KR20170043955A - Voice conversation system and method of thereof - Google Patents

Abstract

A voice communication system and a voice communication method in the system are disclosed.
In this system, the dialogue understanding part performs language processing including morphological analysis, lexical analysis, and semantic analysis on a query sentence input from a user, and segments the keyword into units of key words for pattern inference matching. For a key word that can be interpreted as a plurality of semantic entries. Wherein the dialog management unit performs a conversation pattern and a semantic area inference based on the keyword segmented by the conversation understanding unit, generates a multiple semantic pattern composed of a combination of the plurality of semantic entries when the keyword is composed of a plurality of semantic entries, And performs matching between the multiple semantic patterns and predetermined pattern reasoning rules. The dialog generation unit generates a response corresponding to the pattern inference rule matched by the dialogue management unit.

Description

TECHNICAL FIELD [0001] The present invention relates to a voice conversation system and a voice conversation method in the system.

The present invention relates to a speech dialogue system based on sentence pattern inference using multiple semantic patterns and a speech dialogue method in the system.

As natural language speech recognition, natural language based processing technology, and speech synthesis technology rapidly develop, there is an increasing demand for a technology that combines them. The user can retrieve and command the natural language voice directly from the typing of inputting characters directly from the smart device, process the information based on the natural language voice, and receive the voice. The technology field of these technologies is becoming an issue with the technology of voice conversation agent which is intelligent conversation based on natural speech and the scale of conversation processing market is also increasing every year. Recently, IoT (internet of Things), which can be remotely controlled and monitored by voice conversation, is being developed.

In the field of voice conversation processing, a variety of techniques for providing answers to user query sentences are being researched. Basic conversation methods include word exact matching method, whole sentence matching method, and sequential pattern matching method.

In these methods, there is a problem that it is difficult to judge the intention of the user by the single meaning tagging of the word. For that reason, exceptions such as the versatility of words can occur in the main meaning of a word. Also, there is a problem that it is difficult to accept multiple semantic patterns based on semantic multiplicity of words. Also, the matching method has problems due to the problem of having to collect and train a lot of similar sentences for matching with the question sentences, and the inconsistency of the vocabulary that occurs when matching the question sentences.

Therefore, there is a need for a method that can respond to the query by accepting the multiplicity representing the various meanings of the word.

SUMMARY OF THE INVENTION The present invention has been made in view of the above problems, and it is an object of the present invention to provide a voice conversation system capable of accurately determining a user's query intention through a semantic pattern and stochastically matching pattern inference rules using the same, Of the user.

According to one aspect of the present invention,

A language processing including morpheme analysis, lexical analysis and semantic analysis is performed on a query sentence input from a user, and segmented into a keyword unit for pattern inference matching, and a key word that can be interpreted as one or more meanings for each word in keyword extraction A conversation understanding part constituting a plurality of meaning entries; The method comprising: generating a plurality of semantic patterns composed of a combination of the plurality of semantic entries when the key word is composed of a plurality of semantic entries, and performing the semantic domain inference based on the key words segmented by the conversation understanding section, A dialogue management unit for matching the pattern with a predetermined pattern inference rule; And a dialog generation unit for generating a response corresponding to the pattern inference rule matched by the dialogue management unit.

Here, the dialogue understanding unit may include a morpheme analysis unit for segmenting and analyzing the morpheme of the query sentence into a morpheme, which is a minimum unit based on the morpheme dictionary, and extracting a proverb and a verb; A semantic analysis unit for analyzing the meaning of a word by referring to a predefined semantic dictionary for words morphed by the morpheme analysis unit and constructing a plurality of semantic entries for the word when the word has a plurality of semantics; And a keyword extracting unit for extracting a keyword using the result of the semantic analysis unit.

Also, the dialogue management unit may include: a multi-pattern generation unit for generating multiple semantic patterns by combining a plurality of semantic entries of the key words extracted by the dialogue understanding unit; A pattern inference unit for matching a pattern inference rule of a knowledge base that is constructed in advance among the multiple semantic patterns generated by the multi-pattern generation unit and extracting matched pattern inference rules; And a reasoning analyzer for analyzing the knowledge base based on a pattern inference rule result extracted from the pattern reasoning unit and analyzing a response corresponding to the query sentence.

In addition, the dialog generation unit may generate a response sentence corresponding to the pattern inference rule matched by the dialogue management unit and knowledge information to be searched in the knowledge base, as an answer script structurally describing the answer sentence, and provide the answer script to the user.

In addition, the multi-pattern generation unit may generate multiple semantic patterns using a Cartesian product, which links all the semantic entries in order with respect to the column vectors corresponding to the plural semantic entries per word, .

In addition, the pattern reasoning unit may calculate a Bayes rule used to calculate a conditional probability by an approximate probability calculation method in pattern inference rules in a process of matching with pattern inference rules constructed among multiple semantic patterns of the query sentence, Rule) is used.

In addition, if the semantic entries of the keyword extracted from the dialogue understanding unit are independent, the pattern reasoning unit obtains the matching result of the pattern reasoning rule by multiplying the result of pattern matching with the pattern reasoning rule of each semantic belonging to the semantic entry .

In addition, the pattern reasoning unit applies a smoothing technique to prevent a semantic entry having a frequency from being included in a semantic dictionary but not observed in a pattern inference rule, .

