KR20190046062A - Method and apparatus of dialog scenario database constructing for dialog system - Google Patents

Abstract

The present invention relates to the construction of a dialogue scenario database, and more particularly to a dialogue scenario database which automatically searches, collects, refines, and learns a dialogue scenario in SNS, radio, and broadcast to automatically expand the dialogue area or improve the conversation quality To a dialog scenario database construction method for a dialog system applied to an ontology dialog network.
According to another aspect of the present invention, there is provided a method for constructing a dialog scenario database for an interactive system applied to an ontology dialog network, the method comprising the steps of: (a) extracting a sentence from a conversation- (b) classifying the sentence extracted in the step (a) as a question and an answer; (c) selecting whether the question and answer sentences classified in step (b) are linked and connected to each other by emotion, intention, affirmation, and negation, and creating the selected question and answer sentence as a list-type scenario; (d) converting a question and an answer sentence of a scenario, which is a list type generated in step (c), into a semantic vector and learning; And (e) databaseing the scenarios learned in step (d) into consecutive scenario semantic vectors.

Description

FIELD OF THE INVENTION [0001] The present invention relates to a method and apparatus for constructing a dialogue scenario database for a dialogue system,

The present invention relates to the construction of a dialogue scenario database, and more particularly, to an ontology dialogue service in which dialogue scenarios are automatically searched, collected, refined, and learned in SNS, radio, And a dialogue scenario database construction method for an interactive system applied to a network.

Humans have long dreamed to communicate naturally with people other than humans. Artificial Intelligence (AI) and Big Data are driving the fourth industrial revolution, and at the core of artificial intelligence is the natural dialog communication between humans and objects.

However, until now, the dialogue scenarios have been costly, time-consuming, and labor-intensive, since people manually build the scenarios by refining them. This allows people to collect and refine their own scenarios, which will require a relatively long time, which will allow them to engage in conversations in the past, allowing immediate conversations about current incidents or accidents, trends or issues. I did not.

That is, since only a limited dialogue is used in a limited area, there is a limitation in expanding the dialog area or improving the conversation quality.

It is an object of the present invention to provide an ontology dialogue network in which dialogue scenarios are automatically searched, collected, refined, and learned in SNS, radio, and broadcasting, The present invention provides a method for constructing a dialogue scenario database for an interactive system applied to a dialogue scenario.

According to another aspect of the present invention, there is provided a method of constructing a dialog scenario database for an interactive system applied to an ontology dialog network, the method comprising the steps of: (a) extracting a sentence from a conversation- (b) classifying the sentence extracted in the step (a) as a question and an answer; (c) selecting whether the question and answer sentences classified in step (b) are linked and connected to each other by emotion, intention, affirmation, and negation, and creating the selected question and answer sentence as a list-type scenario; (d) converting a question and an answer sentence of a scenario, which is a list type generated in step (c), into a semantic vector and learning; And (e) databaseing the scenarios learned in step (d) into consecutive scenario semantic vectors.

Preferably, the step (e) further comprises the step of displaying the continuous semantic vector converted into the database on the ontology multidimensional space as one of a semantic word, a semantic node and a semantic cube.

Preferably, the interactive speech file of step (a) is extracted after being converted into text through speech recognition.

Preferably, the step (a) extracts a specific word (word) from a conversation-type article, and parses and extracts the posts and comments retrieved with the specified search term.

Preferably, the sentence extracted in step (a) is automatically converted into a semantic vector.

According to another aspect of the present invention, there is provided a sentence extracting unit for extracting a sentence. A sentence analyzing unit for classifying the sentence extracted by the sentence extracting unit into a question and an answer; A dialog scenario generation unit for generating a scenario in which a sentence classified by the sentence analysis unit corresponds to each other and a dialog is connected; A dialog scenario learning unit for learning a sentence of a dialogue scenario generated by the dialogue scenario generating unit; A dialogue scenario database for storing the dialogue scenarios learned by the dialogue scenario learning unit as semantic vectors and storing them; And an ontology relationship mapping unit for displaying successive semantic vectors of the conversation scenario stored in the conversation scenario database as one of semantic words, meaning nodes, and semantic cubes on the ontology multidimensional space.

According to the present invention, it is possible to reduce the construction cost of the dialogue scenario by automating the dialogue scenario collection, refinement, and construction.

In addition, if a person directly collects scenarios, it takes a lot of time to refine and construct them, so that conversations are made in the past. However, since conversation scenarios are automatically collected and reflected in conversations, So that the conversation quality can be improved.

In addition, since conversation scenarios are constantly collected on various topics, it is possible to have conversations on topics reflecting various perspectives.

1 is a flowchart for explaining a method of constructing a dialogue scenario database for an interactive system according to the present invention;
Figure 2 schematically illustrates a database building apparatus for a dialog scenario for an interactive system according to the present invention;
3 is a detailed view of a database construction apparatus of a dialogue scenario for the dialog system according to FIG.
FIG. 4 to FIG. 5 are views illustrating an example of sentence extraction during the construction of a dialogue scenario database for the dialog system according to the present invention;
6 is a screen showing an example of sentence analysis during construction of a dialogue scenario database for the dialog system according to the present invention;
7 is a screen showing an example of dialogue scenario generation during dialogue scenario database construction for an interactive system according to the present invention;
8 is a screen showing an example of conversation scenario learning during construction of a conversation scenario database for the conversation system according to the present invention.
9 is a diagram illustrating an example of a dialog scenario database for an interactive system constructed in accordance with the present invention displayed on an ontology dialogue network;
Figure 10 is an illustration of a dialog scenario stored in the dialog scenario database according to the present invention;
11 is a screen in which the dialog scenario stored in the dialog scenario data face according to the present invention is expressed by a three-dimensional authoring tool.
12 is an example of a scenario input screen in which a dialog scenario stored in the dialog scenario database according to the present invention is expressed by a three-dimensional authoring tool;
13 is an example of a scenario modification screen in which the dialog scenario stored in the dialog scenario database according to the present invention is expressed by a three-dimensional authoring tool.
14 is an example of a scenario deletion screen in which a dialog scenario stored in the dialog scenario database according to the present invention is expressed by a three-dimensional authoring tool;
FIG. 15 is a block diagram illustrating a sequence in which continuous conversation is performed using the conversation scenario automatic collection and ontology conversation network according to the present invention; FIG.
FIG. 16 is a flowchart showing a sequence in which continuous conversation is performed using an automatic conversation scenario collection and ontology conversation network according to the present invention; FIG.
FIG. 17 is a flowchart illustrating a method of mapping a new conversation sentence into an ontology conversation network, when a new conversation sentence is input, in a continuous conversation using the ontology conversation network according to the present invention.
FIG. 18 is a block diagram illustrating a sequence for implementing conversation quality improvement using continuous conversation using an ontology conversation network according to the present invention; FIG.
19 is a diagram showing a configuration of a continuous conversation system using an ontology conversation network according to the present invention;
FIG. 20 illustrates an embodiment of an ontology talk network structure according to the present invention; FIG.
FIG. 21 illustrates an embodiment of a general conversation classification structure in an ontology conversation network according to the present invention; FIG.
22 illustrates an embodiment of a professional conversation classification structure in an ontology conversation network according to the present invention;
23 illustrates an example of a conversation classification structure that is exchanged with an agent in a hospital call center during a professional conversation in an ontology conversation network according to the present invention.
24 illustrates an example of a classification structure in which a daily conversation, an emotional conversation, and a professional conversation are connected in an ontology conversation network according to the present invention.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. Prior to this, terms and words used in the present specification and claims should not be construed as limited to ordinary or dictionary terms, and the inventor should appropriately interpret the concepts of the terms appropriately It should be interpreted in accordance with the meaning and concept consistent with the technical idea of the present invention based on the principle that it can be defined. Therefore, the embodiments described in this specification and the configurations shown in the drawings are merely the most preferred embodiments of the present invention and do not represent all the technical ideas of the present invention. Therefore, It is to be understood that equivalents and modifications are possible.

1 is a flowchart illustrating a method for constructing a dialogue scenario database for a dialog system according to the present invention.

In the dialogue scenario database building method for the dialog system according to FIG. 1, a sentence is extracted first (S110). Extraction of sentences is extracted from conversational voice files or conversational posts.

If a sentence is extracted in step S110, the extracted sentence is classified into a question and an answer (S120).

