KR100904049B1 - System and Method for Classifying Named Entities from Speech Recongnition - Google Patents

Abstract

According to another aspect of the present invention, there is provided a statistical semantic classification system for speech recognition, comprising: a speech recognition unit configured to extract input speech signals into a predefined text corpus and classify them in word units; A morpheme analyzer for segmenting the text corpus classified into word units into morpheme units; A mapping unit which introduces a concept dictionary in which words having the same concept are semantically grouped, and groups each morpheme segmented by the morpheme analyzer into a representative word; And a semantic classification unit for assigning the representative word to the vector generated by the statistical-based semantic classifier and automatically assigning the semantic category to each semantic category.

The statistical semantic classification system and method for speech recognition according to the present invention stabilizes dictionary size by applying a morpheme analyzer to segment each morpheme of Korean word by word, and applying a stopword dictionary and a conceptual dictionary to the search process. It has the effect of improving the performance of speech recognition and semantic classification.

 Voice, Morphology, Vector Extraction, Concept Dictionary

Description

System and Method for Classifying Named Entities from Speech Recongnition}

1 is a block diagram of a conventional speech recognition system.

2 is a schematic diagram of a speech recognition and semantic classification system according to a preferred embodiment of the present invention.

3 is a block diagram of a semantic classification system according to a preferred embodiment of the present invention.

4 is a block diagram illustrating a semantic classifier of a semantic classification system according to an exemplary embodiment of the present invention.

5 is a diagram illustrating a structure of a neural network based integrated speech recognizer according to the present invention.

FIG. 6 is a diagram illustrating an example in which the neural network based integrated speech recognizer of FIG. 5 is applied.

7 is a flowchart of a statistical semantic classification system for speech recognition according to the present invention.

8 is a flowchart of a semantic classifier of a statistical semantic classification system for speech recognition according to the present invention.

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

10: speech recognition unit 15: learning data unit

20: morpheme analyzer 22: morpheme phonetic dictionary DB

24: management module 26: classification module

28: recognition module 30: mapping unit

32: Conceptual Dictionary DB 34: Grouping Module

40: Terminology Removal Unit 42: Terminology Dictionary DB

44: stop-word removal module 50: feature selection unit

60: vector extraction unit 70: vector learning unit

80: semantic classification

The present invention relates to a statistical semantic classification system and method for speech recognition, and more particularly, to automatically classify the semantic categories of utterances spoken by users by morphemes, thereby improving performance of Korean speech recognition and semantic classification. The present invention relates to a statistical semantic classification system and method for speech recognition.

In general, speech recognition technology enables an interface between a person and a computer through a person's voice. A computer analyzes a voice of a person having a specific frequency according to pronunciation, converts it into an electric signal, and then extracts frequency characteristics of the voice signal. This is the skill of understanding pronunciation and performing tasks according to the understood voice. Such voice recognition technology is currently commercialized and applied to various fields such as telephone dialing, toy control, language learning or home appliance control.

In the conventional speech recognition technology, as illustrated in FIG. 1, when a speaker speaks through a specific terminal 100, the spoken speech signal is transferred to the speech recognition system 100 to extract and calculate information. Finally, the voice signal spoken by the speaker is converted into text 300.

In addition, the modules constituting the conventional speech recognition system 100 usually perform learning and computation in five ways. As shown in FIG. 1, the feature extractor 110, an acoustic model unit 212, a pronunciation model unit 222, a language model unit 232, and a post-processing unit 240 are included.

The feature extractor 110 is a process of extracting useful features from a speech signal, extracting feature representations that are perceptually meaningful, and robust features to various noise environments, speakers, and channel variations. do.

The acoustic model unit 212 shows how the speech signal from the speech database 210 can be represented. Recently, the most widely used acoustic model in speech recognizer is based on HMM (hidden Markov model). The basic unit of an acoustic model is a phoneme or similar phoneme unit. Each model represents one acoustic model unit and usually consists of three states. Usually only transitions between states from left to right are allowed. The observed probabilities of the speech feature vectors in each state are expressed as discrete probability distributions or continuous probability density functions (pdf).

The pronunciation model unit 222 is a model for changing the phonemes to phonemes because the phonemes to be actually learned are phonetic phones, not phonetic phones. The model is usually implemented by establishing a phonetic dictionary DB (database) 220 by defining simple rules based on standard phonetics or by taking into account specific environments, characteristics of speakers and dialects.

