KR102117287B1 - 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 specifically, it is to automatically search, collect, refine, and learn dialogue scenarios on SNS, radio, and broadcasting to automatically expand a dialogue area or improve dialogue quality. It relates to a method of constructing a dialogue scenario database for a dialogue system applied to an ontology dialogue network.
A feature of the present invention for achieving the above object is a method for constructing a dialogue scenario database for a dialogue system applied to an ontology dialogue network, (a) a sentence by voice or ID from a dialogue-type voice file or dialogue-type post, comment Extracting; (b) expressing the dialogue sentence in the form of a tree by ID or name according to the conversation development of the post or comment from the sentence extracted in step (a); (c) extracting a dialogue scenario from each tree of step (b); (d) expressing each sentence of the dialogue scenario generated in step (c) as a semantic vector; (e) expressing each sentence of the dialogue scenario generated in step (d) as an intention pattern; (f) database of the dialogue scenario into an intention pattern and a semantic vector; And (g) creating a new intention pattern from the sentence without the existing intention pattern in step (e).

Description

Method and apparatus of dialog scenario database constructing for dialog system

The present invention relates to the construction of a dialogue scenario database, and more specifically, an ontology dialogue for automatically expanding, conversing, or improving dialogue quality by automatically searching, collecting, refining, and learning dialogue scenarios on SNS, radio, and broadcasting. It is related to a method of constructing a dialogue scenario database for a dialogue system applied to a network.

Humans have had a long dream of naturally communicating with non-human subjects. Currently, the fourth industrial revolution that AI and big data have driven is underway, and the core of artificial intelligence is natural conversation communication between humans and objects.

However, to date, since the dialogue scenario is constructed by human beings, it is expensive, time and labor. As a person collects and refines and builds the scenario in person, it takes a relatively long time, and as a result, it is necessary to have a conversation in the past, so it is not possible to have an immediate conversation about events, accidents, and various trends or issues. Did not.

In other words, since only limited conversation is spoken in a limited area, there is a limitation that the conversation area cannot be expanded or the conversation quality can be improved.

KR 10-0444568 KR 0326931 KR 10-2012-0146174

The present invention was devised to solve this problem, and ontology dialogue network that automatically searches, collects, refines, and learns conversation scenarios on SNS, radio, and broadcasting to automatically expand the conversation area or improve conversation quality. An object of the present invention is to provide a method for constructing a dialogue scenario database for a dialogue system applied to the system.

The feature according to the present invention for achieving the above object is a method for constructing a dialogue scenario database for a dialogue system applied to an ontology dialogue relational network, (a) by voice or dialogue type post or comment by ID or name Extracting a sentence; (b) expressing the dialogue sentence in the form of a tree by ID or name according to the conversation development of the post or comment from the sentence extracted in step (a); (c) extracting a dialogue scenario from each tree in step (b); (d) expressing each sentence of the dialogue scenario generated in step (c) as a semantic vector; (e) expressing each sentence of the dialogue scenario generated in step (d) as an intention pattern; (f) database of the dialogue scenario into an intention pattern and a semantic vector; (g) creating a new intention pattern in the sentence without the existing intention pattern in step (e); And (h) when the intention pattern representing the corresponding conversation sentence is not matched with the conversation sentence with respect to the analyzed conversation sentence during the continuous conversation with the user, (h1) deriving a semantic vector of the conversation sentence; (h2) classifying the intention pattern corresponding to the dialogue sentence from the semantic vector; And (h3) when the intention pattern corresponding to the conversation sentence is classified, the intention pattern is determined as the intention pattern of the conversation sentence, matched to the conversation sentence, and stored in the ontology conversation network network database, and corresponding to the conversation sentence. When the intention pattern is not classified, the intention pattern mapped to the node closest to the node pointed to by the semantic vector is identified from the ontology dialogue network network database, and the determined intention pattern is determined as the intention pattern of the dialogue sentence, and the conversation is performed. And matching the sentence and storing it in the ontology conversation network database.

Preferably, after the step (e), further comprising the step of displaying the semantic vector of the database in the ontology multidimensional space as a semantic word or semantic node.

Preferably, the speech file of the conversational form of step (a) is extracted after changing to text through speech recognition.

Preferably, the extraction from the conversational post of step (a) is to designate a specific word as a search term, and to parse and extract posts and comments searched for by the specified search term.

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

Another feature of the present invention for achieving the above object is a dialogue scenario extraction unit for extracting a dialogue scenario; A conversation scenario learning unit for learning each sentence of the conversation scenario generated by the conversation scenario extraction unit; A conversation scenario intention pattern classification unit that classifies an intention pattern of each sentence of the conversation scenario using the semantic vector generated by the conversation scenario learning unit; An intention pattern generation unit that creates a new intention pattern for an intention pattern not classified by the dialogue scenario intention pattern classification unit; A dialogue scenario database storing the dialogue scenario intention pattern and the semantic vector; And an ontology relationship mapping unit displaying a continuous semantic vector and an intention pattern of the conversation scenario stored in the conversation scenario database on an ontology multidimensional space, and the intention pattern classification unit is provided to the analyzed dialogue sentence during the continuous conversation with the user. In contrast, if the intention pattern representing the corresponding conversation sentence is classified as not matching the conversation sentence, the intention pattern generation unit derives a meaning vector of the corresponding conversation sentence, and from the meaning vector, the intention pattern corresponding to the conversation sentence When the intention pattern corresponding to the conversation sentence is classified, the intention pattern is determined as the intention pattern of the conversation sentence, matched to the conversation sentence, and stored in the ontology conversation network network database, and corresponds to the conversation sentence. When the intention pattern is not classified, the intention pattern mapped to the node closest to the node pointed to by the semantic vector is identified from the ontology dialogue network network database, and the determined intention pattern is determined as the intention pattern of the dialogue sentence, and the conversation is performed. It matches the sentence and stores it in the ontology dialogue network.