Then, it is determined whether the question and answer sentences classified in step S120 correspond to each other and connected to each other by the sentiment, intention, affirmation, and negation (S130), and the selected question and answer sentence is generated as a list-type scenario ).

Subsequently, the scenario and the question and answer sentences of the generated scenario, which are the list forms, are converted into semantic vectors to learn a scenario (S150), and the learned scenario is converted into a database with a continuous scenario meaning vector (S160).

The consecutive semantic vectors of the database-dated scenario are displayed in the ontology multi-dimensional space as one of semantic words, meaning nodes, and semantic cubes (S160).

The method for constructing a dialog scenario database for the dialog system of the present invention shown in FIG. 1 will be described with reference to a dialogue scenario database establishing apparatus for the dialog system of FIGS.

FIG. 2 is a view schematically showing an apparatus for constructing a dialogue scenario database for an interactive system according to the present invention, and FIG. 3 is a detailed view illustrating an apparatus for constructing a dialogue scenario database for the interactive system according to FIG.

2 to 3, an apparatus 100 for constructing a dialogue scenario database for an interactive system according to the present invention includes a sentence extraction unit 110 for extracting a sentence, a sentence extraction unit 110 for extracting a sentence extracted by the sentence extraction unit 110, A dialogue scenario generation unit 130 for generating a scenario in which sentences classified by the sentence analysis unit 120 correspond to each other and conversations are connected to each other, A dialogue scenario learning unit 140 for learning the sentence of the dialogue scenario generated by the dialogue scenario generation unit 130, a dialogue scenario storage unit 140 for storing a dialogue scenario, which is learned by the dialogue scenario learning unit 140, The scenario database 150 and the continuous semantic vectors of the dialog scenario stored in the dialog scenario database 150 are stored in the ontology multidimensional space with semantic words or semantic nodes Hitting include ontology mapping relationship unit 160 to display one of the mean cube.

The sentence extracting unit 110 extracts a sentence from an interactive voice file or a conversation type of text. The sentence extracting unit 110 includes a voice file extracting module 111 for extracting an interactive voice file, 3, an interactive voice file, such as a call center, a radio, and a TV broadcast, is sent to a voice file extraction module 111 through a voice file extraction module 111. The voice file extraction module 111 extracts a sentence When a search word is selected in the search word generator 113, a sentence is extracted because the search is performed by the postexpression module 112 and the posting is parsed. For example, it is possible to extract a sentence from a text after speech recognition of an existing voice recording file provided by a call center, or extract a sentence from text after speech recognition of a consultation content between a customer and an agent in a call center in real time. On the other hand, SNS extracts sentences from posts that talk about various subjects on Twitter or Facebook, searches for SNS with the prepared search words, extracts each of the searched search links (specific words) From the beginning of the first entry pointed to by the posting to the last posting of the discussion starting from that posting as a conversation topic, extract all the posts for that conversation topic.

The sentence analyzing unit 120 classifies sentences extracted from the sentence extracting unit 110 as questions and answers. If an error word occurs in a sentence extracted as text after speech recognition before classification, Restore the error-resolved sentences or filter inappropriate conversations to analyze and classify whether the sentence is a question or an answer. The sentence analysis unit 120 may be configured based on rules, machine learning, statistics, or a combination of at least one of rule based, machine learning based, and statistical based. When the sentence analyzing unit 120 is based on a machine learning system, learning is performed by learning a map, and the input sentence is analyzed to determine whether the input sentence is a question or an answer based on the language model data constructed as a result of the instruction.

The dialogue scenario generation unit 130 selects whether the question and the answer sentence analyzed through the sentence analysis unit 120 correspond to each other and connected to each other by emotion, intention, affirmation, and negation, and the selected question and answer sentence are connected to each other And the scenario is created by using the semantic vector data of the scenario sentence for the scenario generation and by the map learning based on the machine learning.

The dialogue scenario learning unit 140 changes the question and answer sentences of the scenario generated by the dialogue scenario generation unit 130 into a semantic vector and learns by map learning based on a machine learning basis.

Generally, supervised learning is machine learning that produces functions from training data. The training data consists of a pair of inputs (typically a vector) and a desired output. The output of the function may be a continuous value or it may predict the classification name of the input object. The job of learning a map is to predict the value of a function for a valid input target by reporting only a small number of training examples, the input pair and the target outputs. To do this, the learner has to generalize from the current data to the invisible situation in a reasonable way.

The dialogue scenario database 150 stores the dialogue scenario database as a semantic vector, and the scenario mean vector learned in the dialogue scenario learning unit 140 is stored. At this time, the dialogue scenario consecutively stores semantic vectors composed of sentence-level questions and answers. Sentences, and scenarios are stored as one or more semantic words to help human be intuitively grasped.

The ontology relationship mapping unit 160 performs a function of mapping and displaying the dialog scenario of the continuous semantic vector stored in the ontology dialog network when the dialog scenario database 150 storing the continuous semantic vector of the dialog scenario is constructed . The ontology dialogue network is a multidimensional space composed of 300 to 600 vectors, but since the multidimensional vectors can not be represented physically, they are compressed in three dimensions and displayed. The ontology conversation network expresses one point on the three-dimensional space as a node, and each node can be represented from a word semantic vector, a sentence semantic vector, and a scenario semantic vector stored in the database 150. Here, the word semantic vector and the sentence semantic vector use the previously disclosed word2vec and sent2vec machine learning algorithms, and the scenarios can be expressed in scenario2vec as the scenario semantic vector, and the scenario can be expressed as the semantic vector of the question and answer sentences. .

The ontology dialogue network can upload the continuous semantic vector of the dialogue scenario stored in the dialogue scenario database 150 via the ontology relationship mapping unit 160. The scenario itself can be made into a semantic vector and uploaded, When searching similar scenarios, it is possible to map the scenario closest to the input scenario by distance calculation on the semantic vector space. At this time, words, sentences, and scenarios can be expressed on the semantic vector space, but words, sentences, and scenarios each have a separate semantic space. In addition, words, sentences, and scenarios can be expressed in space while mapping the semantic space of words, sentences, and scenarios to a single semantic space. In the method of mapping, the sentence is expressed as the sum or product of word vectors, and the scenario is expressed as the sum or product of the sentence vectors. The semantic vector of the sentence scenario is used to help the intuitive understanding of the human being And is represented by one or more semantic words.

On the other hand, the dialog scenario stored in the dialog scenario database 150 can directly input the dialog scenario in the ontology dialogue network through the ontology relationship mapping unit 160, and the relationship of the sentences constituting the scenario is described in the following [Embodiment 1], as shown in Fig.

[Example 1]

Scenario 1 = (Question 1) - (Answer 1) - (Question 2) - (Answer 2) .... (Question N) - (Answer N)

Scenario 1 = (Where did you see it?) - (Are you curious about my name?) - (Not where I saw you) - (Where did we last meet? I think I saw it last time at the last Yaliki Seminar.)

Scenario 1 = (name + ambiguous) - (name + confirmation) - (place + ambiguous) - (place + confirmation) .... (alias + accept) - (alias + answer)

 Here, the semantic word " name ambiguity " is a new semantic vector that is a combination of the semantic word " name " and the semantic vector of the semantic word " ambiguity ", and is a semantic word representing a sentence such as "

One semantic word represents one or more sentences semantically, and it can also be called a representative word representing many sentences with the same meaning.

In addition, semantic vocabulary such as "semantic ambiguity" and "name resolution" represent the continuity of each question and answer constituting one scenario, such as "(Question 1)" and "(Answer 1)".

4 to 5 are views showing examples of sentence extraction during database construction of a dialogue scenario for the dialog system according to the present invention.

As shown in FIG. 4, the sentence extraction extracts a word from a DB pedia or WordNet, uses the word as a search word, searches for SNS posts and comment data Is parsed through the web. This is possible because when a post is posted on Twitter, the comment is similar to the format in which you are retweeting and exchanging conversations about a specific topic.

FIG. 5 retweets the article of the user who originally posted the article, and the other user continues to comment. At this time, the conversation development is similar to the tree form. Here, each tree is viewed as a dialog scenario, and a dialog scenario is expressed for each comment ID in order to facilitate understanding, as in the following [second embodiment].