The language model unit 232 may be referred to as a grammar of a speech recognizer. This is a process of extracting grammar from text corpus DB 230 to select sentences that match grammar rather than random sentences during learning and searching. The language model unit 232 may reduce the search space of the speech recognizer and contribute to the improvement of the recognition rate since the language model unit 232 increases the probability of sentences that match the grammar.

Although the post-processing unit 240 may not be considered in some cases, the candidate text having a high recognition rate is usually selected through a recognizer, and then another processed linguistic information or an error pattern is learned and applied to the most suitable text. 300 is the process of finding.

Meanwhile, in order for the computer system to understand the text 300 selected by the post-processing unit 240, the concept of the word expressed in the spoken speech and the "relationship between the concepts" to be expressed in the sentence should be identified. . To this end, conventional semantic classification methods include rule-based methods for designing and processing semantic grammars and automatic semantic categorization of speech based on probabilistic statistics.

Although semantic grammar can be usefully used in a limited area because it combines syntax and semantic processing of speech, it has a disadvantage in that it is difficult to design a semantic grammar rule and generalize rules and is difficult to extend.

On the other hand, the statistical semantic automatic semantic categorization can reduce the development cost of rule development because the system can automatically learn the regularity of the language, provided that learning data is given. The statistics-based semantic classifier is a system that is produced by statistically automatically learning the learning data separated by human as much as the semantic category predefined from the collected utterances by machine learning algorithm. The conventional semantic classifier model recognizes speech spoken by a user by a statistical speech language recognizer, receives the recognized result as text, classifies it into a semantic category class through a statistical semantic classifier, and automatically assigns user speech by semantic category.

However, since most conventional semantic classifier models are designed for English, the semantic classifier model is suitable for selecting recognition words composed of English recognition units and floating units, but not for processing morphological analysis, synonyms, and the like.

In particular, since Korean is a deadlock, a word is composed of real morphemes and hints such as surveys and endings. A method of stabilizing and increasing the speech recognition rate has been proposed.

In order to solve the problems of the prior art as described above, an object of the present invention is to stabilize the dictionary size by applying a morpheme analyzer for segmenting each morpheme in Korean by word units, and applying a stopword dictionary and a conceptual dictionary to the search process. The present invention provides a statistical semantic classification system and method for speech recognition to improve the performance of speech recognition and semantic classification.

In order to achieve the above object, the present invention provides a statistical semantic classification system for speech recognition, comprising: a speech recognition unit configured to extract input speech signals into predefined text corpus and classify them in word units; A morpheme analyzer for segmenting the text corpus classified into word units into morpheme units; A mapping unit which introduces a concept dictionary in which words having the same concept are semantically grouped, and groups each morpheme segmented by the morpheme analyzer into a representative word; And a semantic classification unit for assigning the representative word to the vector generated by the statistical-based semantic classifier and automatically assigning the semantic categories to each semantic category.

The semantic classifier may include: a learning data unit in which a text corpus generated by the speech recognition unit is collected; A morpheme analyzer for segmenting the text corpus collected in the learning data unit into morpheme units; A mapping unit which introduces a concept dictionary in which words having the same concept are semantically grouped, and groups each morpheme segmented by the morpheme analyzer into a representative word; A feature selection unit for selecting a feature useful for categorizing classification from the representative words; And a vector extraction unit for extracting the features selected by the feature selection unit into a specific vector by a machine learning algorithm.

In addition, the morpheme analyzer is a morpheme phonetic dictionary DB each morpheme is classified into basic units; A management module for managing the morpheme phonetic dictionary DB; A classification module for classifying the text corpus generated by the speech recognition unit into each morpheme using the management module; And a recognition module for recognizing the morphemes classified by the classification module.

The mapping unit may include a concept dictionary DB grouping semantically identical words; And a grouping module for substituting each morpheme segmented by the morpheme analyzer into a conceptual dictionary DB and grouping the representative morpheme into a representative word.

The present invention also provides a statistical semantic classification system for speech recognition, comprising: a stopword removing unit for removing stopwords which do not give any other information among the representative words grouped by the mapping unit.