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

In addition, if a person collects a scenario by himself, it takes a lot of time to refine and build it, so he will have a conversation in the past. It has the effect of improving the quality of conversation because it enables instant conversation.

In addition, since the dialogue scenarios are continuously collected on various topics, it is possible to have a dialogue on topics that reflect various viewpoints.

1 is a flowchart for explaining a method for constructing a dialogue scenario database for a dialogue system according to the present invention.
FIG. 2 schematically shows a database construction apparatus of a dialogue scenario for a dialogue system according to the present invention.
FIG. 3 is a detailed diagram of a database construction apparatus of a dialogue scenario for the dialogue system according to FIG. 2;
4 to 5 are screens showing an example of sentence extraction during the construction of a dialogue scenario database for a dialogue system according to the present invention.
6 is a screen showing an example of sentence analysis during the construction of a dialogue scenario database for a dialogue system according to the present invention.
7 is a screen showing an example of learning a conversation scenario while building a conversation scenario database for a conversation system according to the present invention.
8 is a diagram illustrating an example in which a dialogue scenario database for a dialogue system constructed according to the present invention is displayed in an ontology dialogue relationship network.
9 is an example of a conversation scenario stored in a conversation scenario database according to the present invention.
10 is a screen in which a dialogue scenario stored in a dialogue scenario database according to the present invention is represented by a 3D authoring tool.
11 is an example of a scenario input screen in which a dialogue scenario stored in a dialogue scenario database according to the present invention is represented by a 3D authoring tool.
12 is an example of a scenario modification screen in which a dialogue scenario stored in a dialogue scenario database according to the present invention is represented by a 3D authoring tool.
13 is an example of a scenario deletion screen in which a dialogue scenario stored in a dialogue scenario database according to the present invention is represented by a 3D authoring tool.
14 is a block diagram illustrating a sequence of automatic conversation scenario collection according to the present invention and a continuous conversation using an ontology conversation network.
15 is a flow chart illustrating a sequence in which a conversation is automatically performed using a dialogue scenario automatic collection and an ontology dialogue network according to the present invention.
16 is a flowchart illustrating a method of mapping a ontology dialogue network when a new dialogue text is input during continuous conversation using the ontology dialogue network according to the present invention.
17 is a block diagram illustrating a sequence for implementing dialogue quality improvement using continuous dialogue using an ontology dialogue network according to the present invention.
18 is a diagram showing the configuration of a continuous dialogue system using an ontology dialogue relationship network according to the present invention.
19 is a diagram illustrating an embodiment of an ontology dialogue network structure according to the present invention.
20 is a diagram illustrating an embodiment of a general conversation classification structure in an ontology dialogue network according to the present invention.
21 is a diagram showing an embodiment of a specialized conversation classification structure in an ontology dialogue network according to the present invention.
22 is a diagram illustrating an embodiment of a conversation classification structure exchanged with an agent in a hospital call center during a specialized conversation in an ontology conversation network according to the present invention.
23 is a diagram illustrating an embodiment of a classification structure in which daily conversation, emotional conversation, and professional conversation are connected in an ontology dialogue 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, the terms or words used in the present specification and claims should not be construed as being limited to ordinary or lexical meanings, and the inventor appropriately explains the concept of terms in order to explain his or her invention in the best way. Based on the principle that it can be defined, it should be interpreted as meanings and concepts consistent with the technical spirit of the present invention. Therefore, the configuration shown in the embodiments and drawings described in this specification is only one of the most preferred embodiments of the present invention and does not represent all of the technical spirit of the present invention, and thus can replace them at the time of application. It should be understood that there may be equivalents and variations.

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

The method for constructing a dialogue scenario database for the dialogue system according to FIG. 1 first extracts a sentence (S110). Extraction of sentences extracts sentences by ID or name from conversational voice files or conversational posts and comments.

When a sentence is extracted in step S110, a dialogue sentence is displayed in a tree form by ID or name according to the conversation development of a post or a comment from the extracted sentence (S120).

Then, a dialogue scenario is extracted from each tree expressed in step S120 (S130), and each sentence of the generated dialogue scenario is expressed as a semantic vector (S140).

Subsequently, each sentence of the generated dialogue scenario is expressed as an intention pattern (S150), and the dialogue scenario is databased with an intention pattern and a semantic vector (S160).

Here, a sentence without an existing intention pattern creates a new intention pattern (S170).

The semantic vector and semantic vector of the databased scenario are continuously displayed on the ontology multidimensional space (S160).

The method for constructing a dialogue scenario database for the dialogue system of the present invention according to FIG. 1 will be described through the apparatus for constructing a dialogue scenario database for the dialogue systems of FIGS. 2 to 3 as follows.

