KR20080026782A - Method and apparatus for comprehending speech using information extraction method - Google Patents

Abstract

A method and an apparatus for comprehending spoken words by using an information extraction method are provided to comprehend essential elements selectively in comprehending the spoken words on the basis of a meaning structure suitable for each specific domain, thereby improving a degree of comprehension for the spoken words. A method for comprehending spoken words by using an information extraction method comprises the following steps of: standardizing the meaning structure of the spoken words previously(210); embodying the standardized meaning structure to be suitable for a specific domain(220); inputting spoken words recognized through a voice recognition unit(230); performing the natural language processing of the inputted spoken words(240); selecting quality with a specific meaning for determining the meaning structure by a result analyzed through the natural language processing(250); performing mechanical studying by using the selected quality(260); and comprehending the spoken words based on the meaning structure formed by determining corresponding elements configuring the meaning structure through the mechanical studying(270).

Description

Method and apparatus for comprehending speech using information extraction method

1 is a block diagram of a speech language understanding system to which a speech language understanding apparatus using an information extraction technique according to the present invention is applied.

FIG. 2 is a flowchart illustrating an operation of a voice language understanding unit of FIG. 1.

3 is a block diagram of an apparatus to which the voice language understanding method of FIG. 2 is applied.

<Description of the symbols for the main parts of the drawings>

11: speech recognition unit 12: speech language understanding unit

13: conversation management unit 310: speech input unit

320: natural language processing unit 330: feature extraction unit

340: speech act classification unit 350: driving classification unit

360: core element extraction unit 370: semantic structure forming unit

The present invention relates to a method and apparatus for understanding a speech language using an information extraction technique. More particularly, the present invention relates to a speech language understanding method using an information extraction technique. The present invention relates to a method and apparatus for understanding a speech language using an information extraction technique for constructing a semantic structure and extracting information necessary for the semantic structure from a speech recognition result to understand the meaning of the speech language using natural language processing techniques and machine learning methods. .

In general, there are two types of speech language understanding methods in a speech language recognition system.

One method is to understand the phonetic language through the semantic grammar of manual construction. However, this method of understanding speech through semantic grammar requires expensive development cost, and there is a problem in that proper speech understanding cannot be achieved when users are ignorant about grammar rules supported by a specific system.

Another approach to understanding phonetic language is to use statistical methods to assign word strings directly to predefined semantic structures. In this method, hand-organized grammars or rules are replaced with statistical models that are automatically trained from the training data, which can be used to understand speech languages.

In addition, various speech language understanding methods have been proposed for recognizing speech languages with higher accuracy by applying detailed techniques.

SUMMARY OF THE INVENTION The present invention has been made in an effort to provide a speech semantic understanding method using an information extraction technique having a high understanding of a speech language used in a specific domain in a speech recognition result using a speech recognizer.

In addition, another technical problem of the present invention is to provide an apparatus for understanding speech semantics using an information extraction technique that can be efficiently applied to understand the speech language used in a specific domain in a speech recognition result through a speech recognizer.

Speech language understanding method using the information extraction technique to achieve the above technical problem,

Formulating a semantic structure for a speech language in advance;

Embodying the formalized semantic structure as described above for a specific domain;

Inputting a recognized speech language through a speech recognizer;

Performing natural language processing on the input voice language;

Selecting a feature having a characteristic meaning to determine a semantic structure based on a result analyzed through natural language processing;

Machine learning using the selected qualities; And

Through the machine learning, it characterized in that it comprises the step of understanding the speech language based on the semantic structure formed by determining the corresponding elements constituting the semantic structure.

In addition, the formalizing of the semantic structure may include setting a speech act to classify the sentence structure of the general aspect into a semantic structure, driving which is a behavior pattern intended in a specific domain, and key elements supporting the behavior pattern of the driving. It is done.

In addition, the step of materializing the semantic structure may define the speech act as a sentence, question, request, etc. as a sentence form, and support one or more detailed behavioral patterns and the detailed behavioral patterns intended by the specific domain in driving. It is characterized by defining the detailed elements included in the core element.

In addition, the specific domain is characterized in that the purpose of the domain is distinctive, such as 'Tee program guide', 'car road guide'.

In addition, the natural language processing is characterized by performing a morpheme analysis, a sphere structure analysis, a syntax structure analysis for the speech language.

In addition, an apparatus for understanding speech language using information extraction technique for achieving another technical problem of the present invention,

A speech input unit for receiving a speech language to be understood from an operation result of the speech recognizer;

A natural language processor analyzing the recognized speech language inputted into the speech input unit according to a natural language processing process;

A feature extraction unit for selecting a feature having a characteristic meaning to be used as a basis for determining a meaning structure later from the result of the natural language processing unit;

A speech act classification unit for classifying a speech act in a semantic structure into a predefined sentence form by applying a feature set selected by the feature extractor to a machine learning technique;

A driving classification unit for applying the feature set selected by the feature extracting unit to a machine learning technique to classify the driving in a semantic structure into a predefined detailed behavior pattern;

A core element extracting unit for applying the feature set selected by the feature extracting unit to a machine learning technique to classify the key elements supporting the behavior of the driving into detailed sub-elements in the meaning structure; And

The speech act classification unit, the driving classification unit, and the core element extracting unit may include a semantic structure forming unit for forming a semantic structure possessed by the voice language as detailed elements classified into respective categories.

Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings.

1 is a block diagram of a speech language understanding system to which a speech language understanding apparatus using an information extraction technique according to the present invention is applied.

In FIG. 1, when the speech recognized input voice is recognized by the speech recognition unit 11, the speech language understanding unit 12 understands the language of the speech output from the speech recognition unit 11, and understands the speech language. The conversation manager 13 manages the conversation.

As described above, the technology applied by the speech language understanding unit 12 is a technology that provides an intermediate step of connecting the speech recognition technology and the conversation management technology, and outputs the speech of the speaker through the speech recognition unit 11 in the form of a speech notation. It converts this output form into a specific semantic structure that can be recognized later by the conversation manager.

In addition, in performing this process, by applying information extraction technology to understand the speech language, the problem of speech language understanding is approached as an information extraction problem for the purpose of extracting essential elements from the speech language. do.

2 is a flowchart illustrating a method of understanding a speech language according to the present invention.

In FIG. 2, in order to understand the speech language according to the present invention, first, a general semantic structure is formulated to understand the speech language (step 210).

As used herein, a speech language also includes specialized speech and general speech used in a particular domain.

Taking the case of the sentence in which the speech is characterized as an example, step 210 is described in detail below.

In order to understand the meaning of the phonetic language, the meaning structure is formalized in many aspects. That is, the semantic structure is added to specify the form of the sentence (action act) to which the utterance belongs in general terms, the style of behavior intended for the specific domain (driving), and the meaning of the action intended in the specific domain. As the component (core element), it is divided into three elements.

As described above, after shaping the semantic structure, the semantic structure is embodied for each specific domain (step 220).

As the semantic structure is formalized in step 210, the speech act is classified in general terms and thus has a common semantic structure in various areas. However, in the case of driving and key elements, different structures are required according to specific domains.

In step 220, the acts, driving, and key elements of the formalized semantic structure are classified and defined according to the specific domain.

For example, if a specific domain is the domain of the program guide, if you have already embodied a semantic structure that is formalized by speech acts, driving, and key elements,

Speech acts; Statement, question, request, ...

Drive: Turn on TV, turn off TV, find program, find program time, ...

Key elements: genre, program name, start time, channel name, actor name, ...

For example, it defines what kinds of detailed elements are included in the dialogue act, driving, and key elements in a specific domain.

As such, driving and key elements defined in the 'Tip Program Guide' domain have characteristics that are independent of the driving and key elements used in domains with different inherent purposes, such as the 'car path finding' domain. Can be made easier.

After the semantic structure is embodied, a speech-recognized speech language to be understood is input (step 230).

When the recognized speech language to be understood is input as a text, first, a natural language processing process such as morphological analysis, spherical structure analysis, and syntax structure analysis is performed (step 240).

In operation 240, a feature extraction part having a characteristic meaning to play a decisive role in determining a future semantic structure based on the natural language processing result of the speech is selected from the recognized speech language (250). step). Since the process of selecting a feature having a characteristic meaning from the speech language is also applied to a general speech language understanding method, a detailed description thereof will be omitted.

Subsequently, machine learning is performed using the selected qualities in step 250 to finally determine a semantic structure corresponding to speech acts, driving, and key elements (step 260).

Here, as a machine learning method, various commonly known techniques such as a maximum entropy model, support vector machines, and conditional random fields may be applied.

Although a flowchart of a method of understanding a speech language is briefly illustrated in FIG. 2, steps 210 and 220 need to be performed only once for a specific domain in order to understand the speech language in a specific domain. For the input speech language, it may also be performed as a step for pre-machine learning whether the corresponding elements in the semantic structure belong to the detailed elements defined by speech acts and driving, or which detailed elements defined in the core elements.

In addition, after completing the machine learning for the speech language in a specific domain, if you want to understand the speech language by substantially applying the results learned for that specific domain, step 230 of inputting a speech to form a semantic structure Only steps 270 to 270 need to be performed. Therefore, the formalization and specification of the semantic structure is performed in advance, and the speech is input as a machine learning step for the speech language of the dictionary, and the detailed details of speech acts, driving, and key elements are defined for each speech based on the detailed semantic structure. After the machine learning is completed to match the element, the speech structure is understood by analyzing the semantic structure of the input speech substantially through steps 240 to 270.

For example, when the voice language recognized from the voice recognition unit 11, “Where do Kim Tae-hee come out” is input, by performing steps 230 to 270,

Speech acts: question

Drive: Find a program

Key Elements: Hero-Kim Tae Hee, Genre-Drama

The result of the analysis of the semantic structure in the form of. In this way, by performing the machine learning of 260 steps to form the analyzed semantic structure by integrating speech acts, driving, and key elements determined (step 270), the meaning of the first input speech is understood based on this semantic structure.