In addition, the stopword removal unit is a stopword dictionary DB consisting of stopwords, such as phrases, words onomatopoei do not give much information; And a stopword removal module for removing stopwords that do not give any other information among the representative words by using the stopword dictionary DB.

The feature selection unit may include a feature selection module configured to extract a mutual information scale and chi-square statistics about the representative word; And a word extraction module for applying and removing data having a low semantic classification feature value from the feature selection module to the stopword removal unit.

The vector extractor has a neural network structure including an input layer, a hidden layer, and an output layer, and generates a target output value for an input pattern by adjusting neural network weights between layers through the machine learning algorithm based on the neural network structure. And fusion to extract a speech feature vector for the speech signal.

The present invention also provides a statistical semantic classification system for speech recognition, comprising a vector learning unit for automatically learning the vector by applying the machine learning algorithm to the vector extracted from the vector extracting unit.

In addition, the machine learning algorithm provides a statistical semantic classification system for speech recognition, characterized in that the ANNs (Artificial Neural Networks) algorithm.

Also, in the statistical meaning classification method for speech recognition using the statistical meaning classification system for speech recognition, a speech recognition step of extracting and classifying the text corpus used for speech recognition from the input speech signal by word unit ; A morpheme analysis step of segmenting the text corpus classified into word units into morpheme units; A mapping step of grouping each morpheme segmented in the morpheme analysis step into a representative word by introducing a concept dictionary in which words having the same concept are semantically grouped; And a semantic classification step of assigning the representative word to the vector generated by the semantic classifier and automatically assigning the semantic word to each semantic category.

In addition, the semantic classifier is a learning data collection step of collecting the text corpus generated by the speech recognition unit; A morpheme analysis step of segmenting the text corpus classified into word units into morpheme units; A mapping step of grouping each morpheme segmented in the morpheme analysis step into a representative word by introducing a concept dictionary in which words having the same concept are semantically grouped; A stopword removal step of removing stopwords which do not give any other information among the representative words grouped in the mapping step; A feature selection step of selecting a feature useful for categorizing classification among the representative words from which the stopword is removed in the stopword removal step; And a vector extracting step of extracting the feature into a speech feature vector by the machine learning algorithm.

The present invention also provides a statistical semantic classification method for speech recognition, comprising: removing a stopword which does not give any other information among the representative words grouped in the mapping step.

The present invention also provides a statistical semantic classification method for speech recognition, comprising a vector learning step of automatically learning the vector by applying the machine learning algorithm to the vector extracted in the vector extracting step.

Hereinafter, a statistical semantic classification system and method for speech recognition according to a preferred embodiment of the present invention will be described in more detail with reference to the accompanying drawings.

2 is a schematic diagram of a speech recognition and semantic classification system according to a preferred embodiment of the present invention.

Referring to FIG. 2, when a speaker speaks through a specific terminal 100, the spoken voice signal is transmitted to the voice recognition system 100 to extract and calculate information. The speech recognition system 100 has a conventional feature extractor 110, an acoustic model unit 212, a pronunciation model unit 222, a language model unit 232, and a post-processing unit 240, similar to the conventional (FIG. 1). It is configured to include, which is the same configuration as the conventional speech recognition system 100, a more detailed description will be omitted.

The text 300 selected by the post processor 240 is automatically assigned to each semantic category of the text 300 so that the computer system can understand the meaning of the text 300 through the semantic classification system 5. .

3 is a block diagram of a semantic classification system according to a preferred embodiment of the present invention.

Referring to FIG. 3, the semantic classification system 5 of the present invention is a system for automatic semantic categorization processing of speech based on probabilistic statistics. The speech recognition unit 10, the morpheme analyzer 20, and the mapping unit 30 are described. ), The stopword removal unit 40 and the semantic classification unit 80 is configured.

The speech recognizer 10 extracts the text 300 selected by the speech recognition system of FIG. 2 into a predefined text corpus, and classifies the text 300 in word units. Separate and recognize as a specific text corpus.

The morpheme analyzer 20 analyzes each morpheme, which is a minimum unit having a meaning, from a single word in which a bundle of various morphemes appears in the form of a surface, and grasps the original structure of a word used in an actual sentence. The corpus is segmented from word units to morpheme units. The segmented morphemes group words having the same concept through the mapping unit 30, and remove unnecessary words such as unnecessary affixes through the stopword removal unit 40. .