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

2 to 3, the dialogue scenario database construction apparatus 100 for a dialogue system according to the present invention is extracted by the sentence extraction unit 110 and the sentence extraction unit 110 for extracting sentences The scenario extracting unit 120 for extracting a dialogue scenario from a sentence, the semantic vector extracting unit 130 for extracting a semantic vector from each sentence extracted by the scenario extracting unit 120, and the semantic vector extracting unit 130 The intention pattern classification unit 140 that classifies the intention of each sentence, the intention pattern generation unit 150 that generates an intention pattern not classified by the intention pattern classification unit 140, and a dialogue scenario as an intention pattern and a semantic vector. It includes a dialogue scenario database 160 to store, and an ontology relation mapping unit 170 that displays continuous intention patterns and semantic vectors of dialogue scenarios stored in the dialogue scenario database 160 on an ontology multidimensional space.

A semantic vector is extracted from each sentence extracted by the sentence extraction unit 110, the scenario extraction unit 120 for extracting a dialogue scenario from the sentences extracted by the sentence extraction unit 110, and the scenario extraction unit 120 The semantic vector extracting unit 130 and the intention pattern classifying unit 140 that classify the intention of each sentence by the semantic vector extracting unit 130 and the intention pattern not classified by the intention pattern classifying unit 140 are generated. The intention pattern generating unit 150, the dialogue scenario database 160 storing the dialogue scenario as the intention pattern and the semantic vector, and the continuous intention pattern and the semantic vector of the dialogue scenario stored in the dialogue scenario database 160 are ontology multidimensional space And an ontology relation mapping unit 170 displayed on the image.

The sentence extraction unit 110 extracts a sentence by ID or name from a conversational voice file or a conversational post or comment, and the sentence extraction unit 110 is a voice file extraction module that extracts a conversational voice file Includes (111), a post extraction module 112 for extracting sentences from conversational posts and comments, and as shown in FIG. 3, voice files of conversational type such as call center, radio, and TV broadcasting are extracted as voice files. The sentence is extracted through the module 111, and when the search term is selected in the search term generator 113, the sentence in the conversation form posted on SNS, etc. is searched for by posting or commenting by post or ID by the post extraction module 112, Parse comments to extract sentences. For example, after voice recognition of an existing voice recording file provided in a call center, a sentence may be extracted from the text, or a voice call may be extracted from the text after real-time voice recognition of a consultation between a customer and an agent in the call center. On the other hand, SNS extracts sentences from postings on various topics on Twitter or Facebook, etc. First, it searches for SNS with prepared search terms, extracts each of the post links found with this search term (specific words), and then extracts each link. From the beginning of the first input post pointed to by the post to the last post (comment) for the discussion that started from the post, it is regarded as a conversation topic, and all posts (comments) for the conversation topic are extracted.

The scenario extraction unit 120 extracts a dialogue sentence by ID or name from a voice file in a conversation format or a post or comment in a conversation format by the sentence extraction unit 110.

For example, in the case of Twitter, more than one comment can be attached to one post, and another comment is continuously derived from the bottom like a tree trunk. In this way, depending on the development of the dialogue of posts and comments, dialogue sentences are displayed consecutively by ID or name in a tree shape, and in one tree starting with one post, there are as many scenarios as there are comments, and there are scenarios as many as the next number of comments. It increases as you hit.

In other words, if there are three comments in one post and two comments in one comment, the total dialogue scenario is four.

The semantic vector extracting unit 130 expresses each sentence constituting the dialogue scenario as a semantic vector. Conventionally, a technique of expressing a word or sentence as a semantic vector by word2vec and sentence2vec algorithms is a known technique. By utilizing this technique, each sentence constituting the dialogue scenario can be expressed as a semantic vector.

The intention pattern classification unit 140 classifies the intention pattern of each sentence constituting the dialogue scenario. The method of classifying the intention pattern can learn an intention pattern classified as a correct answer using algorithms such as DNN, CNN, and SVM, generate a learning model, and classify the intention pattern for newly input sentences by the learning model. .

The intention pattern generation unit 150 serves to newly generate an intention pattern that is not classified by the intention pattern classification unit 140. In this case, when the intention pattern is not classified, the probability value (eg 0.5 or less) classified as the specific intention pattern is too low than the expected value (eg 0.8) when classified as a specific intention pattern. At this time, the predicted value is 10-20% lower than the probability value average (0.95) classified as the correct answer when the algorithm learned by the learning sentence is tested by the test sentence.

The intention pattern generating unit 150 extracts a candidate word for a region and a candidate for an action (action) from a single sentence, and uses a wordnet or the like to use a combination of a region word and an operation word composed of the higher word and the lower word information. Create intent patterns.

In the dialogue scenario database 160, the dialogue scenario is databased with an intention pattern and a semantic vector. At this time, in the dialogue scenario, intention patterns and semantic vectors composed of sentence-based questions and answers are stored continuously. The semantic vectors of sentences and scenarios are also expressed as intention patterns to help humans intuitively grasp.

The ontology relation mapping unit 170 maps the dialogue scenarios of the continuous intention patterns and the semantic vectors stored in the ontology dialogue network when the dialogue scenario database 160 storing the continuous intention patterns and the semantic vectors of the dialogue scenario is built. Display function. The ontology dialogue network is a multi-dimensional space composed of 300 to 600 vectors, but because it cannot physically represent a multi-dimensional vector, it is compressed and displayed in three dimensions. The ontology dialogue network represents one point in a three-dimensional space as a node, and each node can be represented from an intention pattern and a semantic vector stored in the database 160.