The answer script may further include a pattern tag indicating pattern inference rules; An action tag indicating an action command for the user; A response tag indicating a response sentence to the query sentence; And a result tag indicating a list of knowledge information searched in the knowledge base.

A voice input unit for recognizing a user voice for conversation, analyzing a voice to be recognized, extracting features, converting the extracted voice into a corresponding query sentence, and transmitting the extracted query to the query understanding unit; And a speech synthesis unit for artificially synthesizing the speech corresponding to the response generated by the dialog generation unit and outputting the synthesized speech to the user.

According to another aspect of the present invention,

A language processing including morpheme analysis, lexical analysis and semantic analysis is performed on a query sentence input from a user, and segmented into a keyword unit for pattern inference matching, and a key word that can be interpreted as one or more meanings for each word in keyword extraction Constructing a plurality of semantic entries for; The method includes generating a plurality of semantic patterns composed of a combination of the plurality of semantic entries if the key word is composed of a plurality of semantic entries, performing a dialog pattern and a semantic area inference based on the segmented key words, Performing matching with a rule; And generating a response corresponding to the matched pattern inference rule and outputting the generated response to the user.

The step of constructing the plurality of meaning entries may include segmenting and analyzing the morpheme of the query sentence into a morpheme, which is a minimum unit based on the morpheme dictionary, and extracting a proverb and a verb; Analyzing the meaning of the word by referring to a predefined semantic dictionary for the morpheme-analyzed words, and constructing a plurality of semantic entries for the word if the word has a plurality of semantics; And extracting the key word using the semantic analysis result of the word.

The step of performing matching with the pattern inference rule may include generating a plurality of semantic patterns by combining a plurality of semantic entries of the extracted key words; Matching pattern inference rules of a knowledge base constructed in advance among the multiple semantic patterns and extracting matched pattern inference rules; And analyzing the knowledge base by the extracted pattern inference rule result and analyzing the response corresponding to the query sentence.

여기서,

는 패턴 추론 규칙

의

번째 패턴 규칙이고,

는 질의 문장

의 다중 의미 패턴 집합이며,

는

중

번째 다중 의미 패턴이고,

는 질의 문장

의

번째에서

번째 단어의 의미 엔트리임을 사용하여 패턴 추론 규칙을 추출하는 것을 특징으로 한다.In addition, the step of extracting the pattern reasoning rule may include a step of matching the pattern reasoning rule constructed among multiple semantic patterns of the query sentence,

here,

The pattern inference rule

of

Th pattern rule,

Query sentence

Is a set of multiple semantic patterns,

The

medium

Th multiple semantic pattern,

Query sentence

of

From the

Th word, and the pattern inference rule is extracted using the semantic entry of the first word.

가

개의 단어

로 구성되고, 상기 질의 문장

중에서 연속된 두 단어

와

에 대해 의미 엔트리가 각각

,

인 경우, 두 단어의 모든 의미 엔트리들에 대해 카티션 곱을 이용하여

의 다중 의미 패턴을 생성하는 것을 특징으로 한다.In addition,

end

Word

, Said query sentence

Two consecutive words

Wow

Lt; RTI ID = 0.0 >

,

, Then a Cartesian product is used for all the semantic entries of the two words

And generates a plurality of semantic patterns.

According to the present invention, it is possible to accurately answer a user's query intention through a semantic pattern and stochastically match a pattern inference rule using the pattern inference rule, thereby providing a response closer to the query intention.

Also, by using multiple meanings of words, it is possible to solve the problems caused by the storage space and vocabulary mismatch of many similar sentences for training.

In addition, it is possible to respond to semantic similar query patterns, and it is useful to extend the scope of query response.

1 is a block diagram showing a schematic configuration of a voice chat system according to an embodiment of the present invention
Fig. 2 is a diagram showing a specific configuration of the dialogue understanding portion shown in Fig. 1. Fig.
3 is a diagram showing a specific configuration of the dialogue management unit shown in Fig.
4 is a view for explaining contents constituting a plurality of semantic entries of words in a speech communication system according to an embodiment of the present invention.
FIG. 5 is a diagram illustrating an example of constructing a plurality of meaningful words in a speech dialog system according to an embodiment of the present invention.
6 is a diagram illustrating a process of generating multiple semantic patterns in a speech communication system according to an embodiment of the present invention.
FIG. 7 is a diagram showing an example of generating multiple semantic patterns using the example of FIG.
8 is a diagram showing an example of matching with pattern inference rules using the example of FIG.
9 is a diagram showing an example of generating an answer script using the example of FIG.

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings so that those skilled in the art can easily carry out the present invention. The present invention may, however, be embodied in many different forms and should not be construed as limited to the embodiments set forth herein. In order to clearly illustrate the present invention, parts not related to the description are omitted, and similar parts are denoted by like reference characters throughout the specification.

Throughout the specification, when an element is referred to as "comprising ", it means that it can include other elements as well, without excluding other elements unless specifically stated otherwise. Also, the terms " part, "" module," and " module ", etc. in the specification mean a unit for processing at least one function or operation and may be implemented by hardware or software or a combination of hardware and software have.

Hereinafter, a voice communication system according to an embodiment of the present invention will be described with reference to the drawings.

