the current input sentence is 4.7, the total emotion value for 'anger' is 0.7, and the total emotion value for 'fear' is 0.6.

Therefore, since the number of emotion types having an emotion level of 'strong' is one, the mood type determination module 233 directly operates the mood type determination module 233 without comparing the total emotional values calculated for each emotion type. It is possible to determine the mood type for 'sadness' and output it to the control module 231.

When the emotion type of the current input sentence calculated by the operation of the emotion type extraction module 232 is 'no emotion', the emotion value of the current sentence may be '0'.

In this way, when the mood type for the current conversation is calculated, the mood type determination module 233 outputs the determined mood type to the control module 231 and the control module 231 outputs the determined mood type to the response generation module 234. , so that a response to the current input sentence can be generated.

Therefore, as described above, the control module 231 controls the type of category stored in the storage unit 24, the current input sentence stored in the user input table, and the current conversation input from the mood type determination module 233. The mood type of and the intention type of the current input sentence may be input to the response generating module 234 .

The response generating module 234 generates a response to the current input sentence by taking the type of category, the current input sentence, the type of mood, and the type of intention applied from the control module 231 as inputs, and transmits the response to the user terminal 10. there is.

However, in an alternative example, control module 231 may output the determined intent type to response generation module 234, and mood type determination module 233 may output the determined mood type to response generation module 234. As a result, the response generation module 234 may generate a response to the current input sentence using the intention type and the mood type respectively input from the control module 231 and the mood type determination module 233. In this case, the response generation module 234 may also be connected to the mood type determination module 233 to receive a mood type.

Response generation module 234 of the present example may use a machine learning algorithm, such as deep learning learning; for example, the seq2seq algorithm may be used.

Also, in this example, an Anaconda 3 virtual environment was built to develop an automatic response generation algorithm using the seq2seq algorithm in Ubuntu 20.04.

In addition, since NVIDIA CUDA is installed in the response generation module 234, algorithm learning time can be reduced.

Since the response generation module 234 can crawl and use Twitter's tweets as learning data, and the publisher retrieves and uses tweets in the form of mention and reply among the crawled tweets, the response generation module 234 is Tweets that do not exist in the form of mention or reply may not be used as training data.

The original tweet and related tweets in the form of mention and reply are regarded as conversations and used as independent learning data.

The response generation module 234 may extract two dialogue scripts for one subject from entertainment, dramas, and movies and use them as learning data.

In addition, AIHub can use wellness conversation script dataset, Korean conversation dataset, Twitter-based daily conversation dataset, and interactive Korean agent dataset as learning data.

The response generation module 234 may classify each conversation dataset into eight emotions (ie, joy, sadness, trust, anger, fear, expectation, surprise, and displeasure), and tag the emotions represented by the conversation data, and Based on conversation data tagged with emotions, a response generation algorithm can be trained.

In this way, since the response generation module 234 uses the mood type of the current conversation determined through the mood type determination module 233 as an input value of the seq2seq algorithm, the conversation participant has a suitable response for the current input sentence suitable for the current mood. response can be generated.

The technical features disclosed in each embodiment of the present invention are not limited to the corresponding embodiment, and unless incompatible with each other, the technical features disclosed in each embodiment may be merged and applied to other embodiments.

In the above, the embodiments of the present invention have been described. The present invention is not limited to the above-described embodiments and accompanying drawings, and various modifications and variations will be possible from the viewpoint of those skilled in the art to which the present invention belongs. Therefore, the scope of the present invention should be defined by not only the claims of this specification but also those equivalent to these claims.