The ontology dialogue network can upload a continuous intention pattern and a semantic vector of a dialogue scenario stored in the dialogue scenario database 160 through the ontology relation mapping unit 170, and make the scenario itself as a semantic vector and upload it. When searching for similar sentences and similar scenarios, the scenario closest to the entered scenario can be mapped by means of distance calculation in the semantic vector space. At this time, sentences and scenarios can be expressed on the semantic vector space, but sentences and scenarios each have a separate semantic space. In addition, the meaning space of sentences and scenarios can be mapped to one meaning space, and at the same time, sentences and scenarios can be expressed in space. Among thought methods, sentences are expressed as sums or products of word vectors, scenarios are expressed as sums or products of sentence vectors, etc., meaning vectors of sentence scenarios help humans to intuitively grasp For this, it is indicated by one or more intention patterns.

On the other hand, the conversation scenario stored in the conversation scenario database 160 can be directly entered by the administrator in the ontology conversation relationship network through the ontology relationship mapping unit 170, and the relationship between sentences constituting the scenario is described in the following [Example] 1] as the sequence of the semantic pattern.

[Example 1]

Scenario 1 = (question 1)-(answer 1)-(question 2)-(answer 2) .... (question N)-(answer N)

Scenario 1 = (Where did you see it?)-(You wondered my name?)-(No where I saw you)-(Where did we last meet?) .... (Yeah)-(Perhaps, I think I think we saw it last at the last Yali host seminar.)

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

 Here, the intention pattern "name ambiguity" is a combination of the domain word "name" and the action word "ambiguity", and is an intention pattern representing a sentence such as "Where have you seen it?".

A single intention pattern expresses more than one sentence semantically, and can also be said to be a representative word representing numerous sentences with the same meaning.

In addition, semantic patterns such as "meaning ambiguous" and "name verification" also represent the continuity of each question and answer constituting one scenario, such as "(question 1)" and "(answer 1)".

4 to 5 are screens illustrating an example of sentence extraction during database construction of a dialogue scenario for a dialogue system according to the present invention.

As illustrated in FIG. 4, the sentence extraction extracts a word from Divpedia or DBNet, uses the word as a search term, and uses SNS post and comment data by crawling, etc. Parse through the web. This is possible because when a post is posted from Twitter, a retweet comment is sent in a similar way to a conversation on a specific topic.

5, another user continues to comment while retweeting the post of the user who posted the first post, and the development of the conversation is similar to the tree form. Here, if each tree is viewed as a dialogue scenario and the dialogue scenario is expressed by comment ID to help understanding, it is as follows [Example 2].

[Example 2]

Scenario 1: ffebreze-hatter365-fffebreze

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

FIG. 6 is a screen showing an example of intention pattern analysis during the construction of a dialogue scenario database for a dialogue system according to the present invention, and after parsing a post as illustrated, a map of machine learning what the post implies in a conversation scenario Classify conversation intention patterns by learning.

7 is a screen showing an example of a conversation scenario learning while building a conversation scenario database for a conversation system according to the present invention. Questions and answers do not appear alternately on a single topic, and may occur due to overlapping questions and answers (eg, Question 1, Answer 1, Question 2, Question 2, Answer 2, Question 3, Answer 3, Answer 3, Answer 3, etc.) Among these questions and answers, questions and answers suitable for the scenario are selected, and the screened dialogue scenario as shown in FIG. 7 shows one question and repeated learning of one answer to the question.

8 is a screen illustrating an example in which a dialogue scenario database for a dialogue system constructed according to the present invention is mapped to an ontology dialogue relationship network.

Scenarios are represented as sequences of semantic vectors consisting of sentence-based questions and answers. Each question and answer constituting one scenario has a multidimensional semantic vector, and each question and answer sentence is automatically mapped to one intention pattern.

When a large number of scenarios are input as a multidimensional semantic vector value, the vector distance between the inputted semantic vector value and the existing intention pattern is compared, and when it is within a certain value, an existing intention pattern may be assigned to the input semantic vector value. have.

When a large number of scenarios is inputted as well as a multidimensional semantic vector value and an intention pattern and a dialogue sentence are entered at the same time, it is necessary to map the dialogue scenario (each question sentence and answer sentence) to each intention pattern space. At this time, the intention pattern can be extracted by automatically analyzing the "conversational sentence" of the scenario, or a person can attach an intentional pattern suitable for the "conversational sentence" directly.

[Example 3]

Scenario 1 = [Intentional Pattern] [Conversational Text] [Meaning Vector], [Intentional Pattern] [Conversational Text] [Meaning Vector] ....

Scenario 1 = [Name Ambiguity] [I think I saw it] [2.382, 6.108, ...], [Name Check] [Are you curious about my name?] [8.730, 1,383, ....] ....

The method of automatically classifying [Intention Pattern] can be analyzed based on rule-based, statistical-based, or machine-learning among natural language processing methods. Intention patterns are classified.

9 is an example in which a conversation scenario stored in a conversation scenario database according to the present invention is expressed, and the conversation scenario is usually represented by two turns or more (question 1-answer 1-question 2-answer 2 relationship), and is flat in one space The number of scenarios that can be displayed is also limited, and scenarios that are spatially close to the current scenarios are unknown. In addition, even if a scenario is input or modified, there is a problem in that a correlation in semantic space with other scenarios is not known at all.