1 is a block diagram showing a schematic configuration of a voice conversation system 10 according to an embodiment of the present invention

1, the voice conversation system 10 includes a voice input unit 100, a conversation understanding unit 200, a conversation management unit 300, a conversation generation unit 400, and a voice synthesis unit 500 .

The voice input unit 100 recognizes a user voice for conversation, analyzes the voice to be recognized, extracts features, and converts the voice into a corresponding text sentence. Here, the user's dialogue input sentence is input in a natural language using a sentence of voice or text. At this time, the voice conversation system 10 may output a greeting or an input request message prior to the input

The dialogue understanding unit 200 performs language processing such as morphological analysis, lexical analysis, and semantic analysis on a text sentence transformed by the speech input unit 100, thereby segmenting the text sentence into major key word units for pattern inference matching.

The dialogue management unit 300 performs conversation pattern and semantic domain inference based on the key words segmented by the dialogue understanding unit 200. [ More specifically, the conversation management unit 300 performs conversation semantic analysis, pattern analysis, and inference rule matching based on the keyword.

The dialog generation unit 400 generates a response suitable for the pattern reasoning rules matched by the dialogue management unit 300. [ This response is provided by describing the response data and the knowledge information optimized in various databases based on the pattern inference rule as the search response script.

The speech synthesizer 500 artificially synthesizes the speech corresponding to the response generated by the dialog generator 400 and outputs it to the user.

In the embodiment of the present invention, the dialogue understanding unit 200 extracts a plurality of meanings for a key word that can be interpreted as one or more meanings when extracting a key word from an input sentence.

In addition, the dialogue management unit 300 generates multiple semantic patterns composed of a plurality of semantic combinations for each key word extracted in the multiple meanings in the dialogue understanding unit 200, Matching of rules is performed to generate optimized responses.

Each of the components 100, 200, 300, 400, 500 of the voice chat system 10 illustrated in FIG. 1 above or a partially combined configuration thereof may be configured as hardware for executing software, As shown in FIG. Of course, hardware and software may be combined and configured. When configured as software, an application program implemented for all or some of the functions of the illustrated components 100, 200, 300, 400, 500 is recorded in a storage medium and read Lt; / RTI > device.

2 is a diagram showing a specific configuration of the dialogue understanding unit 200 shown in FIG.

2, the dialogue understanding unit 200 includes a morphological analysis unit 210, a semantic analysis unit 220, and a key word extraction unit 230. [

The morpheme analysis unit 210 segments and analyzes the morpheme of the text sentence converted by the speech input unit 100 into a morpheme that is the minimum unit based on the morpheme dictionary 240 having information such as the part of speech of the word, .

The semantic analysis unit 220 interprets the meaning of the word by referring to the predefined semantic dictionary 250 for the morpheme analyzed words by the morpheme analysis unit 210. [ At this time, if each word has a plurality of meanings, a plurality of meaning entries are formed for the corresponding word. For example, when the word segmented by the morpheme analysis unit 210 is "leg ", the" leg "can be interpreted as" leg / leg " 220) can construct two semantic entries for the word "leg "," leg "and" bridge ".

The keyword extracting unit 230 extracts a keyword using the result of the semantic analyzer 220. The extraction of these key words is well known, so a detailed description thereof will be omitted here.

3 is a diagram showing a specific configuration of the dialogue management unit 300 shown in FIG.

3, the dialogue management unit 300 includes a multi-pattern generation unit 310, a pattern inference unit 320, and an inference analysis unit 330. The multi-

The multi-pattern generating unit 310 generates multiple semantic patterns by combining the entries of at least one semantic pattern of the key words extracted by the conversation understanding unit 200. [ Such a multiple semantic pattern generation method will be described later.

The pattern reasoning unit 320 extracts pattern matching rules and pattern matching rules of the knowledge base 340 constructed in advance from among multiple semantic patterns generated by the multi-pattern generating unit 310. In this case, since it is virtually impossible to generate only a single pattern of a conversation sentence based on a single meaning corresponding to a word, a result of a high probability value among multiple sentence patterns composed of an entry combination of semantics interpreted in various meanings of words The appropriate pattern rules and conversation sentences have a meaning.

The inference analysis unit 330 analyzes the database and the knowledge base according to the pattern inference rule results extracted from the pattern inference unit 320 and obtains the inference results. Here, in order to solve the problem of scanning a large database and a knowledge base, a pattern inference rule and a matching process based on the pattern inference rule are analyzed and a knowledge information linked thereto is derived.

Therefore, the dialogue generation unit 400 can generate a response script through a suitable pattern-response template using the knowledge information generated by the pattern inference analysis in the inference analysis unit 330, , And a conversation script based on key words. Here, the conversation script includes various information such as a response time, a response sentence, a subject area, a search count, and action information. In particular, when the action information is a command operated by the client, the action code "show_weather" is transmitted in the case of "Tell the weather", so the client recognizes the action code to provide weather information and images, Please inform me. "In this case," find_path_map "is transmitted to provide information such as map image.

Hereinafter, a method for generating multiple semantic patterns in the voice conversation system 10 according to the embodiment of the present invention will be described.