[Example 2]

Scenario 1: ffebreze - hatter365 - fffebreze

Scenario 2: ffebreze - ffebreze - Teahya - ffebreze - Teahya

FIG. 6 is a screen showing an example of sentence analysis during the construction of a dialogue scenario database for the dialog system according to the present invention. As shown in FIG. 6, after parsing a posting, it is determined whether the corresponding posture corresponds to a question in the dialogue scenario none. Therefore, classify question and answer types by instructional learning of machine learning.

FIG. 7 is a view showing an example of scenario generation during construction of a dialogue scenario database for the dialog system according to the present invention, and FIG. 8 is a screen showing an example of dialogue scenario learning during dialogue scenario database construction for the dialogue system according to the present invention . As shown in FIG. 7, a question and an answer may not be displayed alternately for one subject, and a question and an answer may overlap (for example, question 1, answer 1, question 2, question 2, answer 2, question 3 , Answer 3, answer 3, answer 3, etc.) Among the questions and answers, questions and answers suitable for the scenario are selected, and the selected dialogue scenario as shown in FIG. 8 is displayed as one question and one answer It shows the iterative learning.

FIG. 9 is a screen showing an example in which a dialogue scenario database for an interactive system constructed according to the present invention is mapped to an ontology dialogue network.

The scenario is represented by a sequence of semantic vectors consisting of sentence-level questions and answers. Each question and answer that constitutes a scenario has a multidimensional semantic vector, and each question and answer sentence is automatically mapped to a dialogue intention (meaning word).

The dialogue intention (semantic word) expresses the currently used word and the semantic vector value of the word in a three-dimensional space in advance using word2vec or the like, (Meaning words) on a three-dimensional space.

When a large number of scenarios are input as a multidimensional semantic vector value, the distance between the input semantic vector value and the semantic vector value of the existing dialog intention (semantic word) is compared. When the input semantic vector value (Semantic word or semantic node or semantic cube) to the user. (The input semantic vector value is named as the name of the existing conversation)

When a large number of scenarios are inputted at the same time as the conversation intention (meaning word or semantic node or semantic cube) and dialogue sentence as well as the multidimensional semantic vector value, the space of each conversation intention (semantic word or semantic node or semantic cube) (Each question and answer). At this time, the semantic word may be automatically assigned to a semantic word by automatically analyzing the "conversational sentence" of the scenario, or a person may directly attach a semantic word suitable for the "conversational sentence".

[Example 3]

Scenario 1 = [Semantic word] [Conversation sentence] [Semantic vector], [Semantic word] [Conversation sentence] [Semantic vector] ....

Scenario 1 = [name ambiguity] [Where do you think] [2.382, 6.108, ...], [Name resolution] [Are you curious about Jay's name?] [8.730, 1,383, ....] ....

The method of automatically analyzing [semantic words] can be analyzed based on natural language processing methods, rule-based, statistical-based, and machine-based. In the case of machine learning based on the learning model, The intent of the conversation (intention word) is classified.

FIG. 10 shows an example of a dialog scenario stored in the dialogue scenario database according to the present invention. The dialogue scenario is usually displayed at 2 turn (relation of question 1 - answer 1 - question 2 - answer 2) The number of scenarios that can be displayed is also limited, and scenarios that are spatially close to the current scenario are also unknown. In addition, even if a scenario is input or modified, there is a problem that the correlation in the semantic space with other scenarios can not be known at all.

Dialogue scenario means and displayed as a continuous sequence of words, figure like the first scenario, the top, "the name is ambiguous", and certain characters in order to avoid duplication, such as "check C Name", "C yes", "KE name of IU" It can also be combined and displayed.

FIG. 11 is a screen in which a dialog scenario stored in the dialog scenario database according to the present invention is expressed by a three-dimensional scenario authoring tool, and the scenario can be displayed in a three-dimensional space as shown in FIG. Each question and answer in the scenario is represented by a semantic node (meaning cube or semantic word). Since semantic closeness between semantic nodes belonging to each scenario is displayed in the near space and displayed in the distant space, it is much easier to grasp the semantic relationship between the nodes constituting the scenario.

FIG. 12 is an example of a scenario input screen in which a dialog scenario stored in the dialog scenario database according to the present invention is expressed by a three-dimensional scenario authoring tool. As shown in FIG. 12, the scenario includes a right semantic word (semantic node or semantic cube) Dragging to the left margin space, one scenario is represented by a sequence of semantic words (eg, weekend schedule - schedule answer - hobby question - hobby answer). When the scenario input is completed, the scenario is automatically mapped in the three dimensional dialogue network . In addition, the right semantic word pattern has a multidimensional semantic vector value, and the semantic vector value is extracted by learning the combination of the semantic word (weekend + schedule) separately, or the semantic vector sum or product of the existing semantic word It is possible.

The right semantic word (eg, weekend schedule) represents various sentences, and it has various sentence structures which implies "weekend schedule".

Meaning word pattern Sentence example Weekend Schedule What is your weekend schedule? Please let me know your weekend schedule. I am curious about the weekend schedule. Tell me your weekend schedule. Do you have a weekend schedule?

FIG. 13 is an example of a scenario modification screen in which a dialog scenario stored in the dialog scenario database according to the present invention is expressed by a three-dimensional scenario authoring tool. The scenario may be searched by a search function, and a function of modifying a searched scenario is provided . When the scenario is modified, the modified scenario immediately reflects in the 3D space.

FIG. 14 is an example of a scenario deletion screen in which a dialog scenario stored in the dialog scenario database according to the present invention is represented by a three-dimensional scenario authoring tool. The scenario can be deleted, and the deleted scenario is completely deleted in the three- Can not.

FIG. 15 is a block diagram showing a sequence in which continuous conversation is performed using the conversation scenario automatic collecting and ontology conversation network according to the present invention. FIG. 16 is a block diagram showing a sequence diagram of automatic conversation scenario collection and on- As a sequence diagram.

As described above, the ontology dialogue network is a multidimensional vector space composed of 300 to 600 vectors. However, since the multidimensional vector can not be represented physically, spatial compression (using a space compression algorithm such as PCA) is displayed in three dimensions. The ontology dialogue network expresses one point on the 3D space as nodes, and each node can represent a word semantic vector, a sentence semantic vector, and a scenario semantic vector. The term 'ontology dialogue network' expresses the entire structure of such a node and the configuration of a scenario by connection between nodes. When 'ontology dialogue network database' is used, the storage unit storing such data and structure relation is expressed, Hereinafter, they will be used in a mixed manner with no significant discrimination.

Also, as described above, the semantic word is a word representing a plurality of sentences having a similar meaning, and may be a single word or a combination of several words. For example, the semantic word " name ambiguity " is a new meaning vector in which the semantic vector of the semantic word " name " is combined with the semantic vector of the semantic word " ambiguity ", and the semantic word representing " Same as. That is, the semantic word can be represented by a semantic vector in the ontology dialog network space, and has one node.

Scenarios can be displayed on a three-dimensional ontology dialogue network space, and each question and answer in a scenario is represented by a semantic node (meaning cube or semantic word). Since semantic closeness between semantic nodes belonging to each scenario is displayed in the near space and displayed in the distant space, it is much easier to grasp the semantic relationship between the nodes constituting the scenario. Hereinafter, the semantic node will also be referred to simply as a 'node'.

Hereinafter, a continuous conversation process using the ontology conversation network will be described with reference to the block diagram sequence of FIG. 15 and the flowchart of FIG.

First, when the user speaks, the user's voice is input to the continuous conversation system 200 (see FIG. 19) using the ontology conversation network of the present invention (S210). The continuous conversation system 200 using the ontology conversation network analyzes a sentence to understand a sentence of the input voice (S220). The continuous conversation system 200 using the ontology conversation network determines whether a semantic word that is a word or combination of words representing the conversation sentence is matched with the conversation sentence from the ontology conversation relationship network database S230).

If the semantic word is not matched (S240), a semantic word is determined for the conversation sentence, and the semantic word is matched to the conversation sentence and stored in the ontology conversation network database (S250), which will be described later in detail with reference to FIG.

A meaning word is not matched to the conversation sentence, and a new meaning word is determined and matched to the conversation sentence, or when a meaning word already matches the conversation sentence, a meaning word is matched to the conversation sentence (Hereinafter referred to as 'second semantic word') following the semantic word (hereinafter referred to as 'first semantic word') matched to the conversational sentence, in accordance with the conversation scenario in the ontology conversational network database, (S260), selects one of the conversation sentences included in the corresponding word (S270), and outputs the selected conversation sentence to the user (S280). For example, the first semantic word may be a semantic word representing a question conversation sentence that the current user has spoken, and the second semantic word may be a semantic word that the contiguous conversation system 200 using the ontology conversation network decided to answer Answer It can be a semantic word representing a conversation sentence.