In this way, the word expressing the contents of the utterance can be selected from the word unit to the morpheme unit through the word separation process through morphological analysis, so only the roots necessary for the content analysis can be selected. It becomes possible.

The morpheme analyzer 20 has a morpheme sounding dictionary module 22 in which each morpheme is classified into basic units, a management module 24 for managing the morpheme sounding dictionary module 22, and the morpheme sounding dictionary module 22. And a classification module 26 for classifying the text corpus into respective morphemes, and a recognition module 28 for recognizing the morphemes classified by the classification module 26. Looking at the specific operation of these, first, the classification module 26 is linked to the morpheme phonetic dictionary DB 22 composed of morphemes as a basic phoneme unit group, and classifies and models the text corpus into each morpheme. The recognition module 28 recognizes the text corpus classified by the classification module 26 in morpheme units.

The mapping unit 30 groups the morphemes segmented by the morpheme analyzer 20 into representative words through the grouping module 34 introducing the concept dictionary DB 32 in which the words having the same concepts are grouped together. It is composed. By the mapping unit 30, the text corpus is remapped to a representative entity name having the same semantic concept or a newly defined entity name representing the meaning, while filtering multiple text corpus having the same meaning to one text corpus. Will be. Here, the concept dictionary DB (32) is a database in which a dictionary is grouped of semantically identical concepts except for similar pronunciation words, and solves an insufficient sparse data problem frequently found in existing statistical models. It can be used as a tool to make up for the weaknesses of existing statistical models.

Terminology removal unit 40 is a group of representative words grouped in the mapping unit 30 using the stopword dictionary DB 42 and the stopword dictionary DB 42 composed of stopwords, such as phrases, words, and the like that do not give much information. It is composed of a stopword removal module 44 for removing stopwords that do not give much information, so as to obtain data that removes stopwords that do not give much information, that is, meaningless phrases, words, onomatopoeia, etc., from utterance data. As a result, data obtained by excluding unnecessary words is obtained from the utterance data that has undergone the concept dictionary application step, thereby reducing the cost of data analysis that is unnecessary when extracting keyword feature values. And such stopword removal unit 40 may not be considered in some cases.

The semantic classification unit 80 substitutes the representative word into a vector generated by the statistical-based semantic classifier 82 so that the representative word is automatically allocated for each semantic category. The semantic classifier 82 will be described in detail with reference to FIG. 4.

As such, the speech signal uttered through the specific terminal 100 is selected as text through the speech recognition system 100, and the selected text is recognized by the speech recognition unit 10 as a text corpus, followed by the morpheme analyzer 20, Through the mapping unit 30, the stop word removing unit 40, and the semantic classification unit 80, the semantic category for the text is automatically assigned so that the computer system can understand the meaning of the text.

4 is a block diagram illustrating a semantic classifier of a semantic classification system according to an exemplary embodiment of the present invention.

4, the semantic classifier 82 of the semantic classification system 5 according to the preferred embodiment of the present invention includes a learning data unit 15, a morpheme analyzer 20, a mapping unit 30, and a stopword removal unit ( 40), the feature selection unit 50, the vector extraction unit 60 and the vector learning unit 70.

The learning data unit 15 is configured to collect a plurality of text corpus generated by the voice recognition unit 10 of FIG. 3.

The morpheme analyzer 20 analyzes each morpheme, which is a minimum unit having a meaning, from a single word in which a bundle of various morphemes is represented in a surface form, and grasps the original structure of a word used in an actual sentence. The text corpus collected in (15) is segmented from word units to morpheme units using a morpheme analyzer, and the segmented morphemes are grouped with words having the same concept through the mapping unit 30, and the stopword unit 40 ) To eliminate unnecessary words such as unnecessary macros.

In addition, since the morpheme analyzer 20, the mapping unit 30, and the stop word removing unit 40 are the same as those of FIG. 2, detailed description thereof will be omitted.

The feature selection unit 50 is for keyword extraction, and is composed of a feature selection module and a word extraction module for enhancing semantic categorization performance of speech.

The feature selection module is a mutual information scale for the representative words appearing in the training data to select words that can be usefully used for categorization among the representative words generated by the mapping unit 30 or the stopword removal unit 40. ) And Chi-Square, and only words with a large amount of information are selected and used for the semantic classifier.