Prior to the description, the Naive Bayes model, which is used for the probabilistic matching of multiple semantic-based patterns and pattern inference rules in the embodiment of the present invention, is a well-known classification method. . These Naive Bayes models are used in text classification as a popular method when mapping multiple documents to a classification. In the embodiment of the present invention, a pattern inference rule based on a semantic pattern matching is derived from a given query sentence, and the knowledge base is found and provided in the knowledge base 340. In this way, We use the Naive Bayes model to generate multiple semantic patterns and to calculate the probabilities assigned to pattern inference based on these multiple semantic patterns.

First, we explain the general method of predicting between query sentences and pattern inference rules in a basic way.

는

개의 단어

로 구성되고, 패턴 추론 규칙

은

개의 패턴 엔트리

로 구성될 때, 각 단어

와 패턴 엔트리

(1≤

≤

)가 매칭된다고 가정하자. 그러면, 단어

이 나타날 때, 패턴 추론 규칙

로 표현될 확률을 나타낼 수 있다. 이러한 확률은 다음 [수식 1]과 같이 근사될 수 있다. Given query sentence

The

Word

And the pattern inference rule

silver

Pattern entries

, Each word

And pattern entry

(1?

≤

) Are matched. Then,

, The pattern inference rule

Can be expressed as the probability of being expressed as. This probability can be approximated as in the following [Equation 1].

[Equation 1]

개의 가능한 확률적 패턴 결과들 중에서 가장 높은 확률값

을 계산해야 한다. 하지만, 다음과 같은 문제점을 갖는다. 1) [수식 1]은 단어의 수

개가 많은 경우나 학습 데이터가 미흡한 경우 확률 패턴 규칙에 베이즈 모델을 바로 적용하기 어려움이 있다. 2) 하나의 단어가 여러 의미를 갖는 다의성과 여러 단어가 하나의 의미를 갖는 동의성의 경우 질의의 의미 해석에 모호성의 문제가 발생할 수 있다. 마지막으로, 3) 정확한 패턴 규칙을 도출하기 위해서는 주어진 문장과 유사한 학습 데이터가 방대하게 있어야 하므로 이를 갖추어야 하는 조건이 쉽지 않다. here,

Of the possible probabilistic pattern results, the highest probability value

. However, it has the following problems. 1) [Equation 1] is the number of words

If there are many dogs or if the training data is insufficient, it is difficult to apply the Bayesian model directly to the probability pattern rule. 2) There is ambiguity in the semantic interpretation of the query in the case of the ambiguity in which a word has multiple meanings and the correspondence in which several words have a single meaning. Finally, in order to derive an accurate pattern rule, it is not easy to have the necessary data because there is a huge amount of learning data similar to the given sentence.

In order to solve these problems, the embodiment of the present invention uses a matching method between multiple semantic patterns and pattern inference rules. The matching method of inference rule is a method of finding corresponding pattern inference rule in query sentence. First, a given sentence is segmented into word of morphological unit and then interpreted into several meanings, and multiple semantic patterns are generated by combination of them. Next, we derive a suitable pattern based on pattern inference rules in multiple semantic patterns, and deduce the meaning of the sentence. In particular, we propose a systematic method to calculate the probabilistic matching between them. The main notations used in the embodiment of the present invention, and definitions and meanings thereof, are shown in Table 1 for the matching of the multiple semantic patterns of the query sentence and the pattern inference rules.

[Table 1]

The following is a description of how to tag the main meaning, which is a pattern component of a sentence, as a preparation process for matching pattern inference rules using multiple semantic patterns of a query sentence.

Key meanings of words Tagging is the task of assigning mean features of words in relation to the entries made in the pattern of sentences, and can be provided as an important clue to deduce patterns.

는

개의 단어

로 구성되며, 특정 단어

는 의미 패턴의 엔트리

를 가진다고 가정한다. 이 과정에서, 단어가 가지는 확률값은 사전에 형태소 분석을 통해 질의 문장에서 단어 열로 분절되고 패턴 의미 분석 사전을 통해 주요 단어의 패턴 의미가 태깅된 학습 데이터에 의하여 계산하여 결정할 수 있다. 다시 말해,

개의 단어 중 특정 단어

가 의미 패턴의 엔트리

로 나올 확률은

로 나타낼 수 있으며, 이를 통해 의도하는 패턴 추론을 계산할 수 있다. 하지만, 특정 단어가 갖는 의미는 다양한 의미 엔트리들로 정의할 수 있으며, 실제

를 계산하는 것은 쉽지 않다. 이에 따라, 본 발명의 실시예에서는 단어의 여러 의미 엔트리들을 수용하면서 사용자 의도를 명확하게 파악하는 방법에 대해 설명한다. In order to determine the semantic pattern of the query sentence,

The

Word

And a specific word

Is an entry of the semantic pattern

. In this process, the probability value of a word is segmented into a word sequence in the query sentence through morphological analysis in advance, and the pattern meaning of the main word can be calculated by the tagged learning data through the pattern semantic analysis dictionary. In other words,

Specific word

The entry of the semantic pattern

The probability of

, Which allows the calculation of the intended pattern inference. However, the meaning of a particular word can be defined by various meaning entries,

It is not easy to calculate. Accordingly, in the embodiment of the present invention, a method of clearly grasping the intention of the user while accepting various semantic entries of words will be described.

In the embodiment of the present invention, it is assumed that each word in the query sentence is assumed to be independent of each other and independent of the order in which the word appears before applying the Naive Bayes model. Based on this, we start by tagging the main meaning of a word.