Dialogue scenarios with specific characters in order to avoid duplication as intended and displayed as a continuous sequence of pattern, illustration, like the first scenario, the top, "the name is ambiguous", "C Check Names", "C yes", "KE name of IU" It can also be displayed in combination.

FIG. 10 is a screen in which a dialogue scenario stored in a dialogue scenario database according to the present invention is represented by a 3D scenario authoring tool. As shown in FIG. 10, a scenario may be displayed on a 3D space. Each question and answer in the scenario is represented by a pattern of intent. Depending on the semantic closeness between the intention patterns belonging to each scenario, it is displayed in the near space or in the far space, so it is much easier to grasp the semantic relationship between the intention patterns constituting the scenario.

11 is an example of a scenario input screen in which a dialogue scenario stored in a dialogue scenario database according to the present invention is represented by a 3D scenario authoring tool. As shown in FIG. 11, the scenario is intended by dragging the right intention pattern to the left margin space A scenario is expressed by a sequence of patterns (eg, weekend schedule-schedule answer-hobby query-hobby answer), and when the scenario input is completed, the scenario is automatically mapped into the 3D conversation network. In addition, the right intention patterns all have multi-dimensional semantic vector values, and can be made by extracting domain words and combining them (weekends + schedules) or by semantic vector sums or products of many sentences constituting the intention patterns.

The right intention pattern (eg, weekend schedule confirmation) represents various sentences, and has various sentence structures as shown in [Table 1] below, which implies "weekend schedule confirmation".

Intention pattern Sentence example Check weekend schedule What is your weekend schedule? Please tell me the weekend schedule. I'm curious about the weekend schedule. Tell me about the weekend schedule. Do you have a weekend schedule?

12 is an example of a scenario modification screen in which a dialogue scenario stored in a dialogue scenario database according to the present invention is represented by a 3D scenario authoring tool, and a scenario can be searched by a search function, and a function for correcting a searched scenario is provided. . When the scenario is modified, the modified scenario is immediately reflected in the 3D space.

13 is an example of a scenario deletion screen in which a dialogue scenario stored in the dialogue scenario database according to the present invention is represented by a 3D scenario authoring tool, the scenario can be deleted, and the deleted scenario is completely deleted even in the 3D space, and is no longer visible. It becomes impossible.

14 is a block diagram illustrating a sequence of automatic conversation scenario collection according to the present invention and continuous conversation using an ontology conversation network, and FIG. 15 illustrates automatic conversation scenario collection according to the present invention and continuous conversation using an ontology conversation network. It is a diagram showing a sequence made of a flow chart.

As described above, the ontology dialogue network is a multidimensional vector space composed of 300 to 600 vectors, but since it cannot physically represent a multidimensional vector, it is represented by spatial compression (using a spatial compression algorithm such as PCA) in three dimensions. The ontology dialogue network represents a point in 3D space as a node, and each node can represent a word meaning vector, a sentence meaning vector, and a scenario meaning vector. When referring to the 'Ontology Dialogue Network', the overall structure of such a node and the scenario composition by connection between nodes is expressed, and when referring to the 'Ontology Dialogue Network', the storage unit that stores such data and structure relations is represented. Hereinafter, it will be used interchangeably without any significant difference.

In addition, as described above, the intention pattern is a phrase representing several sentences with similar meanings, and may consist of a single word or a combination of several words. For example, the intention pattern "name ambiguity confirmation" is a phrase that combines the region word "name", the region word "ambiguity", and the action word "confirmation". The semantic vectors of "name", "ambiguity", and "confirmation" It is also a combined new semantic vector. It is as described above that it is a phrase that represents a sentence such as "I saw it anywhere." That is, the intention pattern can also be represented by a semantic vector in the ontology dialogue network space and has one node.

Scenarios can be displayed on a three-dimensional ontology dialogue network space, and each question and answer in the scenario can be represented by a single semantic vector. Depending on the semantic closeness between the semantic vectors belonging to each scenario, it is displayed in the near space or in the far away space, so it is much easier to understand the semantic relations between nodes constituting the scenario.

Hereinafter, a process of performing a continuous conversation using an ontology dialogue network will be described with reference to the block diagram sequence of FIG. 14 and the flowchart of FIG. 15.

First, when the user speaks, the user's voice is input to the continuous conversation system 200 (see FIG. 18) using the ontology conversation network of the present invention (S210). The continuous dialogue system 200 using the ontology dialogue network analyzes the sentence for the sentence understanding of the received voice (S220). The continuous dialogue system 200 using the ontology dialogue relation network determines, from the ontology dialogue relation network database, whether the intention pattern representing the dialogue phrase matches the analyzed dialogue phrase (S230).

If the intention pattern is not matched (S240), the intention pattern is determined for the corresponding conversation sentence, matched to the conversation sentence, and stored in the ontology conversation relation network database (S250), which will be described later in detail with reference to FIG. 16.