Hereinafter, a speech language understanding apparatus to which the above-described speech understanding method is applied will be described. 3 is a block diagram of an apparatus to which the voice language understanding method of FIG. 2 is applied.

As shown in FIG. 3, the apparatus for understanding a speech language of the present invention includes a speech input unit 310, a natural language processor 320, a feature extractor 330, a speech act classifier 340, a driving classifier 350, and a key element. It is configured to include an extraction unit 360, and semantic structure forming unit 370.

The speech input unit 310 may receive a character of the recognized speech language from the speech recognition unit, and the natural language processor 320 may perform morphological analysis, sphere structure analysis, syntax structure analysis, etc. on the recognized characters of the input speech. Do this.

The feature extractor 330 may include one or more features having a characteristic meaning necessary to determine a future meaning structure from a result of analyzing a morpheme, a sphere structure, and a syntax structure of a speech recognized sentence by the natural language processor 320. Choose.

The act act classification unit 340 classifies the act act included in the semantic structure into a corresponding character form in a pre-set statement, question, request statement, etc. by using the selected qualities in the speech language input from the feature extracting unit 330 in machine learning. do. In addition, the driving classification unit 350 classifies the driving included in the semantic structure into a corresponding behavioral pattern among the detailed behavioral patterns previously defined on a specific domain by performing machine learning using the characteristics selected by the feature extraction unit 330. do.

In addition, the core element extractor 360 is configured to determine which of the detailed elements matched in advance as key elements in the semantic structure by using the qualities selected by the feature extractor 330 in machine learning. .

Finally, in the semantic structure forming unit 370, speech acts, driving, and key elements with respect to speech input in a specific domain through the speech act classification unit 340, the driving classification unit 350, and the core element extractor 360 The semantic structure is formed from the detailed elements in to understand the meaning of the first input speech.

Although the present invention has been described with reference to the embodiments shown in the drawings, those skilled in the art will understand that various modifications and equivalent other embodiments are possible therefrom. Therefore, the true technical protection scope of the present invention will be defined by the technical spirit of the appended claims.

As described above, the voice transition method and apparatus using the information extraction method according to the present invention can be efficiently applied to understand the speech language used in a specific domain, and in particular, the speech language based on a semantic structure suitable for each specific domain. In understanding this, it is possible to selectively understand the key elements (driving, speech acts, key elements), thereby providing an effect of improving the understanding of the speech language.

Claims (7)

Formulating a semantic structure for a speech language in advance; Embodying the formalized semantic structure as described above for a specific domain; Inputting a recognized speech language through a speech recognizer; Performing natural language processing on the input voice language; Selecting a feature having a characteristic meaning to determine a semantic structure based on a result analyzed through natural language processing; Machine learning using the selected qualities; And And a step of understanding the speech language based on the semantic structure formed by determining the corresponding elements constituting the semantic structure through the machine learning. The method of claim 1, wherein the formalizing of the semantic structure comprises: a speech act for classifying a sentence form of a general aspect into a semantic structure, a driving that is intended in a specific domain, and a key element supporting the behavioral pattern of the driving; A speech language understanding method using information extraction technique characterized in that the setting. The method of claim 1, wherein the embodying the semantic structure defines the speech act as a sentence, a question, a request, and the like, and includes one or more detailed behavioral patterns and the detailed behavioral patterns intended for the specific domain in driving. How to understand the speech language using information extraction technique characterized in that it defines the sub-elements included in the key elements that support them. The method of claim 1, wherein the specific domain is a domain having a distinct purpose, such as a program guide or a car road guide. The method of claim 1, wherein the natural language processing performs morphological analysis, sphere structure analysis, and syntax structure analysis on the speech language. A speech input unit for receiving a speech language to be understood from an operation result of the speech recognizer; A natural language processor analyzing the recognized speech language inputted into the speech input unit according to a natural language processing process; A feature extraction unit for selecting a feature having a characteristic meaning to be used as a basis for determining a meaning structure later from the result of the natural language processing unit; A speech act classification unit for classifying a speech act in a semantic structure into a predefined sentence form by applying a feature set selected by the feature extractor to a machine learning technique; A driving classification unit for applying the feature set selected by the feature extracting unit to a machine learning technique to classify the driving in a semantic structure into a predefined detailed behavior pattern; A core element extracting unit for classifying a feature set selected by the feature extracting unit into a machine learning technique to classify the core elements supporting the behavior of the driving into detailed sub-elements in a semantic structure; And The speech using speech information extraction technique, characterized in that the speech act classification unit, the driving classification unit and the core element extracting unit, comprising a semantic structure forming unit for forming a semantic structure of the speech language as the detailed elements classified into each Language understanding device. The apparatus of claim 6, wherein the natural language processor performs morphological analysis, sphere structure analysis, and syntax structure analysis on the speech language.

Images