에 대해 여러 의미 엔트리 m_j의 일대 다의 관계로 정의될 수 있다. 여기서, 특정 단어

에 대해 단어의 의미 엔트리 m_j의 값이 널(NULL) 값이면 의미가 존재하지 않아 제외 가능하다. Referring to FIG. 4, the meaning of a word defined in the pattern semantic dictionary is analyzed with words segmented by morpheme to construct a multiple meaning entry for each word. In this case, unlike existing methods that define one-to-one relationship between word and meaning entry after semantic tagging,

Multiple entries can be defined as meaning one-to-many relationship of m _j for. Here,

If the value of the semantic entry m _j of the word is a null value, it is possible to exclude it because there is no meaning.

For example, referring to FIG. 5, when the query sentence S is "legs are too sick ", the word" legs " segmented by morphemes is semantically referred to as " "Too" can be interpreted as "too / so," "too," and the words "sick" can be interpreted as "sick" or "hurt.

Next, a method for generating multiple semantic patterns based on semantic tagging according to an embodiment of the present invention will be described.

에 대해

개의 패턴 의미 엔트리

를 기반으로 구성하는 것에 대해 이미 설명하였다. 그런 후, 질의 문장의 패턴 추론을 위한 준비 과정으로, 각 단어에서 해석된 의미 엔트리를 기반으로 다중 의미 패턴들을 구조화하는 방법에 대해 설명한다. We propose a method for structuring multiple semantic patterns in matching process with pattern inference rules. That is, the meaning of the words segmented by the morpheme is extracted,

About

Pattern Meaning Entry

Has already been described. Then, we describe how to structure multiple semantic patterns based on semantic entries interpreted in each word, as a preparation process for pattern inference of query sentences.

에서 각 단어

에 대한

개의 패턴 의미 엔트리들

를 기반으로 의미 엔트리들의 조합을 재귀적으로 유도하여 생성된 여러 결과를 의미한다. 다시 말해, 단어 열

중에서 연속된 임의의 두 단어

와

에 대하여, 단어의 태깅된 의미 엔트리는 각각

와

이다. 여기서,

와

는 각각 1ㅧ

또는 1ㅧ

의 열 벡터를 갖는다. 이들 열 벡터들에 대하여, 모든 의미 엔트리들을 순서대로 옆으로 연결하여 새로운 관계를 만드는 카티션 곱(Cartesian product)을 이용하여 유도하면,

과 같이 여러 벡터들이 구조화 된다. 이를 기반으로 의미 기반의 패턴을 추출하면 다중 의미 패턴

로 생성할 수 있으며, 이를 적절히 사용하면 사용자 질의 의도가 근접하게 반영될 수 있게 된다. Referring to FIG. 6, the structure of the multiple semantic patterns is a word sequence

In each word

For

Pattern Meaning Entries

And the result is a result obtained by recursively deriving a combination of semantic entries. In other words,

≪ / RTI >

Wow

, The tagged semantic entries of the words are

Wow

to be. here,

Wow

Respectively.

Or 1

Lt; / RTI > For these column vectors, if we derive all of the semantic entries in order by using a Cartesian product,

Several vectors are structured as shown in Fig. Based on this, extraction of semantic-based patterns can extract multiple semantic patterns

And if properly used, the intention of the user query can be reflected closely.

은 도면의 우측 부분과 같이 "leg+so+sick" 등으로 구성되는 의미 엔트리들을 갖는다. For example, an example of generating multiple semantic patterns is shown in Fig. Referring to FIG. 7, when a sentence such as "the leg is too sick" is given, when each word "leg", "too", "sick" constituting the sentence is analyzed through the morpheme, meaning tagging results such as "/ leg", "bridge /", "too / very", "sick" / "hurt" A combination of these must form a semantic pattern, which must be composed of multiple semantic patterns rather than a semantic pattern due to multiple semantics. Here, the structure of multiple semantic patterns is constructed by recursively guiding the combination of semantic entries of the respective words. Then,

&Quot; has " leg + so + sick "and the like as shown in the right part of the drawing.

Next, matching between multiple semantic patterns and pattern inference rules according to an embodiment of the present invention will be described.

In the process of pattern inference, as described above, when multiple semantic patterns based on the semantic tagging are generated, it is important to search for the meaningful pattern inference rules by searching for them using an appropriate matching method. Accordingly, in the embodiment of the present invention, the Bayes rule used to calculate the conditional probability by the approximate probability calculation method in the pattern inference rule in the process of matching with the pattern inference rule constructed among the multiple semantic patterns of the query sentence, Rule) is used. Thus, we use the Naive Bayes model of [Equation 2] to search for optimal matches in sentence patterns and pattern inference rules.

[Equation 2]

는 질의 문장에서 추출한 의미 패턴이 패턴 추론 규칙

에 속할 확률 값을 의미하며,

는 패턴 추론 규칙

에 속하는 질의 문장의 다중 의미 패턴들에서 임의 의미 패턴

일 확률 값을 의미한다.

는 질의 문장에서 의미 패턴

일 확률 값을 의미한다. here,

Is a semantic pattern extracted from a query sentence,

Quot ;, " probability "

The pattern inference rule

In the case of multiple semantic patterns of a query sentence belonging to a random semantic pattern

Probability value.

In the query sentence,

Probability value.