There may be a case where there are a plurality of semantic words following the first semantic word matched to the conversation sentence according to the conversation scenario in the ontology dialog network database. In this case, it is possible to extract a semantic word having the highest weight given to the scenario for each semantic word as a second semantic word. For example, if the scenario following the first semantic word is the first semantic word -> A (meaning word) (weight 0.7), the first semantic word -> B (semantic word) A (meaning word) '(weighting 0.7) when there are three kinds of meaning words ->' C (meaning word) '(weighting 0.1) You can decide by word. Such a weight can be set in advance through conversation quality control for diagnosing and analyzing a conversation with a user, which will be described later with reference to FIG.

FIG. 17 is a flow chart illustrating a method of mapping a new conversation sentence into an ontology conversation network when a new conversation sentence is input, in a continuous conversation using the ontology conversation network according to the present invention.

When a new scenario is input, the semantic vector values of the question and answer sentences of the scenario are extracted, and the conversation intention of the question answering semantic vector is classified and classified into input scenarios ("conversation intention (semantic word) Vector ") to the conversation network.

That is, when a semantic word is not matched to a conversation sentence input by a user, that is, when a new scenario is input (S240), a semantic vector of the conversation sentence is derived (S251) (S252). ≪ / RTI > Methods for classifying semantic words can be based on one or more of rule based, statistical based, and machine learning based methods. Here, the machine learning corresponds to the 'conversation scenario learning' block in FIG.

If the semantic word corresponding to the conversation sentence is classified (S252), the semantic word is determined to be a semantic word of the conversation sentence, and is matched to the conversation sentence and stored in the ontology conversation network database (S253). At this time, a continuous list of semantic words, conversation sentences, and semantic vectors classified as described above can be stored in the ontology dialog network database.

If the semantic word corresponding to the conversation sentence is not classified (S252), a semantic word mapped to the node closest to the node indicated by the semantic vector is grasped from the ontology dialog network database, (S254). At this time, a continuous list of the classified semantic word, conversation sentence, and semantic vector can also be stored in the ontology dialog network database .

In this case, the semantic word mapped to the node closest to the node indicated by the semantic vector may be a single semantic word, or may be a new semantic word indicated by a semantic vector generated by a combination of semantic words.

A scenario can be expressed as a series of pairs of question answers. Just as each question answer forms a single scenario, a scenario can contain specific dialogue information (education, literature, common sense, everyday conversation, sports, movies, etc.) And scenarios can be represented by a combination of meaning words.

Scenario classifications can be categorized using one or more of the following: rule-based, statistical-based, and machine-based. Therefore, before classifying the conversation intention of each sentence, the scenario classification may be firstly selected and the search intention (semantic word) belonging to the scenario classification may be searched first to reduce the search time.

18 is a block diagram showing a sequence for implementing conversation quality improvement using continuous conversation using the ontology conversation network according to the present invention.

As the user interacts with the system, the user can immediately know whether the user's response is positive or negative. The user's reaction such as positive reaction, negative reaction, neglect reaction, topic change reaction can be analyzed by analyzing the user's conversation sentence, and various analysis methods such as rule based, statistical based, and machine learning based can be used.

When the user response is positive, the current scenario is recognized as the preferred dialog scenario, and the likelihood that the current scenario will be adopted is increased by weight adjustment, so that the preferred scenario is continuously selected, and the non-preferred scenario has priority , And managing the dialogue scenario continuously through the real-time user response and ultimately aiming at improving the conversation quality as the continuous conversation is performed.

This process will be described with reference to FIG. 18. Continuous conversation quality management is performed through a conversation scenario history between the user and the continuous conversation system 200 using the ontology conversation network of the present invention. In other words, the sentence inputted from the user is analyzed and the reaction of the user is analyzed from the analyzed conversation sentence. If the user reaction is analyzed to be positive, the likelihood that the dialog scenario leading to the corresponding sentence sentence or the semantic word of the corresponding sentence is selected is set to be high, and if the user reaction is analyzed as negative, the corresponding sentence or semantic word Is set to a low possibility.

Positive or negative here is a very generic term. As shown in the example of Fig. 18, it can be divided into details such as praise reaction, positive reaction, negative reaction, neglect reaction, topic change reaction, etc. However, it can be explained as a positive or negative reaction including all of these. This would be a more accurate expression of giving a degree or degree, rather than analyzing it with two kinds of reactions. In other words, the more responsive and favorable the user's response to the sentence analysis is, the more likely it is that the dialogue scenario with such response will be selected in the future through feedback, and the more the opposing dialogue scenario The possibility of being selected is lowered.

The setting of the likelihood that a future dialog scenario will be selected may be made by assigning a weight in relation to the preference / non-preference dialog scenario according to predetermined criteria in the dialog scenario in which the user response is made. This increases the probability that a conversation scenario with a higher weight will be selected in the future.

At this time, the analysis of the user reaction may be performed by analyzing the user's conversation sentence according to a rule (rule-based), performing based on the sentence statistics (statistical based), or machine learning Machine learning basis).

Alternatively, the analysis of the user response may be performed by a user preference in which the conversation sentence of the user in the ontology dialog network database is analyzed and mapped to the corresponding semantic word. That is, a semantic word space and a user preference space may be separately provided on the ontology dialog network database, and a specific semantic word node of each semantic word space may be mapped to a user preference of the same node in the user preference space. That is, since the user preference analyzed by the user reaction is mapped to the semantic word of the user, the weight for the conversation can be set therefrom.

19 is a diagram illustrating a configuration of a continuous conversation system using an ontology conversation network according to the present invention.

The control unit 201 controls each of the components of the continuous dialog system using the ontology dialog network to perform a series of processes related to the continuous dialog using the ontology dialog network.

The voice input unit 202 receives the voice of the user.

The sentence analysis unit 203 analyzes sentences from the input speech.

The dialog management unit 204 receives a semantic word (hereinafter referred to as a " first semantic word ") following a semantic word (hereinafter, referred to as 'first semantic word'), which is a combination of words or words representing the conversation sentence, Quot; second semantic word "), and selects one of the conversation sentences included in the corresponding semantic word as an output conversation sentence. In addition, in the case where there are a plurality of semantic words following the first semantic word matched to the conversation sentence in accordance with the conversation scenario in the ontology conversation relationship network database, And extracts the semantic word having the highest weight as the second semantic word.

If the semantic word is not matched with the analyzed conversation sentence, the new scenario processing unit 206 determines a new semantic word for the sentence, stores the new semantic word in the ontology dialog network database, matches the sentence with the sentence, And transmits it to the dialog management unit.

The new scenario processing unit can derive the semantic vector of the corresponding sentence sentence if the sentence sentence is not matched with the semantic word. And classifies the semantic word corresponding to the conversation sentence from the derived semantic vector. If the semantic word corresponding to the conversation sentence is classified, the semantic word is determined to be a semantic word of the conversation sentence, and the semantic word is matched to the conversation sentence and stored in the ontology conversation network database, The semantic vector mapped to the node closest to the node indicated by the semantic vector is grasped from the ontology dialog network database, and the identified semantic word is determined to be a semantic word of the conversation sentence and is matched to the sentence sentence Store it in the ontology dialog network database, and deliver the determined semantic word to the conversation management unit.

As described above with reference to FIG. 17, the semantic word mapped to the node closest to the node indicated by the semantic vector may be a single semantic word or a semantic word indicated by a semantic vector generated by a combination of semantic words.

The dialog quality management unit 205 analyzes the response of the user from the analyzed conversation sentence to determine the possibility that the conversation scenario leading to the corresponding sentence sentence or the semantic word of the corresponding sentence is selected if the user response is analyzed as being positive And when the user reaction is analyzed as negative, the conversation quality management is performed by setting a possibility that the conversation scenario leading to the conversation sentence or the meaning word of the conversation sentence is selected to be low. At this time, as described above with reference to FIG. 18, the setting of the possibility that the conversation scenario is selected may be made by weighting according to predetermined criteria. Also, analysis of the user response may be performed by analyzing the conversation sentence according to a rule (rule-based), based on statistics of a conversation sentence (statistical based), or by machine learning Based). Analysis of the user response may be performed by the user preference that is analyzed and mapped on the corresponding semantic word of the conversation sentence of the user in the ontology conversation network database. This has been described in detail with reference to FIG. 18 .