10: user terminal 20: automatic response providing device
30: database unit 21: communication unit
23: response control unit 23 231: control module
232 Emotion type extraction module 233 Mood type determination module
234: response generation module 24: storage unit
31: New words and emoticon sentiment dictionary database
32: morpheme dictionary database
33: Emotional dictionary database
34: Synonyms and thesaurus database
35: Intent Type Database

Claims (16)

The current input sentence is divided into word units and morpheme units, a final keyword for the current input sentence is extracted using the separated morphemes, an intended word is extracted from the final keyword, and a final keyword extracted as the intended word is extracted. a control module that determines intent types, calculates the number of intent types for each intent type, and determines the intent type having the largest number as the intent type of the current input sentence;
Connected to the control module, extracts neologisms, emoticons, and sentimental words from the current input sentence separated into word units and morpheme units from the control module, and determines the emotion type for each of the extracted neologisms, emoticons, and sentimental words, and determines the current input sentence. an emotion type extraction module for determining a sentence emotion type; and
It is connected to the emotion type extraction module and compares whether the current sentence emotion type input from the emotion type extraction module and the immediately previous sentence emotion type are the same, and calculates different current sentence emotion values according to the comparison result, and then calculates the previous sentence emotion type. Mood type determination module that determines the mood type for the current conversation by reducing the emotion value by a set size and using the size of the sentence emotion value for each emotion type calculated by summing the sentence emotion values for each emotion type
An automatic response providing device comprising a. According to claim 1,
The control module,
Extracting the subject, object, and predicate of the current input sentence using the separated morphemes to extract preliminary keywords for the current input sentence;
The preliminary keyword is compared with a similar word and synonym database in which similar words and synonyms are stored corresponding to representative words, and if there is a preliminary keyword corresponding to a synonym or similar word among the preliminary keywords, a representative word corresponding to the corresponding synonym or similar word is selected. Determine the final keyword of the preliminary keyword,
An automatic response providing apparatus for determining a preliminary keyword that does not correspond to a synonym or a synonym among the preliminary keywords as a final keyword. According to claim 1,
The emotion type extraction module,
Newly coined words and emoticons and each word are compared with the newly coined word and emoticon emotion dictionary database in which emotion types corresponding to each new word and emoticon are stored, and the newly coined word and emoticon included in each word are extracted, and each extracted new word and each emoticon is compared. To determine the kind of emotion for
Sentiment words are extracted from the morphemes by comparing each morpheme with a sentiment word dictionary database in which emotional words and emotion types corresponding to each emotion word are stored, and emotion types for each extracted emotion word are determined;
An automatic response providing device that calculates the total number of emotions for each emotion type from the emotion types for newly coined words and emoticons and the emotion types for emotional words, and determines the emotion type having the largest number as the emotion type for the current sentence. According to claim 1,
The emotion type extraction module removes stopwords included in each word before comparison with the newly coined word and the emoticon emotion dictionary database. According to claim 1,
wherein the mood type determination module determines, as the mood type for the current conversation, an emotion type corresponding to a sentence emotion value having the largest magnitude among total sentence emotion values equal to or greater than a predetermined value among sentence emotion values for each emotion type. According to claim 5,
The mood type determination module determines the mood type for the current conversation as stable when all sentence emotion values for each emotion type have values less than a predetermined value. According to claim 1,
A response generation module connected to the control module.
Including more,
The control module additionally receives a conversation category and outputs the conversation category, the intention type, and the current input sentence to the response generation module;
The response generation module generates a response to the current input sentence by machine learning using the input conversation category, the intention type, the current input sentence, and the mood type. According to claim 1,
The emotion type is one of joy, sadness, trust, anger, fear, expectation, surprise, and displeasure. separating, by the control module, the current input sentence into word units and morpheme units;
extracting, by the control module, a final keyword for the current input sentence using the separated morphemes;
extracting, by the control module, an intended word from the extracted final keyword, and determining an intended type of the final keyword extracted from the extracted final keyword;