Using the following [Equation 3], if there are several semantic patterns of the query character, the probability can be calculated using the corresponding Bayes rule for each pattern inference rule, and the query pattern calculated with the highest probability can be calculated using the corresponding query sentence This implication can be deduced as a pattern inference rule.

[Equation 3]

는

개의 의미 패턴

과 각 의미 패턴

에 속해 있는

개의 의미 엔트리

로 구성될 때, [수식 4]에서와 같이 확률적으로 패턴 추론 규칙

을 추정하는 것이 가능하다. In addition, a set of multiple semantic patterns extracted from a query sentence

The

Meaning patterns

And each semantic pattern

Belonging to

Meaning entries

, The probability of the pattern inference rule is stochastically determined as in [Equation 4]

Can be estimated.

[Equation 4]

와 패턴 추론 규칙

의 확률을 계산하면, 특정 의미 패턴이 특정한 패턴 추론 규칙에 적합한지를 결정할 수 있다. 즉, 특정 의미 패턴이 확률이 가장 높은 패턴 추론 규칙에 매칭되는 가능성이 가장 높은 것이다. Here,

And pattern inference rules

, It is possible to determine whether a particular semantic pattern is suitable for a particular pattern inference rule. That is, a particular semantic pattern is most likely to match the pattern inference rule with the highest probability.

The Naive Bayes model can calculate the matching probability of the inference rule by taking into account the assumption that each word is independent in the Bayes rule of [Equation 4].

}으로 이루어진 특정 의미 패턴

이라고 가정한다. 그러면,

는 [수식 5]에 의해 나이브 베이즈 모델의 조건부 확률을 손쉽게 계산하여 패턴 추론 규칙에 확률적으로 매칭할 수 있다. 이에 따라, 각각의 단어의 의미 엔트리들은 독립적이므로 [수식 5]의 조건부 확률은 다음과 같이 계산된다. First, when a query word is composed of several words, each word having one or more semantic entries, the semantic entries are classified into a set derived by a Cartesian product {

} ≪ / RTI >

. then,

Can easily calculate the conditional probability of the Naive Bayes model and can stochastically match the pattern inference rule with [Equation 5]. Accordingly, since the semantic entries of each word are independent, the conditional probability of [Equation 5] is calculated as follows.

[Equation 5]

Therefore, the matching probability method for calculating the pattern reasoning rule in the semantic-based multiple semantic pattern with the Bayes rule is as follows.

[Equation 6]

Here, when calculating the pattern inference rule probability with the pattern generated in the query sentence, there occurs a problem that the probability becomes zero when the meaning of the new word exists. In the query sentence, meanings existing in the pattern inference rule may occur among the meanings of the whole words, so that the probability value can be calculated using the smoothing method in the embodiment of the present invention. This will be described in detail later.

Hereinafter, a method of matching pattern inference rules after generating a semantic pattern in a given sentence will be described.

을 위해 사용된다. First, a knowledge base 340 as shown in [Table 2] is required. The knowledge base 340 includes a conditional probability calculating unit 370 for calculating which pattern inference rule matches a query sentence

.

The components of knowledge base 340 are pattern inference, common_leg_pain, head_pain, find_around_hosp, ... And the pattern inference rules (Pattern List) are composed of leg + sick, head + sick, near + hosp, etc., and categories such as Pain and Hospital There is an answer. Here, the answer sentence can provide useful answers by substituting the key words in the answer using meta data such as $ {leg}, $ {head}.

[Table 2]

Now, when given the query sentence "When the legs are too sick", explain how the query sentence with the three words "bridge, too," hurts through the morphological process and matches the pattern inference rule.

As described above, the matching process of inference rules is performed by segmenting words in a query sentence into words through a morpheme, tagging various meanings of words using a pattern semantic dictionary, and generating multiple semantic patterns based on the semantic dictionary. Then, multiple semantic patterns consisting of a combination of semantic entities of words are generated as shown in [Table 3].

[Table 3]

이 패턴 추론 comm_leg_pain일 확률은

이며, head_pain일 확률은

이 된다. Next, we calculate the probabilities of which pattern inference rules match multiple semantic patterns. Any semantic pattern of multiple semantic patterns

The probability of this pattern inference comm_leg_pain is

, And the probability of head_pain is

.

이면 comm_leg_pain로 근사 될 수 있다. 그러므로,

은 패턴 추론 규칙 comm_leg_pain 중 의미 패턴

에서 의미 엔트리로 분류하면,

(leg, so, sick|comm_leg_pain)이며, 의미 엔트리 leg, so, sick는 서로 연관 관계가 없이 독립적이라는 가정 하에

(leg|comm_leg_pain)*

(so|comm_leg_pain)*

(sick|comm_leg_pain)으로 계산할 수 있다. 예를 들어, comm_leg_pain에서 총 패턴 엔트리 개수는 3개이고,

에서 의미 엔트리의 개수는 comm_leg_pain이면서 leg인 것이 1번, so인 것이 0번 그리고 sick인 것이 1번으로 다음의 [수식 7]과 같이 확률 값이 계산된다. if,

It can be approximated by comm_leg_pain. therefore,

Is the semantic pattern of the pattern inference rule comm_leg_pain

In the case of a semantic entry,

(leg, so, sick | comm_leg_pain), and the semantic entries leg, so, and sick are independent of each other

(leg | comm_leg_pain) *

(so | comm_leg_pain) *

(sick | comm_leg_pain). For example, in comm_leg_pain, the total number of pattern entries is 3,

The number of semantic entries is comm_leg_pain, the leg is 1, the so is 0, and the sick is 1, and the probability value is calculated as shown in [Equation 7].