The word extraction module applies and removes data having a low semantic classification feature value to the stopword removal unit 40 in the feature selection module.

Table 1, Equation 1, and Equation 2 are examples of the feature selection module of the feature selection unit 50. First, in Table 1, A includes feature candidate words among utterances belonging to a semantic category. B is the number of utterances, and B is the number of utterances containing the candidate candidate words among utterances not belonging to the semantic category. C is the number of utterances that do not include the candidate candidate words among the utterances belonging to the semantic category, and D is the number of utterances that do not include the candidate candidate words among the utterances which do not belong to the semantic category.

TABLE 1

Meaning Category NOT semantic category Qualification candidate words A B NOT qualifier candidate words C D

The chi-square statistic of Equation 1 and the mutual information scale of Equation 2 are calculated as statistics by using the number of occurrences of A, B, C, and D in Table 1 to form a word list having excellent qualities. Select.

[Equation 1]

Here, N means the total number of ignitions.

[Equation 2]

Here, N means the total number of ignitions.

The vector extraction unit 60 extracts a feature selected by the feature selection unit 50 as a specific vector, and uses a vector space model to express the characteristics of the speech using the words selected above. The vector space model is composed of a machine learning algorithm to which neural network theory is applied. Using this machine learning algorithm, the learning speech data can be represented as a vector using only a word list having a large amount of information selected by the feature selection unit 50. The machine learning algorithm is composed of ANNs (Artificial Neural Networks) algorithm.

The vector learning unit 70 is configured to automatically learn the vector by applying the machine learning algorithm to the vector extracted from the vector extracting unit 60, which will be described below with reference to FIG. 5.

The vector learning unit 70 trained by the neural network-based integrated speech recognizer uses a speech recognition system as shown in FIG. It will be able to identify and provide services that meet the user's purpose.

5 is a diagram illustrating a structure of a neural network based integrated speech recognizer according to the present invention. Referring to FIG. 5, the vector extracted by the vector extractor 60 of FIG. 4 is configured to be trained by an ANNs (Artificial Neural Networks) algorithm, which is a machine learning algorithm to which a neural network theory is applied, wherein the ANNs algorithm inputs data. Inputting an input layer, a hidden layer for inputting a signal from the input layer and an output signal of a previous layer, and inputting a signal from the input layer and an output signal of the hidden layer to perform learning and outputting a final result. For the output layer. And through the machine learning algorithm based on the neural network structure, the neural network weights between the layers are adjusted to generate and fuse a target output value for the input pattern to extract the speech feature vector for the speech signal.

FIG. 6 illustrates an example in which the neural network based integrated speech recognizer of FIG. 5 is applied, wherein X1 to Xn are input values of the neural network, that is, an input layer, h1 to hk are hidden layers, and y1 to ys are output results for respective input values. That is, the output layer is shown. Here, the input value is a semantic vector value of speech composed of qualities, and each Ys is an output of a statistical value at which the speech belongs to Ys. Learning is done by determining the maximum value of Ys as a semantic category for speech.

7 is a flowchart of a statistical semantic classification system for speech recognition according to the present invention.

Referring to the drawings, first, the voice is recognized and generated as text through the speech recognition system 100, and the generated text is extracted by the speech recognition unit 10 into text corpus and classified into word units (S110). The text corpus classified in word units is segmented in morpheme units by the morpheme analyzer 20 (S120).

The text corpus segmented in the morpheme unit is grouped as a representative word by the mapping unit 30 grouping the words that are semantically identical in concept, and in this case, the text corpus is segmented in the morpheme analyzer 20 by the grouping module 34. Each morpheme is substituted into the conceptual dictionary DB 32 to be grouped into representative words (S130). Unnecessary words such as unnecessary affixes among the cannon words may be removed through the stop word removing unit 40.

The mapping unit 30 and the stop word removing unit 40 filter the text corpus segmented in morpheme units, and are used to reduce the comparison data when the feature selection unit 50 analyzes excellent feature values.

The representative words grouped by the mapping unit 30 are assigned to the vector generated by the statistical-based semantic classifier 82 and automatically assigned to each semantic category (S140).

8 is a flowchart of a semantic classifier of a statistical semantic classification system for speech recognition according to the present invention.