The dialogue output unit 207 outputs the dialogue sentence selected by the dialogue management unit to the user through voice or the like.

20 is a diagram showing an embodiment of the ontology talk network structure according to the present invention.

That is, referring to the embodiment of FIG. 20, the conversation of the ontology conversation network can be classified into a general conversation domain, a professional conversation domain, or another domain by setting another new domain as necessary. That is, a large number of nodes such as semantic words mapped in the ontology dialog network space may already be set in areas such as general conversation and professional conversation, respectively. That is, when a semantic vector for each sentence is obtained, the node indicated by each semantic vector according to the meaning of the sentence may be configured to point to an area of the classification determined on the ontology dialog network space. Of course, the same method can be used for detailed classification at the bottom of a general conversation or a professional conversation.

General conversations (daily conversations, common sense conversations, topic conversations, etc.) develop conversations based on conversation scenarios in a conversation network for various topics.

Specialized conversations (advance purchase, reservation, purchase, etc.) acquire specific information (Argument) according to a specific purpose of the strategy, and perform dialogue for the performance of the task. For example, in the case of KTX reservation, the user is asked about the information necessary for the reservation such as the reservation time, the number of the companion, the type of the seat, and the reservation is finished after the information is satisfied. Therefore, it is necessary to follow a dialogue procedure (Flow) for a specific purpose.

In this way, general conversation and professional conversation can be freely performed in the ontology conversation network.

21 to 24 show examples of conversations classified in the ontology dialog network as described above.

21 is a diagram illustrating an example of a general conversation classification structure in an ontology conversation network according to the present invention.

General conversations can include daily conversations, common conversations, topic conversations, emotional conversations, topic conversations, and conversations are possible by constantly moving conversation focus on various topics.

22 is a diagram illustrating an embodiment of a professional talk classification structure in an ontology talk network according to the present invention.

Professional conversation refers to conducting a specific purpose-based conversation, such as booking, reserving, or purchasing.

23 is a diagram illustrating an example of a conversation classification structure exchanged with an agent in a hospital call center during a professional conversation in an ontology conversation network according to the present invention.

24 is a view showing an embodiment of a classification structure in which a daily conversation, an emotional conversation, and a professional conversation are connected in an ontology conversation network according to the present invention.

Continuous conversation can be done while freely coming in everyday conversation, emotional conversation, and professional conversation.