[Equation 7]

(so|comm_leg_pain)에서 "so"는 의미 사전에는 존재하지만, 패턴 추론 규칙에서 관측이 되지 않은 엔트리로 단순히 빈도수를 가지고 매칭 확률을 계산할 경우 0이 되어 문제가 발생한다. 이러한 문제는 스무딩(Smoothing) 기법을 적용하여 해소할 수 있으며, 이를 적용한

계산식은 [수식 8]과 같다. By the way,

(so | comm_leg_pain), "so" exists in the semantic dictionary, but when the matching probability is calculated with a simple frequency as an entry that is not observed in the pattern inference rule, a problem arises. This problem can be solved by applying a smoothing technique,

The equation is shown in [Equation 8].

[Equation 8]

는 모든 단어에서 중복되지 않은 유일한 개수이다. [수식 8]을 이용하여, 재계산한 패턴 추론 규칙 comm_leg_pain의 매칭 확률 값과 다른 패턴 추론 규칙 head_pain과의 매칭 확률 값은 다음과 같다. 특히, 질문 문장에서 특정 성격을 대표하는 주제단어일 경우 가중치를 부여할 수 있다. here,

Is the only number not duplicated in all words. The matching probability value of the recalculated pattern inference rule comm_leg_pain and the matching probability value of another pattern inference rule head_pain using Equation (8) are as follows. In particular, weights can be given in the case of a subject word representing a specific personality in a question sentence.

Based on the above result, it is determined that the pattern inference result of the given sentence "legs are so sick" is highly probable with the pattern inference rule comm_leg_pain having the closest probability value as a result of calculation using the matching probability of multiple semantic patterns and pattern inference rules .

Next, a method by which the dialogue generation unit 400 generates an answer script according to an embodiment of the present invention will be described.

As described above, when the probabilistic matching process of the semantic-based multi-pattern and the pattern inference rule is performed, the knowledge information necessary for providing the answer to the query should be described in the answer script.

FIG. 8 is a diagram illustrating a process of matching multiple semantic patterns of a question sentence and pattern matching rules matched with the multiple semantic patterns of a question sentence in a speech dialog system according to an embodiment of the present invention.

Referring to FIG. 8, in the left part, a semantic pattern is extracted from a query sentence, a pattern inference rule is found based on the extracted multiple semantic patterns, and on the right side, optimization of response sentences, And finds the knowledge information in the knowledge base 340 resident in the memory. Based on this, the knowledge information found is provided to the user as a structured answer script.

Answers scripts are configured and delivered in JSON format for delivery to remote clients. This is a script generated with the results of the knowledge information retrieved from the database on the basis of category, pattern reasoning, etc. in the knowledge base 340.

An example of the JSON form of the response script is a formalized information delivery format, which is shown in FIG.

9, the dialog script retrieves information extracted from the knowledge base 340 and adds a pattern tag, an action tag, a response tag, an answer tag, a result tag, to be. In addition, various information such as the response time and the number of times of retrieval can be included as needed. Here, the pattern tag in the script is a pattern inference rule (e.g., "comm_leg_pain"), the action tag is a client action command (e.g., showList "), And the result tag represents a list of knowledge and related information (eg," category: private hospital, title: dasarang surgery hospital, ... ") retrieved from the database.

Such a dialogue script can be outputted to the user on a display or the like to provide answer and visualization information as well.

As described above, in the embodiment of the present invention, a method of providing an answer script describing answer and knowledge information through a process of analyzing the meaning of user query and matching with pattern inference rules has been described. Particularly, in this embodiment, a method of generating multiple semantic-based semantic patterns and performing probabilistic matching between them and pattern inference rules has been described. This method can provide a closer answer to the query intention by stochastically matching the pattern inference rules by judging the user's query intention accurately through the semantic pattern.

The proposed multi-semantic-pattern-based pattern reasoning rule matching can more accurately determine the intention of the user including the semantic multiplicity of the word and is superior to the conventional method in terms of space and management due to the collection of similar sentences and lexical inconsistency .

Meanwhile, the multiple semantic-based pattern inference matching according to the embodiment of the present invention can be applied to various methods of calculating the probability matching. In addition to the subject inference of the differentiated category using the inferred pattern rule, It can be applied to various application fields such as the way of maintaining the context of conversation by deducing pattern rules based on knowledge matching by using inference rules.

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 belongs to the scope of right.