If the intention pattern is not matched to the dialogue sentence and a new intention pattern is determined and matched to the dialogue sentence, or the intention pattern is already matched to the dialogue sentence, the intention pattern is matched to the dialogue sentence An intention pattern (hereinafter referred to as a 'second intention pattern') followed by an intention pattern (hereinafter referred to as a 'first intention pattern') matched to the conversation sentence according to a conversation scenario in the ontology conversation network network database, after being stored in the ontology conversation network network database. Is extracted) (S260), and one of the dialogue sentences included in the second intention pattern is selected (S270), and the selected dialogue sentence is output to the user by voice (S280). For example, the first intention pattern may be an intention pattern representative of the question dialogue sentence input by the current user, and the second intention pattern is determined by the continuous conversation system 200 using the ontology dialogue network to answer this. The answer may be an intention pattern representing a dialogue sentence.

Depending on the dialogue scenario in the ontology dialogue network, there may be a case where a plurality of intention patterns following the first intention pattern matched to the dialogue sentence exist. In this case, the intention pattern having the highest weight assigned to the scenario of each of the plurality of intention patterns may be extracted as the second intention pattern. For example, the scenario following the first intention pattern is: first intention pattern-> 'A' (weight 0.7), first intention pattern-> 'B' (weight 0.2), first intention pattern-> 'C' ( When there are three such as weight 0.1), it is possible to determine the intention pattern A as the second intention pattern by determining the first intention pattern-> 'A' (weight 0.7). Such a weight can be set in advance through conversation quality management that diagnoses and analyzes a conversation with a user, which will be described later with reference to FIG. 17.

16 is a flowchart illustrating a method of mapping a ontology dialogue network when a new dialogue text is input during continuous conversation using the ontology dialogue network according to the present invention.

When a new scenario is entered, the semantic vector values of the question and answer sentences of the corresponding scenario are extracted, and the dialogue intention of the corresponding question answer semantic vector is classified and input scenarios ("dialog intention (intention pattern), dialogue sentence, meaning" Vector "in a contiguous list).

That is, when the intention pattern is not matched to the conversation sentence input by the user, that is, when a new scenario is input (S240), a semantic vector of the corresponding conversation sentence is derived (S251), and the conversation sentence is derived from the semantic vector The intention pattern corresponding to is classified (S252). As a method of classifying intention patterns, one or more of rule-based, statistical-based, and machine-learning-based methods may be used. Here, machine learning corresponds to the block 'conversation scenario learning' in FIG. 14.

When the intention pattern corresponding to the conversation sentence is classified (S252), the intention pattern is determined as the intention pattern of the conversation sentence and matches to the conversation sentence and stored in the ontology conversation relationship network database (S253). In this case, as described above, a continuous list of classified intention patterns, a continuous list of dialogue sentences, and a continuous list of semantic vectors may be stored in the ontology dialogue network network.

When the intention pattern corresponding to the dialogue sentence is not classified (S252), the intention pattern mapped to the node closest to the node pointed to by the semantic vector is identified from the ontology dialogue network network, and the identified intention pattern is communicated. Determine the sentence intention pattern and match it to the conversation sentence and store it in the ontology conversation relationship network database (S254). At this time, a continuous list of classified intention patterns, a continuous list of conversation sentences, and a continuous list of semantic vectors Can be stored in the ontology conversation network database.

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

A scenario can be expressed as a continuous pair of question and answer. Just as each question and answer gathers to form a scenario, a scenario provides information about a specific conversation area (education, literature, common sense, everyday conversation, sports, movies, etc.). Scenarios can also be expressed as a combination of intention patterns.

Scenario classification can be classified into each scenario using one or more of rule-based, statistical-based, and machine-learning-based methods. Accordingly, before classifying the conversation intention of each sentence, the scenario classification is first selected, and then the conversation intention (intention pattern) belonging to the corresponding scenario classification is first searched to shorten the search time.

FIG. 17 is a block diagram illustrating a sequence for implementing dialogue quality improvement using continuous dialogue using an ontology dialogue network according to the present invention.

The user can instantly know whether the user's reaction is positive or negative as the user proceeds with the system. User reactions such as positive reactions, negative reactions, neglect reactions, and topic change reactions can be known by analyzing the user's dialog, and at this time, 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 can be recognized as the preferred dialogue scenario, and the probability of the current scenario being adopted is increased through weight adjustment, etc., the preferred scenario is continuously selected, and the non-preferred scenario is given priority. The goal is to continuously manage conversation scenarios through real-time user reactions, and ultimately, the more continuous conversations, the better the conversation quality.

When such a process is described with reference to FIG. 17, continuous conversation quality management is performed through a conversation scenario history between a normal user and a continuous conversation system 200 using the ontology conversation network of the present invention. That is, the sentence inputted from the user is analyzed, and the user's reaction is analyzed from the analyzed dialogue. If the user's response is analyzed as positive, the likelihood of selecting a conversation scenario leading to the corresponding conversation sentence or the intention pattern of the corresponding conversation sentence is set high. It is to set the likelihood that the dialogue scenario leading to will be selected is low.

Here, positive or negative is a very comprehensive term. As shown in the example of FIG. 17, it may be divided into details such as a praise reaction, a positive response, a negative response, a neglect reaction, and a topic change reaction, but it can be described as a positive or negative response by covering all of these. This would not be analyzed with two kinds of reactions, but rather a more accurate expression of giving a kind of degree or degree. That is, the more responsive and good the user's response by sentence analysis is, the higher the likelihood that a dialogue scenario in which such a reaction will occur through feedback will be selected in the future. It makes the possibility of being selected lower.