calculating, by the control module, the number of each intent type for the determined intent types, and determining the intention type having the largest number as the intention type of the current input sentence;
The emotion type extraction module receives the current input sentence divided into word units and morpheme units from the control module, and extracts neologisms, emoticons, and emotional words;
determining, by the emotion type extraction module, a current sentence emotion type by determining an emotion type for each of the extracted neologism, emoticon, and emotional word;
comparing, by the mood type determination module, whether the emotion type of the current sentence and the emotion type of the immediately previous sentence input from the emotion type extraction module are the same;
calculating, by the mood type determination module, different current sentence emotion values according to a comparison result between a current sentence emotion type and an immediately previous sentence emotion type;
reducing, by the mood type determination module, a previous sentence emotion value by a set size after calculating the current sentence emotion value; and
determining, by the mood type determination module, the mood type for the current conversation by using the size of the sentence emotion value for each emotion type calculated by adding up the sentence emotion values for each emotion type;
Control method of an automatic response providing device comprising a. According to claim 9,
The step of extracting the final keyword,
extracting, by the control module, a preliminary keyword for the current input sentence by extracting the subject, object, and predicate of the current input sentence using the separated morphemes;
comparing the preliminary keyword with a similar word and synonym database in which similar words and synonyms are stored corresponding to representative words, and determining, by the control module, whether a preliminary keyword corresponding to a similar word or a synonym exists among the preliminary keywords;
determining, by the control module, a representative word corresponding to the corresponding synonym or synonym as a final keyword of the corresponding preliminary keyword, if a preliminary keyword corresponding to a similar word or a synonym exists among the preliminary keywords; and
determining, by the control module, a preliminary keyword that does not correspond to a synonym or a synonym among the preliminary keywords as a final keyword;
Control method of an automatic response providing device comprising a. According to claim 9,
The step of extracting the neologisms, emoticons and emotional words,
The emotion type extraction module compares each word with a new word and emoticon emotion dictionary database storing new words and emoticons and emotion types corresponding to each new word and emoticon;
extracting, by the emotion type extraction module, a newly coined word or emoticon included in each word, if the newly coined word or emoticon stored in the emoticon emotion dictionary database exists in each word; and
The emotion type extraction module determining the emotion type for each new word and each emoticon extracted from the new word and emoticon emotion dictionary database
Control method of an automatic response providing device comprising a. According to claim 11,
The step of extracting the newly coined word, emoticon, and emotional word further comprises removing stopwords included in each word before comparing each word with a new coined word and an emoticon emotion dictionary database. According to claim 9,
The step of determining the emotion type of the current sentence,
comparing, by the emotion type extraction module, each morpheme with a sentiment word dictionary database in which emotion words and emotion types corresponding to each emotion word are stored;
extracting, by the emotion type extraction module, a sentiment word included in the morpheme if the sentiment word stored in the sentiment word dictionary data exists in each morpheme;
determining, by the emotion type extraction module, an emotion type for the emotion word extracted from the emotion word dictionary database;
calculating, by the emotion type extraction module, a total number of emotion types for each emotion type among emotion types for new words and emoticons and emotion types for emotional words; and
determining, by the emotion type extraction module, the emotion type having the largest number among the total number of calculated emotion types as the current sentence emotion type;
Control method of an automatic response providing device comprising a. According to claim 9,
Determining the mood type for the current conversation includes:
determining whether total sentence emotion values greater than or equal to a predetermined value exist among sentence emotion values for each emotion type;
If there is a total sentence emotional value of more than a predetermined value among sentence emotional values for each emotion type, the emotion type corresponding to the sentence emotional value having the largest magnitude among the total sentence emotional values for each emotion type that is equal to or greater than the predetermined value is determined as the mood type for the current conversation. step; and
Determining the mood type for the current conversation as stable if there is no total sentence emotion value greater than or equal to a predetermined value among the sentence emotion values for each emotion type.
Control method of an automatic response providing device comprising a. According to claim 9,
and outputting, by the control module, the conversation category, the intention type, and the current input sentence to a response generating module. According to claim 9,
The emotion type is one of joy, sadness, trust, anger, fear, expectation, surprise, and displeasure.

Images