Claims (15)

A language processing including morpheme analysis, lexical analysis and semantic analysis is performed on a query sentence input from a user, and segmented into a keyword unit for pattern inference matching, and a key word that can be interpreted as one or more meanings for each word in keyword extraction A conversation understanding part constituting a plurality of meaning entries;
The method comprising: generating a plurality of semantic patterns composed of a combination of the plurality of semantic entries when the key word is composed of a plurality of semantic entries, and performing the semantic domain inference based on the key words segmented by the conversation understanding section, A dialogue management unit for matching the pattern with a predetermined pattern inference rule; And
A dialog generation unit for generating a response corresponding to the pattern inference rule matched by the dialogue management unit,
And a voice communication system. The method according to claim 1,
The conversation understanding unit,
A morpheme analyzing unit for segmenting and analyzing the morpheme of the query sentence into morpheme units based on a morpheme dictionary and extracting a proverb and a verb;
A semantic analysis unit for analyzing the meaning of a word by referring to a predefined semantic dictionary for words morphed by the morpheme analysis unit and constructing a plurality of semantic entries for the word when the word has a plurality of semantics; And
A key word extraction unit for extracting a key word using the result of the semantic analysis unit,
And a voice communication system. The method according to claim 1,
Wherein the dialog management unit comprises:
A multi-pattern generation unit for generating multiple semantic patterns by combining a plurality of semantic entries of key words extracted from the dialogue understanding unit;
A pattern inference unit for matching a pattern inference rule of a knowledge base that is constructed in advance among the multiple semantic patterns generated by the multi-pattern generation unit and extracting matched pattern inference rules; And
Analyzing the knowledge base based on a pattern inference rule result extracted from the pattern inference unit, and analyzing a response corresponding to the query sentence;
And a voice communication system. The method of claim 3,
Wherein the dialog generation unit generates a response sentence corresponding to a pattern inference rule matched by the dialogue management unit and knowledge information that is searched in the knowledge base as an answer script structurally describing the response sentence and provides the answer script to a user. The method of claim 3,
The multi-
A plurality of semantic patterns are generated by using a Cartesian product which connects all the semantic entries side by side in order for the column vectors corresponding to the plural semantic entries per word
And the voice communication system. 6. The method of claim 5,
The pattern reasoning unit may calculate a Bayes' rule to be used when the conditional probability is calculated by the approximate probability calculation method in the pattern inference rule in the process of matching with the pattern inference rule constructed among the multiple semantic patterns of the query sentence, Is used. The method according to claim 6,
If the semantic entries of the keyword extracted from the dialogue understanding unit are independent, the pattern reasoning unit determines the matching result of the pattern reasoning rule by multiplying the result of pattern matching with the pattern reasoning rule of each semantic belonging to the semantic entry And the voice communication system. The method according to claim 6,
The pattern reasoning unit applies a smoothing technique to prevent a semantic entry having a frequency from being included in a semantic dictionary but not observed in the pattern inference rule and to be zero when calculating a matching probability Voice chat system. 5. The method of claim 4,
The answer script includes:
A pattern tag indicating pattern inference rules;
An action tag indicating an action command for the user;
A response tag indicating a response sentence to the query sentence; And
A result tag indicating a list of knowledge information searched in the knowledge base
And a voice communication system. 10. The method according to any one of claims 1 to 9,
A voice input unit for recognizing a user voice for conversation, analyzing a voice to be recognized and extracting features, converting the extracted voice into a corresponding query sentence, and transmitting the extracted query to the query understanding unit; And
A speech synthesis unit for artificially synthesizing a voice corresponding to a response generated by the dialog generation unit and outputting the synthesized voice to a user,
Further comprising: A language processing including morpheme analysis, lexical analysis and semantic analysis is performed on a query sentence input from a user, and segmented into a keyword unit for pattern inference matching, and a key word that can be interpreted as one or more meanings for each word in keyword extraction Constructing a plurality of semantic entries for;
The method includes generating a plurality of semantic patterns composed of a combination of the plurality of semantic entries if the key word is composed of a plurality of semantic entries, performing a dialog pattern and a semantic area inference based on the segmented key words, Performing matching with a rule; And
Generating a response corresponding to the matched pattern inference rule and outputting the generated response to the user
/ RTI > 12. The method of claim 11,
Wherein constructing the plurality of semantic entries comprises:
Segmenting and analyzing the morpheme of the query sentence into a morpheme that is a minimum unit based on a morpheme dictionary, and extracting a proverb and a verb;
Analyzing the meaning of the word by referring to a predefined semantic dictionary for the morpheme-analyzed words, and constructing a plurality of semantic entries for the word if the word has a plurality of semantics; And
A step of extracting a key word using a result of semantic analysis of a word
/ RTI > 12. The method of claim 11,
Wherein the step of performing matching with the pattern inference rule comprises:
Generating multiple semantic patterns by a combination of multiple semantic entries of extracted key words;
Matching pattern inference rules of a knowledge base constructed in advance among the multiple semantic patterns and extracting matched pattern inference rules; And
Analyzing the knowledge base according to the extracted pattern inference rule result and analyzing a response corresponding to the query sentence
/ RTI > 12. The method of claim 11,
Wherein the extracting of the pattern inference rules comprises:
In the matching process with the pattern inference rule constructed among the multiple semantic patterns of the query sentence,

here,

The pattern inference rule

of

Th pattern rule,

Query sentence

Is a set of multiple semantic patterns,

The

medium

Th multiple semantic pattern,

Query sentence

of

From the

Meaning of the second word
And extracting a pattern inference rule using the pattern inference rule. 15. The method of claim 14,
The query sentence

end

Word

≪ / RTI >
The query sentence

Two consecutive words

Wow

Lt; RTI ID = 0.0 >

,

, Then a Cartesian product is used for all the semantic entries of the two words

To generate multiple semantic patterns of
And a voice communication method.

Images