Referring to the drawings, first, the learning data unit 15 collects the text corpus generated by the speech recognition unit 10 (S210). The collected text corpus, that is, the learning data is collected by the typeface analyzer 20. It is divided into morphological units. (S220)

The text corpus segmented in the morpheme unit is grouped as a representative word by the mapping unit 30 to group words that are semantically identical in concept (S230), and the grouping module 34 makes the morpheme analyzer 20. Each morpheme segmented in is substituted into the conceptual dictionary DB 32, and grouped into representative words. (S240) And stopwords such as unnecessary affixes that do not give much information among the cannon words are removed through the stopword removal unit 40. It may be. The mapping unit 30 and the stop word removing unit 40 filter the text corpus segmented in morpheme units, and are used to reduce the comparison data when the feature selection unit 50 analyzes excellent feature values.

The feature selector 50 selects a feature useful for categorization among the representative words. (S250) And the feature selected by the feature selector 50 is extracted as a specific vector by the vector extractor. The extracted vector is automatically learned by the machine learning algorithm applying the ANNs algorithm.

The specific vector learned in this way is used by the semantic classification unit 80 to recognize the speaker's voice, thereby statistically classifying the meaning of the utterance so as to grasp the user's intention and provide a service suitable for the user's purpose. do.

While the invention has been described in detail in the foregoing embodiments, it will be apparent to those skilled in the art that various modifications and variations are possible within the spirit of the invention, and such modifications and variations belong to the appended claims.

As described above, the statistical semantic classification system and method for speech recognition according to the present invention is a dictionary size by applying a morpheme analyzer for segmenting each morpheme in Korean by word units, and applying a stopword dictionary and a conceptual dictionary to a search process. It is effective to stabilize the performance and to improve the performance of Korean speech recognition and semantic classification.

Claims (14)