The possibility of selecting a conversation scenario in the future may be made by assigning a weight related to the preference / non-preferred conversation scenario to a conversation scenario in which the corresponding user reaction has been generated according to a preset criterion. This increases the probability that a dialogue scenario with high weight will be selected in the future.

At this time, the analysis of the user's reaction is performed by analyzing the user's dialogue sentence according to rules (rule-based), based on the statistics of the dialogue sentence (statistic-based), or by machine learning (machine learning) Machine learning based).

Alternatively, the analysis of the user reaction may be performed by a user preference that is pre-analyzed and mapped to an intention pattern corresponding to the user's dialogue sentence in the ontology dialogue network. That is, the intention pattern space and the user preference space may be separately provided on the ontology conversation network database, and user preferences of the specific intention pattern node of each intention pattern space and the same node of the user preference space may be mapped. That is, user preferences analyzed in response to a user's response to the user's intention pattern are pre-mapped, from which the weight for the conversation can be set.

18 is a diagram showing the configuration of a continuous dialogue system using an ontology dialogue relationship network according to the present invention.

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

The voice input unit 202 receives a user's voice.

The sentence analysis unit 203 performs sentence analysis from the input voice.

The dialogue management unit 204, from the ontology dialogue network network database, leads to an intention pattern (hereinafter referred to as a 'first intention pattern') representing an intention pattern (hereinafter referred to as a 'first intention pattern') for the analyzed dialogue sentence. And extracts one of the conversation sentences included in the second intention pattern as an output conversation sentence. In addition, the conversation management unit 204, if there are a plurality of intention patterns following the first intention pattern matched to the conversation sentence according to the conversation scenario in the ontology conversation relational network database, is assigned to each intention pattern scenario The intention pattern having the highest weight may be extracted as the second intention pattern.

When the intention pattern is not matched to the analyzed dialogue sentence, the new scenario processing unit 206 determines a new intention pattern for the sentence, matches the sentence, and stores the determined intention pattern in the ontology dialogue network network database. It plays a role of delivering to the conversation management department.

The new scenario processing unit may derive a semantic vector of the corresponding dialogue sentence if the intention pattern is not matched to the received dialogue sentence. From the derived semantic vector, an intention pattern corresponding to the dialogue sentence is classified. When the intention pattern corresponding to the conversation sentence is classified, the intention pattern is determined as the intention pattern of the conversation sentence, matched to the conversation sentence, and stored in the ontology conversation network network database, and the intention pattern corresponding to the conversation sentence is classified If not, the intention pattern mapped to the node closest to the node pointed to by the semantic vector is identified from the ontology dialogue network database, and the determined intention pattern is determined as the intention pattern of the dialogue sentence and matched to the dialogue sentence. It can be stored in the ontology dialogue relation network database, and transmits the determined intention pattern to the dialogue management unit.

As described above with reference to FIG. 16, the intention pattern mapped to the node closest to the node indicated by the semantic vector may be a new word indicated by a semantic vector generated by a single word or a combination of words.

The dialogue quality management unit 205 analyzes a user's reaction from the analyzed dialogue sentence, and when a user response is analyzed to be positive, a possibility that a conversation scenario leading to the intention pattern of the conversation sentence or the conversation sentence is selected is selected. If it is set as high and the user response is analyzed to be negative, dialogue quality management is performed by setting the probability that a dialogue scenario leading to the corresponding dialogue sentence or intention pattern of the dialogue sentence is selected is low. At this time, as described above with reference to FIG. 17, the setting of the possibility of selecting the conversation scenario may be made by weighting according to a preset criterion. In addition, analysis of the user's reaction is performed by analyzing the dialogue sentence according to rules (rule-based), based on statistics of the dialogue sentence (statistic-based), or by machine learning (machine learning) Running based). Alternatively, the analysis of the user reaction may be performed by a user preference that has been previously analyzed and mapped to an intention pattern corresponding to the user's conversational sentence in the ontology conversational network database, which has been described in detail with reference to FIG. 17. .

The conversation output unit 207 serves to output a conversation sentence selected by the conversation management unit to a user through voice or the like.

19 is a diagram illustrating an embodiment of an ontology dialogue network structure according to the present invention.

That is, referring to the embodiment of FIG. 19, the conversation in the ontology conversation relational network may be classified into such a domain by setting a general conversation domain, a specialized conversation domain, or another new domain as required. That is, a large number of nodes, such as an intention pattern mapped to the ontology dialogue network space, may already be set to areas such as general conversation and specialized conversation, respectively. That is, when obtaining a semantic vector for each sentence, a node pointed to by each semantic vector according to the meaning of the sentence may be configured to point to an area of a predetermined classification in the ontology dialogue network space. Of course, the same may be applied to the case where detailed classification is performed under a general conversation or a specialized conversation.

The general conversation (daily conversation, common sense conversation, topic conversation, etc.) develops conversations on various topics according to the conversation scenario in the conversation network.

Specialized conversations (reservations, reservations, purchases, etc.) acquire specific information (Argument) from the user according to a specific purpose conversation strategy, and conduct conversations for performing tasks. For example, in the case of KTX reservation, the user is asked for information necessary for reservation, such as reservation time, number of passengers, and seat type, and the reservation is completed only after the information is satisfied. Therefore, it is necessary to follow a conversation procedure (Flow), which is suitable for a specific purpose (Task).

In this way, general and professional conversations can be freely conducted in the ontology dialogue network.