While the present invention has been particularly shown and described with reference to exemplary embodiments thereof, it is to be understood that the invention is not limited to the disclosed exemplary embodiments. It will be understood that various modifications and changes may be made without departing from the scope of the appended claims.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. Prior to this, terms and words used in the present specification and claims should not be construed as limited to ordinary or dictionary terms, and the inventor should appropriately interpret the concepts of the terms appropriately It should be interpreted in accordance with the meaning and concept consistent with the technical idea of the present invention based on the principle that it can be defined. Therefore, the embodiments described in this specification and the configurations shown in the drawings are merely the most preferred embodiments of the present invention and do not represent all the technical ideas of the present invention. Therefore, It is to be understood that equivalents and modifications are possible.
1 is a flowchart illustrating a method for constructing a dialogue scenario database for a dialog system according to the present invention.
In the dialogue scenario database building method for the dialog system according to FIG. 1, a sentence is extracted first (S110). Extraction of sentences is extracted from conversational voice files or conversational posts.
If a sentence is extracted in step S110, the extracted sentence is classified into a question and an answer (S120).
Then, it is determined whether the question and answer sentences classified in step S120 correspond to each other and connected to each other by the sentiment, intention, affirmation, and negation (S130), and the selected question and answer sentence is generated as a list-type scenario ).
Subsequently, the scenario and the question and answer sentences of the generated scenario, which are the list forms, are converted into semantic vectors to learn a scenario (S150), and the learned scenario is converted into a database with a continuous scenario meaning vector (S160).
The consecutive semantic vectors of the database-dated scenario are displayed in the ontology multi-dimensional space as one of semantic words, meaning nodes, and semantic cubes (S160).
The method for constructing a dialog scenario database for the dialog system of the present invention shown in FIG. 1 will be described with reference to a dialogue scenario database establishing apparatus for the dialog system of FIGS.
FIG. 2 is a view schematically showing an apparatus for constructing a dialogue scenario database for an interactive system according to the present invention, and FIG. 3 is a detailed view illustrating an apparatus for constructing a dialogue scenario database for the interactive system according to FIG.
2 to 3, an apparatus 100 for constructing a dialogue scenario database for an interactive system according to the present invention includes a sentence extraction unit 110 for extracting a sentence, a sentence extraction unit 110 for extracting a sentence extracted by the sentence extraction unit 110, A dialogue scenario generation unit 130 for generating a scenario in which sentences classified by the sentence analysis unit 120 correspond to each other and conversations are connected to each other, A dialogue scenario learning unit 140 for learning the sentence of the dialogue scenario generated by the dialogue scenario generation unit 130, a dialogue scenario storage unit 140 for storing a dialogue scenario, which is learned by the dialogue scenario learning unit 140, The scenario database 150 and the continuous semantic vectors of the dialog scenario stored in the dialog scenario database 150 are stored in the ontology multidimensional space with semantic words or semantic nodes Hitting include ontology mapping relationship unit 160 to display one of the mean cube.
The sentence extracting unit 110 extracts a sentence from an interactive voice file or a conversation type of text. The sentence extracting unit 110 includes a voice file extracting module 111 for extracting an interactive voice file, 3, an interactive voice file, such as a call center, a radio, and a TV broadcast, is sent to a voice file extraction module 111 through a voice file extraction module 111. The voice file extraction module 111 extracts a sentence When a search word is selected in the search word generator 113, a sentence is extracted because the search is performed by the postexpression module 112 and the posting is parsed. For example, it is possible to extract a sentence from a text after speech recognition of an existing voice recording file provided by a call center, or extract a sentence from text after speech recognition of a consultation content between a customer and an agent in a call center in real time. On the other hand, SNS extracts sentences from posts that talk about various subjects on Twitter or Facebook, searches for SNS with the prepared search words, extracts each of the searched search links (specific words) From the beginning of the first entry pointed to by the posting to the last posting of the discussion starting from that posting as a conversation topic, extract all the posts for that conversation topic.
The sentence analyzing unit 120 classifies sentences extracted from the sentence extracting unit 110 as questions and answers. If an error word occurs in a sentence extracted as text after speech recognition before classification, Restore the error-resolved sentences or filter inappropriate conversations to analyze and classify whether the sentence is a question or an answer. The sentence analysis unit 120 may be configured based on rules, machine learning, statistics, or a combination of at least one of rule based, machine learning based, and statistical based. When the sentence analyzing unit 120 is based on a machine learning system, learning is performed by learning a map, and the input sentence is analyzed to determine whether the input sentence is a question or an answer based on the language model data constructed as a result of the instruction.
The dialogue scenario generation unit 130 selects whether the question and the answer sentence analyzed through the sentence analysis unit 120 correspond to each other and connected to each other by emotion, intention, affirmation, and negation, and the selected question and answer sentence are connected to each other And the scenario is created by using the semantic vector data of the scenario sentence for the scenario generation and by the map learning based on the machine learning.
The dialogue scenario learning unit 140 changes the question and answer sentences of the scenario generated by the dialogue scenario generation unit 130 into a semantic vector and learns by map learning based on a machine learning basis.
Generally, supervised learning is machine learning that produces functions from training data. The training data consists of a pair of inputs (typically a vector) and a desired output. The output of the function may be a continuous value or it may predict the classification name of the input object. The job of learning a map is to predict the value of a function for a valid input target by reporting only a small number of training examples, the input pair and the target outputs. To do this, the learner has to generalize from the current data to the invisible situation in a reasonable way.
The dialogue scenario database 150 stores the dialogue scenario database as a semantic vector, and the scenario mean vector learned in the dialogue scenario learning unit 140 is stored. At this time, the dialogue scenario consecutively stores semantic vectors composed of sentence-level questions and answers. Sentences, and scenarios are stored as one or more semantic words to help human be intuitively grasped.
The ontology relationship mapping unit 160 performs a function of mapping and displaying the dialog scenario of the continuous semantic vector stored in the ontology dialog network when the dialog scenario database 150 storing the continuous semantic vector of the dialog scenario is constructed . The ontology dialogue network is a multidimensional space composed of 300 to 600 vectors, but since the multidimensional vectors can not be represented physically, they are compressed in three dimensions and displayed. The ontology conversation network expresses one point on the three-dimensional space as a node, and each node can be represented from a word semantic vector, a sentence semantic vector, and a scenario semantic vector stored in the database 150. Here, the word semantic vector and the sentence semantic vector use the previously disclosed word2vec and sent2vec machine learning algorithms, and the scenarios can be expressed in scenario2vec as the scenario semantic vector, and the scenario can be expressed as the semantic vector of the question and answer sentences. .
The ontology dialogue network can upload the continuous semantic vector of the dialogue scenario stored in the dialogue scenario database 150 via the ontology relationship mapping unit 160. The scenario itself can be made into a semantic vector and uploaded, When searching similar scenarios, it is possible to map the scenario closest to the input scenario by distance calculation on the semantic vector space. At this time, words, sentences, and scenarios can be expressed on the semantic vector space, but words, sentences, and scenarios each have a separate semantic space. In addition, words, sentences, and scenarios can be expressed in space while mapping the semantic space of words, sentences, and scenarios to a single semantic space. In the method of mapping, the sentence is expressed as the sum or product of word vectors, and the scenario is expressed as the sum or product of the sentence vectors. The semantic vector of the sentence scenario is used to help the intuitive understanding of the human being And is represented by one or more semantic words.
On the other hand, the dialog scenario stored in the dialog scenario database 150 can directly input the dialog scenario in the ontology dialogue network through the ontology relationship mapping unit 160, and the relationship of the sentences constituting the scenario is described in the following [Embodiment 1], as shown in Fig.
[Example 1]
Scenario 1 = (Question 1) - (Answer 1) - (Question 2) - (Answer 2) .... (Question N) - (Answer N)
Scenario 1 = (Where did you see it?) - (Are you curious about my name?) - (Not where I saw you) - (Where did we last meet? I think I saw it last time at the last Yaliki Seminar.)
Scenario 1 = (name + ambiguous) - (name + confirmation) - (place + ambiguous) - (place + confirmation) .... (alias + accept) - (alias + answer)
Here, the semantic word " name ambiguity " is a new semantic vector that is a combination of the semantic word " name " and the semantic vector of the semantic word " ambiguity ", and is a semantic word representing a sentence such as "
One semantic word represents one or more sentences semantically, and it can also be called a representative word representing many sentences with the same meaning.
In addition, semantic vocabulary such as "semantic ambiguity" and "name resolution" represent the continuity of each question and answer constituting one scenario, such as "(Question 1)" and "(Answer 1)".
4 to 5 are views showing examples of sentence extraction during database construction of a dialogue scenario for the dialog system according to the present invention.
As shown in FIG. 4, the sentence extraction extracts a word from a DB pedia or WordNet, uses the word as a search word, searches for SNS posts and comment data Is parsed through the web. This is possible because when a post is posted on Twitter, the comment is similar to the format in which you are retweeting and exchanging conversations about a specific topic.
FIG. 5 retweets the article of the user who originally posted the article, and the other user continues to comment. At this time, the conversation development is similar to the tree form. Here, each tree is viewed as a dialog scenario, and a dialog scenario is expressed for each comment ID in order to facilitate understanding, as in the following [second embodiment].
[Example 2]
Scenario 1: ffebreze - hatter365 - fffebreze
Scenario 2: ffebreze - ffebreze - Teahya - ffebreze - Teahya
FIG. 6 is a screen showing an example of sentence analysis during the construction of a dialogue scenario database for the dialog system according to the present invention. As shown in FIG. 6, after parsing a posting, it is determined whether the corresponding posture corresponds to a question in the dialogue scenario none. Therefore, classify question and answer types by instructional learning of machine learning.
FIG. 7 is a view showing an example of scenario generation during construction of a dialogue scenario database for the dialog system according to the present invention, and FIG. 8 is a screen showing an example of dialogue scenario learning during dialogue scenario database construction for the dialogue system according to the present invention . As shown in FIG. 7, a question and an answer may not be displayed alternately for one subject, and a question and an answer may overlap (for example, question 1, answer 1, question 2, question 2, answer 2, question 3 , Answer 3, answer 3, answer 3, etc.) Among the questions and answers, questions and answers suitable for the scenario are selected, and the selected dialogue scenario as shown in FIG. 8 is displayed as one question and one answer It shows the iterative learning.
FIG. 9 is a screen showing an example in which a dialogue scenario database for an interactive system constructed according to the present invention is mapped to an ontology dialogue network.
The scenario is represented by a sequence of semantic vectors consisting of sentence-level questions and answers. Each question and answer that constitutes a scenario has a multidimensional semantic vector, and each question and answer sentence is automatically mapped to a dialogue intention (meaning word).
The dialogue intention (semantic word) expresses the currently used word and the semantic vector value of the word in a three-dimensional space in advance using word2vec or the like, (Meaning words) on a three-dimensional space.
When a large number of scenarios are input as a multidimensional semantic vector value, the distance between the input semantic vector value and the semantic vector value of the existing dialog intention (semantic word) is compared. When the input semantic vector value (Semantic word or semantic node or semantic cube) to the user. (The input semantic vector value is named as the name of the existing conversation)
When a large number of scenarios are inputted at the same time as the conversation intention (meaning word or semantic node or semantic cube) and dialogue sentence as well as the multidimensional semantic vector value, the space of each conversation intention (semantic word or semantic node or semantic cube) (Each question and answer). At this time, the semantic word may be automatically assigned to a semantic word by automatically analyzing the "conversational sentence" of the scenario, or a person may directly attach a semantic word suitable for the "conversational sentence".
[Example 3]
Scenario 1 = [Semantic word] [Conversation sentence] [Semantic vector], [Semantic word] [Conversation sentence] [Semantic vector] ....
Scenario 1 = [name ambiguity] [Where do you think] [2.382, 6.108, ...], [Name resolution] [Are you curious about Jay's name?] [8.730, 1,383, ....] ....
The method of automatically analyzing [semantic words] can be analyzed based on natural language processing methods, rule-based, statistical-based, and machine-based. In the case of machine learning based on the learning model, The intent of the conversation (intention word) is classified.
FIG. 10 shows an example of a dialog scenario stored in the dialogue scenario database according to the present invention. The dialogue scenario is usually displayed at 2 turn (relation of question 1 - answer 1 - question 2 - answer 2) The number of scenarios that can be displayed is also limited, and scenarios that are spatially close to the current scenario are also unknown. In addition, even if a scenario is input or modified, there is a problem that the correlation in the semantic space with other scenarios can not be known at all.
Dialogue scenario means and displayed as a continuous sequence of words, figure like the first scenario, the top, "the name is ambiguous", and certain characters in order to avoid duplication, such as "check C Name", "C yes", "KE name of IU" It can also be combined and displayed.
FIG. 11 is a screen in which a dialog scenario stored in the dialog scenario database according to the present invention is expressed by a three-dimensional scenario authoring tool, and the scenario can be displayed in a three-dimensional space as shown in FIG. Each question and answer in the scenario is represented by a semantic node (meaning cube or semantic word). Since semantic closeness between semantic nodes belonging to each scenario is displayed in the near space and displayed in the distant space, it is much easier to grasp the semantic relationship between the nodes constituting the scenario.
FIG. 12 is an example of a scenario input screen in which a dialog scenario stored in the dialog scenario database according to the present invention is expressed by a three-dimensional scenario authoring tool. As shown in FIG. 12, the scenario includes a right semantic word (semantic node or semantic cube) Dragging to the left margin space, one scenario is represented by a sequence of semantic words (eg, weekend schedule - schedule answer - hobby question - hobby answer). When the scenario input is completed, the scenario is automatically mapped in the three dimensional dialogue network . In addition, the right semantic word pattern has a multidimensional semantic vector value, and the semantic vector value is extracted by learning the combination of the semantic word (weekend + schedule) separately, or the semantic vector sum or product of the existing semantic word It is possible.
The right semantic word (for example, a weekend schedule) represents various sentences and has various sentence structures such as the following, which implies a "weekend schedule".
FIG. 13 is an example of a scenario modification screen in which a dialog scenario stored in the dialog scenario database according to the present invention is expressed by a three-dimensional scenario authoring tool. The scenario may be searched by a search function, and a function of modifying a searched scenario is provided . When the scenario is modified, the modified scenario immediately reflects in the 3D space.
FIG. 14 is an example of a scenario deletion screen in which a dialog scenario stored in the dialog scenario database according to the present invention is represented by a three-dimensional scenario authoring tool. The scenario can be deleted, and the deleted scenario is completely deleted in the three- Can not.
FIG. 15 is a block diagram showing a sequence in which continuous conversation is performed using the conversation scenario automatic collecting and ontology conversation network according to the present invention. FIG. 16 is a block diagram showing a sequence diagram of automatic conversation scenario collection and on- As a sequence diagram.