delete In the statistical semantic classification system for speech recognition, A speech recognition unit for extracting the input speech signal into a predefined text corpus and classifying the speech unit into word units; A morpheme analyzer for segmenting the text corpus classified into word units into morpheme units; A mapping unit which introduces a concept dictionary in which words having the same concept are semantically grouped, and groups each morpheme segmented by the morpheme analyzer into a representative word; And A semantic classification unit for assigning the representative word to a vector generated by a statistic based semantic classifier and automatically assigning the semantic category to each semantic category; Including, The semantic classifier is: A learning data unit in which a text corpus generated by the speech recognition unit is collected; A morpheme analyzer for segmenting the text corpus collected in the learning data unit into morpheme units; A mapping unit which introduces a concept dictionary in which words having the same concept are semantically grouped, and groups each morpheme segmented by the morpheme analyzer into a representative word; A feature selection unit for selecting a feature useful for categorizing classification from the representative words; And A vector extraction unit for extracting the feature selected by the feature selection unit into a specific vector by a machine learning algorithm; Statistical semantic classification system for speech recognition comprising a. In the statistical semantic classification system for speech recognition, A speech recognition unit for extracting the input speech signal into a predefined text corpus and classifying the speech unit into word units; A morpheme analyzer for segmenting the text corpus classified into word units into morpheme units; A mapping unit which introduces a concept dictionary in which words having the same concept are semantically grouped, and groups each morpheme segmented by the morpheme analyzer into a representative word; And A semantic classification unit for assigning the representative word to a vector generated by a statistic based semantic classifier and automatically assigning the semantic category to each semantic category; Including, The morphological analyzer is: Morpheme phoneme dictionary DB, in which each morpheme is classified into basic units; A management module for managing the morpheme phonetic dictionary DB; A classification module for classifying the text corpus generated by the speech recognition unit into each morpheme using the management module; And A recognition module for recognizing the morphemes classified by the classification module; Statistical semantic classification system for speech recognition comprising a. In the statistical semantic classification system for speech recognition, A speech recognition unit for extracting the input speech signal into a predefined text corpus and classifying the speech unit into word units; A morpheme analyzer for segmenting the text corpus classified into word units into morpheme units; A mapping unit which introduces a concept dictionary in which words having the same concept are semantically grouped, and groups each morpheme segmented by the morpheme analyzer into a representative word; And A semantic classification unit for assigning the representative word to a vector generated by a statistic based semantic classifier and automatically assigning the semantic category to each semantic category; Including, The mapping unit: A conceptual dictionary that semantically groups words with the same concept; And A grouping module for assigning each morpheme segmented by the morpheme analyzer to a conceptual dictionary DB and grouping the representative morpheme into a representative word; Statistical semantic classification system for speech recognition comprising a. In the statistical semantic classification system for speech recognition, A speech recognition unit for extracting the input speech signal into a predefined text corpus and classifying the speech unit into word units; A morpheme analyzer for segmenting the text corpus classified into word units into morpheme units; A mapping unit which introduces a concept dictionary in which words having the same concept are semantically grouped, and groups each morpheme segmented by the morpheme analyzer into a representative word; And A semantic classification unit for assigning the representative word to a vector generated by a statistic based semantic classifier and automatically assigning the semantic category to each semantic category; Including, Statistical term classification system for speech recognition, characterized in that it further comprises a stop-word removal unit for removing stop words that do not give much information among the representative words grouped by the mapping unit. The method of claim 5, wherein The stopword removal unit: Terminology dictionary DB consisting of phrases that do not give much information, and stopwords of words onomatopoeia; And A stopword removal module for removing stopwords which do not give any other information among the representative words using the stopword dictionary DB; Statistical semantic classification system for speech recognition comprising a. The method of claim 2, The feature selection unit: A feature selection module for extracting mutual information scales and chi-square statistics for the representative words; And A word extraction module for applying and removing data having a low semantic classification feature value from the feature selection module to a stop word removing unit for removing stop words that do not give other information among the representative words grouped by the mapping unit; Statistical semantic classification system for speech recognition comprising a. The method of claim 2, The vector extraction unit: The neural network structure includes an input layer, a hidden layer, and an output layer, and adjusts neural network weights between layers through the machine learning algorithm based on the neural network structure to generate and fuse a target output value for an input pattern, thereby generating the voice signal. And extract a speech feature vector for the statistical semantic classification system for speech recognition. The method of claim 2, And a vector learning unit for automatically learning the vector by applying the machine learning algorithm to the vector extracted from the vector extracting unit. The method according to claim 8 or 9, The machine learning algorithm is a statistical semantic classification system for speech recognition, characterized in that consisting of the ANNs (Artificial Neural Networks) algorithm. delete In the statistical meaning classification method for speech recognition using the statistical meaning classification system for speech recognition of claim 2, A speech recognition step of extracting a text corpus used for speech recognition from the input speech signal and classifying it into word units; A morpheme analysis step of segmenting the text corpus classified into word units into morpheme units; A mapping step of grouping each morpheme segmented in the morpheme analysis step into a representative word by introducing a concept dictionary in which words having the same concept are semantically grouped; And A semantic classification step of assigning the representative word to the vector generated by the semantic classifier and automatically assigning each semantic category; Including, The semantic classification step is: Learning data collection step of collecting the text corpus generated by the speech recognition unit; A morpheme analysis step of segmenting the text corpus classified into word units into morpheme units; A mapping step of grouping each morpheme segmented in the morpheme analysis step into a representative word by introducing a concept dictionary in which words having the same concept are semantically grouped; A stopword removal step of removing stopwords which do not give any other information among the representative words grouped in the mapping step; A feature selection step of selecting a feature useful for categorizing classification among the representative words from which the stopword is removed in the stopword removal step; And Extracting the feature into a speech feature vector by the machine learning algorithm; Statistical meaning classification method for speech recognition, comprising the. In the statistical meaning classification method for speech recognition using the statistical meaning classification system for speech recognition of claim 2, A speech recognition step of extracting a text corpus used for speech recognition from the input speech signal and classifying it into word units; A morpheme analysis step of segmenting the text corpus classified into word units into morpheme units; A mapping step of grouping each morpheme segmented in the morpheme analysis step into a representative word by introducing a concept dictionary in which words having the same concept are semantically grouped; And A semantic classification step of assigning the representative word to the vector generated by the semantic classifier and automatically assigning each semantic category; Including, And a stopword removing step of removing stopwords which do not give any other information among the representative words grouped in the mapping step. The method of claim 12, And a vector learning step of automatically learning the vector by applying the machine learning algorithm to the vector extracted in the vector extracting step.

Images