20 to 23 show examples of conversations classified in the ontology conversation network.

20 is a diagram illustrating an embodiment of a general conversation classification structure in an ontology dialogue network according to the present invention.

General conversation can include daily conversation, common sense conversation, topic conversation, emotional conversation, topic conversation, etc. If you constantly move the focus of conversation on various topics, you can talk.

21 is a diagram illustrating an embodiment of a specialized conversation classification structure in an ontology dialogue network according to the present invention.

Professional conversation refers to conducting specific purpose-based conversations such as reservation, reservation, and purchase.

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

23 is a diagram illustrating an embodiment of a classification structure in which everyday conversation, emotional conversation, and professional conversation are connected in an ontology dialogue network according to the present invention.

Conversations can be made continuously while freely coming and going between daily conversation, emotional conversation, and professional conversation.

As described above, although the present invention has been described by a limited number of embodiments and drawings, the present invention is not limited by this, and the technical idea of the present invention and the following will be described by those skilled in the art to which the present invention pertains. Of course, various modifications and variations are possible within the equivalent scope of the claims to be described.

100: conversation scenario database building device
110: sentence extraction unit
111: voice file extraction module
112: post extraction module
113: Query generator
120: sentence analysis unit
130: conversation scenario generation unit
140: conversation scenario learning department
150: conversation scenario database
160: ontology relation mapping unit
200: continuous dialogue system using ontology dialogue network

Claims (6)

As a method of constructing a dialogue scenario database for a dialogue system applied to an ontology dialogue network,
(a) extracting a sentence by ID or name from a conversational voice file or a conversational post or comment;
(b) expressing the dialogue sentence in the form of a tree by ID or name according to the conversation development of the post or comment from the sentence extracted in step (a);
(c) extracting a dialogue scenario from each tree in step (b);
(d) expressing each sentence of the dialogue scenario generated in step (c) as a semantic vector;
(e) expressing each sentence of the dialogue scenario generated in step (d) as an intention pattern;
(f) database of the dialogue scenario into an intention pattern and a semantic vector;
(g) creating a new intention pattern in the sentence without the existing intention pattern in step (e); And,
(h) When the intention pattern representing the corresponding conversation sentence is not matched with the above conversation sentence with respect to the analyzed conversation sentence during the continuous conversation with the user,
(h1) deriving a meaning vector of the corresponding dialogue sentence;
(h2) classifying the intention pattern corresponding to the dialogue sentence from the semantic vector; And
(h3) When the intention pattern corresponding to the conversation sentence is classified, the intention pattern is determined as the intention pattern of the conversation sentence, matched to the conversation sentence, and stored in the ontology conversation relation network database, and the intention corresponding to the conversation sentence If the pattern is not classified, the intention pattern mapped to the node closest to the node pointed to by the semantic vector is determined from the ontology dialogue network database, and the determined intention pattern is determined as the intention pattern of the dialogue sentence to determine the intention pattern Matching and storing in ontology conversation network database
Conversation scenario database construction method for a conversation system comprising a. The method according to claim 1,
The semantic vector created in step (f) with the database is displayed as a semantic word or semantic node.
Method for building a dialogue scenario database for a dialogue system for a dialogue system, further comprising a. The method according to claim 1,
The speech file of the conversational form of step (a) is extracted after changing to text through speech recognition
Method for constructing a dialogue scenario database for a dialogue system, characterized in that. The method according to claim 1,
Extraction from the conversational article of step (a) above,
Specifying a specific word as a search term, and parsing and extracting posts and comments searched for by the specified search term
Method for constructing a dialogue scenario database for a dialogue system, characterized in that. The method according to claim 1,
The sentence extracted in step (a) is automatically converted into a semantic vector
Method for constructing a dialogue scenario database for a dialogue system, characterized in that. A conversation scenario extraction unit for extracting a conversation scenario;
A conversation scenario learning unit for learning each sentence of the conversation scenario generated by the conversation scenario extraction unit;
Conversation scenario intention pattern classification unit to classify the intention pattern of each sentence in the conversation scenario using the semantic vector generated by the conversation scenario learning unit
An intention pattern generation unit that creates a new intention pattern for an intention pattern not classified by the dialogue scenario intention pattern classification unit;
A dialogue scenario database storing the dialogue scenario intention pattern and the semantic vector; And
An ontology relationship mapping unit displaying continuous semantic vectors and intention patterns of the conversation scenarios stored in the conversation scenario database on an ontology multidimensional space.
Including,
When the intention pattern classification unit classifies the intention pattern representing the corresponding conversation sentence for the analyzed conversation sentence during the continuous conversation with the user as not matching the conversation sentence,
The intention pattern generation unit derives a meaning vector of the corresponding conversation sentence, classifies an intention pattern corresponding to the conversation sentence from the meaning vector, and when the intention pattern corresponding to the conversation sentence is classified, recalls the intention pattern Determined as the intention pattern of the conversation sentence, matched to the conversation sentence, stored in the ontology conversation network, and if the intention pattern corresponding to the conversation sentence is not classified, is mapped to the node closest to the node indicated by the semantic vector. An apparatus for constructing a dialogue scenario database for a dialogue system that grasps an intention pattern from an ontology dialogue relational network database, determines the identified intention pattern as the intention pattern of the dialogue sentence, matches the dialogue sentence, and stores it in the ontology dialogue relational network database.

Images