As described above, the ontology dialogue network is a multidimensional vector space composed of 300 to 600 vectors. However, since the multidimensional vector can not be represented physically, spatial compression (using a space compression algorithm such as PCA) is displayed in three dimensions. The ontology dialogue network expresses one point on the 3D space as nodes, and each node can represent a word semantic vector, a sentence semantic vector, and a scenario semantic vector. The term 'ontology dialogue network' expresses the entire structure of such a node and the configuration of a scenario by connection between nodes. When 'ontology dialogue network database' is used, the storage unit storing such data and structure relation is expressed, Hereinafter, they will be used in a mixed manner with no significant discrimination.
Also, as described above, the semantic word is a word representing a plurality of sentences having a similar meaning, and may be a single word or a combination of several words. For example, the semantic word " name ambiguity " is a new meaning vector in which the semantic vector of the semantic word " name " is combined with the semantic vector of the semantic word " ambiguity ", and the semantic word representing " Same as. That is, the semantic word can be represented by a semantic vector in the ontology dialog network space, and has one node.
Scenarios can be displayed on a three-dimensional ontology dialogue network space, and each question and answer in a scenario is represented by a semantic node (meaning cube or semantic word). Since semantic closeness between semantic nodes belonging to each scenario is displayed in the near space and displayed in the distant space, it is much easier to grasp the semantic relationship between the nodes constituting the scenario. Hereinafter, the semantic node will also be referred to simply as a 'node'.
Hereinafter, a continuous conversation process using the ontology conversation network will be described with reference to the block diagram sequence of FIG. 15 and the flowchart of FIG.
First, when the user speaks, the user's voice is input to the continuous conversation system 200 (see FIG. 19) using the ontology conversation network of the present invention (S210). The continuous conversation system 200 using the ontology conversation network analyzes a sentence to understand a sentence of the input voice (S220). The continuous conversation system 200 using the ontology conversation network determines whether a semantic word that is a word or combination of words representing the conversation sentence is matched with the conversation sentence from the ontology conversation relationship network database S230).
If the semantic word is not matched (S240), a semantic word is determined for the conversation sentence, and the semantic word is matched to the conversation sentence and stored in the ontology conversation network database (S250), which will be described later in detail with reference to FIG.
A meaning word is not matched to the conversation sentence, and a new meaning word is determined and matched to the conversation sentence, or when a meaning word already matches the conversation sentence, a meaning word is matched to the conversation sentence (Hereinafter referred to as 'second semantic word') following the semantic word (hereinafter referred to as 'first semantic word') matched to the conversational sentence, in accordance with the conversation scenario in the ontology conversational network database, (S260), selects one of the conversation sentences included in the corresponding word (S270), and outputs the selected conversation sentence to the user (S280). For example, the first semantic word may be a semantic word representing a question conversation sentence that the current user has spoken, and the second semantic word may be a semantic word that the contiguous conversation system 200 using the ontology conversation network decided to answer Answer It can be a semantic word representing a conversation sentence.
There may be a case where there are a plurality of semantic words following the first semantic word matched to the conversation sentence according to the conversation scenario in the ontology dialog network database. In this case, it is possible to extract a semantic word having the highest weight given to the scenario for each semantic word as a second semantic word. For example, if the scenario following the first semantic word is the first semantic word -> A (meaning word) (weight 0.7), the first semantic word -> B (semantic word) A (meaning word) '(weighting 0.7) when there are three kinds of meaning words ->' C (meaning word) '(weighting 0.1) You can decide by word. Such a weight can be set in advance through conversation quality control for diagnosing and analyzing a conversation with a user, which will be described later with reference to FIG.
FIG. 17 is a flow chart illustrating a method of mapping a new conversation sentence into an ontology conversation network when a new conversation sentence is input, in a continuous conversation using the ontology conversation network according to the present invention.
When a new scenario is input, the semantic vector values of the question and answer sentences of the scenario are extracted, and the conversation intention of the question answering semantic vector is classified and classified into input scenarios ("conversation intention (semantic word) Vector ") to the conversation network.
That is, when a semantic word is not matched to a conversation sentence input by a user, that is, when a new scenario is input (S240), a semantic vector of the conversation sentence is derived (S251) (S252). ≪ / RTI > Methods for classifying semantic words can be based on one or more of rule based, statistical based, and machine learning based methods. Here, the machine learning corresponds to the 'conversation scenario learning' block in FIG.
If the semantic word corresponding to the conversation sentence is classified (S252), the semantic word is determined to be a semantic word of the conversation sentence, and is matched to the conversation sentence and stored in the ontology conversation network database (S253). At this time, a continuous list of semantic words, conversation sentences, and semantic vectors classified as described above can be stored in the ontology dialog network database.
If the semantic word corresponding to the conversation sentence is not classified (S252), a semantic word mapped to the node closest to the node indicated by the semantic vector is grasped from the ontology dialog network database, (S254). At this time, a continuous list of the classified semantic word, conversation sentence, and semantic vector can also be stored in the ontology dialog network database .
In this case, the semantic word mapped to the node closest to the node indicated by the semantic vector may be a single semantic word, or may be a new semantic word indicated by a semantic vector generated by a combination of semantic words.
A scenario can be expressed as a series of pairs of question answers. Just as each question answer forms a single scenario, a scenario can contain specific dialogue information (education, literature, common sense, everyday conversation, sports, movies, etc.) And scenarios can be represented by a combination of meaning words.
Scenario classifications can be categorized using one or more of the following: rule-based, statistical-based, and machine-based. Therefore, before classifying the conversation intention of each sentence, the scenario classification may be firstly selected and the search intention (semantic word) belonging to the scenario classification may be searched first to reduce the search time.
18 is a block diagram showing a sequence for implementing conversation quality improvement using continuous conversation using the ontology conversation network according to the present invention.
As the user interacts with the system, the user can immediately know whether the user's response is positive or negative. The user's reaction such as positive reaction, negative reaction, neglect reaction, topic change reaction can be analyzed by analyzing the user's conversation sentence, and various analysis methods such as rule based, statistical based, and machine learning based can be used.
When the user response is positive, the current scenario is recognized as the preferred dialog scenario, and the likelihood that the current scenario will be adopted is increased by weight adjustment, so that the preferred scenario is continuously selected, and the non-preferred scenario has priority , And managing the dialogue scenario continuously through the real-time user response and ultimately aiming at improving the conversation quality as the continuous conversation is performed.
This process will be described with reference to FIG. 18. Continuous conversation quality management is performed through a conversation scenario history between the user and the continuous conversation system 200 using the ontology conversation network of the present invention. In other words, the sentence inputted from the user is analyzed and the reaction of the user is analyzed from the analyzed conversation sentence. If the user reaction is analyzed to be positive, the likelihood that the dialog scenario leading to the corresponding sentence sentence or the semantic word of the corresponding sentence is selected is set to be high, and if the user reaction is analyzed as negative, the corresponding sentence or semantic word Is set to a low possibility.
Positive or negative here is a very generic term. As shown in the example of Fig. 18, it can be divided into details such as praise reaction, positive reaction, negative reaction, neglect reaction, topic change reaction, etc. However, it can be explained as a positive or negative reaction including all of them. This would be a more accurate expression of giving a degree or degree, rather than analyzing it with two kinds of reactions. In other words, the more responsive and favorable the user's response to the sentence analysis is, the more likely it is that the dialogue scenario with such response will be selected in the future through feedback, and the more the opposing dialogue scenario The possibility of being selected is lowered.
The setting of the likelihood that a future dialog scenario will be selected may be made by assigning a weight in relation to the preference / non-preference dialog scenario according to predetermined criteria in the dialog scenario in which the user response is made. This increases the probability that a conversation scenario with a higher weight will be selected in the future.
At this time, the analysis of the user reaction may be performed by analyzing the user's conversation sentence according to a rule (rule-based), performing based on the sentence statistics (statistical based), or machine learning Machine learning basis).
Alternatively, the analysis of the user response may be performed by a user preference in which the conversation sentence of the user in the ontology dialog network database is analyzed and mapped to the corresponding semantic word. That is, a semantic word space and a user preference space may be separately provided on the ontology dialog network database, and a specific semantic word node of each semantic word space may be mapped to a user preference of the same node in the user preference space. That is, since the user preference analyzed by the user reaction is mapped to the semantic word of the user, the weight for the conversation can be set therefrom.
19 is a diagram illustrating a configuration of a continuous conversation system using an ontology conversation network according to the present invention.
The control unit 201 controls each of the components of the continuous dialog system using the ontology dialog network to perform a series of processes related to the continuous dialog using the ontology dialog network.
The voice input unit 202 receives the voice of the user.
The sentence analysis unit 203 analyzes sentences from the input speech.
The dialog management unit 204 receives a semantic word (hereinafter referred to as a " first semantic word ") following a semantic word (hereinafter, referred to as 'first semantic word'), which is a combination of words or words representing the conversation sentence, Quot; second semantic word "), and selects one of the conversation sentences included in the corresponding semantic word as an output conversation sentence. In addition, in the case where there are a plurality of semantic words following the first semantic word matched to the conversation sentence in accordance with the conversation scenario in the ontology conversation relationship network database, And extracts the semantic word having the highest weight as the second semantic word.
If the semantic word is not matched with the analyzed conversation sentence, the new scenario processing unit 206 determines a new semantic word for the sentence, stores the new semantic word in the ontology dialog network database, matches the sentence with the sentence, And transmits it to the dialog management unit.
The new scenario processing unit can derive the semantic vector of the corresponding sentence sentence if the sentence sentence is not matched with the semantic word. And classifies the semantic word corresponding to the conversation sentence from the derived semantic vector. If the semantic word corresponding to the conversation sentence is classified, the semantic word is determined to be a semantic word of the conversation sentence, and the semantic word is matched to the conversation sentence and stored in the ontology conversation network database, The semantic vector mapped to the node closest to the node indicated by the semantic vector is grasped from the ontology dialog network database, and the identified semantic word is determined to be a semantic word of the conversation sentence and is matched to the sentence sentence Store it in the ontology dialog network database, and deliver the determined semantic word to the conversation management unit.
As described above with reference to FIG. 17, the semantic word mapped to the node closest to the node indicated by the semantic vector may be a single semantic word or a semantic word indicated by a semantic vector generated by a combination of semantic words.
The dialog quality management unit 205 analyzes the response of the user from the analyzed conversation sentence to determine the possibility that the conversation scenario leading to the corresponding sentence sentence or the semantic word of the corresponding sentence is selected if the user response is analyzed as being positive And when the user reaction is analyzed as negative, the conversation quality management is performed by setting a possibility that the conversation scenario leading to the conversation sentence or the meaning word of the conversation sentence is selected to be low. At this time, as described above with reference to FIG. 18, the setting of the possibility that the conversation scenario is selected may be made by weighting according to predetermined criteria. Also, analysis of the user response may be performed by analyzing the conversation sentence according to a rule (rule-based), based on statistics of a conversation sentence (statistical based), or by machine learning Based). Analysis of the user response may be performed by the user preference that is analyzed and mapped on the corresponding semantic word of the conversation sentence of the user in the ontology conversation network database. This has been described in detail with reference to FIG. 18 .
The dialogue output unit 207 outputs the dialogue sentence selected by the dialogue management unit to the user through voice or the like.
20 is a diagram showing an embodiment of the ontology talk network structure according to the present invention.
That is, referring to the embodiment of FIG. 20, the conversation of the ontology conversation network can be classified into a general conversation domain, a professional conversation domain, or another domain by setting another new domain as necessary. That is, a large number of nodes such as semantic words mapped in the ontology dialog network space may already be set in areas such as general conversation and professional conversation, respectively. That is, when a semantic vector for each sentence is obtained, the node indicated by each semantic vector according to the meaning of the sentence may be configured to point to an area of the classification determined on the ontology dialog network space. Of course, the same method can be used for detailed classification at the bottom of a general conversation or a professional conversation.
General conversations (daily conversations, common sense conversations, topic conversations, etc.) develop conversations based on conversation scenarios in a conversation network for various topics.
Specialized conversations (advance purchase, reservation, purchase, etc.) acquire specific information (Argument) according to a specific purpose of the strategy, and perform dialogue for the performance of the task. For example, in the case of KTX reservation, the user is asked about the information necessary for the reservation such as the reservation time, the number of the companion, the type of the seat, and the reservation is finished after the information is satisfied. Therefore, it is necessary to follow a dialogue procedure (Flow) for a specific purpose.
In this way, general conversation and professional conversation can be freely performed in the ontology conversation network.
21 to 24 show examples of conversations classified in the ontology dialog network as described above.
21 is a diagram illustrating an example of a general conversation classification structure in an ontology conversation network according to the present invention.
General conversations can include daily conversations, common conversations, topic conversations, emotional conversations, topic conversations, and conversations are possible by constantly moving conversation focus on various topics.
22 is a diagram illustrating an embodiment of a professional talk classification structure in an ontology talk network according to the present invention.
Professional conversation refers to conducting a specific purpose-based conversation, such as booking, reserving, or purchasing.
23 is a diagram illustrating an example of a conversation classification structure exchanged with an agent in a hospital call center during a professional conversation in an ontology conversation network according to the present invention.
24 is a view showing an embodiment of a classification structure in which a daily conversation, an emotional conversation, and a professional conversation are connected in an ontology conversation network according to the present invention.
Continuous conversation can be done while freely coming in everyday conversation, emotional conversation, and professional conversation.
While the present invention has been particularly shown and described with reference to exemplary embodiments thereof, it is to be understood that the invention is not limited to the disclosed exemplary embodiments. It will be understood that various modifications and changes may be made without departing from the scope of the appended claims.

Claims (6)

A method for constructing a dialog scenario database for an interactive system applied to an ontology dialog network,
(a) extracting a sentence from a conversation-type voice file or a conversation-type article;
(b) classifying the sentence extracted in the step (a) as a question and an answer;
(c) selecting whether the question and answer sentences classified in step (b) are linked and connected to each other by emotion, intention, affirmation, and negation, and creating the selected question and answer sentence as a list-type scenario;
(d) converting a question and an answer sentence of a scenario, which is a list type generated in step (c), into a semantic vector and learning; And
(e) databaseing the scenario learned in the step (d) into a continuous scenario meaning vector
The method comprising the steps of: The method according to claim 1,
After step (e), displaying the continuous semantic vector converted into the database on the ontology multidimensional space as one of a semantic word, a semantic node and a semantic cube
Further comprising the steps of: creating a dialog scenario database for a dialog system for a dialog system.
The method according to claim 1,
Wherein the conversation-type voice file in the step (a) is extracted after changing to text through speech recognition
The method comprising the steps of:
The method according to claim 1,
The extraction from the interactive-type article of step (a)
A specific word (word) is specified as a search term, and the data and the posts and comments retrieved by the specified search term are parsed and extracted
The method comprising the steps of:
The method according to claim 1,
The sentence extracted in the step (a) is automatically converted into a semantic vector
The method comprising the steps of:

A sentence extracting unit for extracting a sentence;
A sentence analyzing unit for classifying the sentence extracted by the sentence extracting unit into a question and an answer;
A dialog scenario generation unit for generating a scenario in which a sentence classified by the sentence analysis unit corresponds to each other and a dialog is connected;
A dialog scenario learning unit for learning a sentence of a dialogue scenario generated by the dialogue scenario generating unit;
A dialogue scenario database for storing the dialogue scenarios learned by the dialogue scenario learning unit as semantic vectors and storing them; And
An ontology relationship mapping unit for mapping a continuous semantic vector of the conversation scenario stored in the conversation scenario database on the ontology multidimensional space as one of a semantic word,
Wherein the dialog scenario database includes at least one of:

Images